Liblib
估值超过 US$2 billion、ARR 约 US$300M 的 AI 创意平台独角兽
Liblib 是增长很快的中国 AI 创作平台独角兽,ARR 增速异常亮眼;但估值建立在未经审计指标上,内容安全、版权和聚合器毛利压力都还没消失。
封面要素
公司概况
Liblib(LiblibAI、哩布哩布AI)是一家总部位于北京的生成式 AI 创意平台,2023 年上线,由母公司 Evoken(前身为 Beijing Qidian Xingyu Technology / Yanyu Technology)运营。它是中国最大的 AI 图像创作社区和模型分享市场,建立在 Stable Diffusion / SDXL / Flux / ComfyUI 生态之上,并支持自定义 LoRA 微调。2026 年 6 月,公司以超过 US$2 billion 的估值完成约 US$300M Series B+ 融资,成为 AI 应用层独角兽。
- 成立时间
- 2023-05-16
- 创始人
- Chen Mian (陈冕)
- 创立地点
- Beijing, China
- 总部
- Beijing, China
- 产品
- 免费增值 AI 创意平台:文生图和图生图生成、ControlNet 工作流、自定义 LoRA 模型训练、模型分享市场(500,000+ 个原创模型)、开发者 API,以及较新的 LibTV(AI 视频)、Xingliu(国内设计 Agent)和 Lovart(国际设计 Agent)。
- 客户
- 个人 AI 创作者和设计师(C 端),以及短剧 / 电影、广告、游戏、动画和电商领域的企业 / 专业内容生产者(B 端)。
- 商业模式
- 免费增值 SaaS —— 带积分的免费层、付费订阅和算力积分(C 端),以及 API / 企业授权 / 私有化部署(B 端)。
- 阶段
- Series B+ (private unicorn)
- 融资情况
- 2026 年 6 月以 >US$2B 估值完成约 US$300M Series B+ 融资,由 Granite Asia、Tencent 和 Shunwei Capital 共同领投;累计融资总额超过 US$500M。
执行摘要
主要优势
- 商业化爆发:到 2026 年 5 月,ARR 达到 ~US$300M,据报道同比增长超过 30x,专业 B 端内容生产者是基本盘,覆盖短剧、影视和广告。
- 在中国 AI 图像创作社区里占据头部位置,累计用户超过 30M,原创模型超过 500,000 个,每天生成图片超过 5M 张。
- 投资人阵容一线且战略协同度高(Granite Asia、Tencent、Shunwei、Sequoia China、Ant Group),能带来资金、分发以及云 / 模型资源。
- 产品从图像快速扩到 AI 视频(LibTV)和设计智能体(Xingliu、Lovart),踩上中国 >RMB120B 微短剧需求池。
主要风险
- 内容安全和监管风险高:央视曝光过审核绕过漏洞,可生成色情内容;同时 CAC / AIGC 标识制度(GB 45438-2025)正在收紧。
- 版权 / IP 责任难躲:中国法院在 Ultraman v. Acgnai 案中认定,生成式 AI 平台可能因侵权用户训练模型和输出承担帮助侵权责任。
- 聚合器 / “wrapper” 毛利风险:Evoken 没有训练自有基础模型,依赖第三方模型 API(如 ByteDance Seedance),毛利和防御性都会受压;公司目前仍未盈利。
- 估值依赖公司披露、未经审计的 ARR 和 >30x 增长;中国 AI 独角兽若经历泡沫退潮 / 重估,倍数可能被压缩。
未决问题
- 没有经审计财务:毛利率、净收入留存、烧钱速度和现金跑道均未披露。
- C 端订阅与 B 端 API / 企业业务的 ARR 拆分,以及客户集中度,尚未量化。
- >30x ARR 增长能维持多久、LibTV 经济性(依赖第三方视频模型 API)能否跑通,仍未证明。
- 2025–2026 年快速扩张后的当前员工数和组织规模未披露。
目录
01公司概况
1.1 身份、结构与商业模式
本报告应把 Liblib 视为 Evoken 内部实际运营的 AI 创意平台品牌。Evoken 这家公司集团在 2026 年 6 月更名前,曾使用 Qidian Xingyu / Yanyu Technology 等名称。硬身份事实包括:总部在北京,运营公司成立于 2023 年 5 月,旗舰产品 LiblibAI 最初是 AI 图像生成与模型分享社区。它的商业模式不是通用聊天机器人套壳,而是把创作者社区、模型市场、云端生成、LoRA 训练和付费专业工作流组合在一起。这个结构很关键,因为后续章节可以沿着一条连贯链条评估市场、产品和估值:社区模型和资产输入生成、视频制作和设计 Agent 用例。官方网站只适合作为公司主张使用,因为页面完全依赖 JavaScript;因此,这里的稳定身份基础依赖第三方新闻、类似登记资料的 Baike 页面,以及 CAC 备案生态。[CO001, CO002, CO003, CO004, CO005, CO006]
| 字段 | 数值 / 状态 | 日期 | 置信度 | 缺口 |
|---|---|---|---|---|
| 成立 | 运营主体成立于 2023-05-16;LiblibAI 于 2023 年上线 | 2023-05-16 | 中 | 具体平台上线月份各来源不一,因此用公司成立日期界定法律主体。 |
| 总部 | 中国北京 | 2026-07-15 | 高 | 街道级总部确认仍停留在工商登记层面,并非公司发布。 |
| 母公司 | Evoken;曾用名 Qidian Xingyu / Yanyu Technology | 2026-06 | 高 | 公司更名过程公开,但股权表实体链条未完全披露。 |
| 创始人 | Chen Mian,曾任 ByteDance Jianying / CapCut 商业化负责人 | 2026-07-15 | 高 | 董事会和保护性控制文件未公开。 |
| 赛道 | AI 创意平台 / AI 应用层 | 2026-07-15 | 中 | 与模型实验室和设计 SaaS 的边界属于分析判断。 |
| 阶段 | 未上市 Series B+ 独角兽 | 2026-06-18 | 高 | 第 1 章来源池未发现 IPO 申报。 |
| 用户 | LiblibAI 累计用户 30M+ | 2026-06 | 高 | 活跃用户和付费用户拆分未披露。 |
| ARR | 截至 2026 年 5 月 ARR 超 US$300M | 2026-05 | 高 | 毛利率和留存仍未公开。 |
| 估值 | Series B+ 后投后估值 >US$2B | 2026-06-18 | 高 | 高于该门槛的具体投后估值未披露。 |
快照采用简报中的标准指标;缺口标示缺失的私有证据,而非零值。
[CO001, CO002, CO003, CO004, CO007, CO023]集团结构串起公司身份、创始人控制和四个创意产品界面。
股权为披露的创始人持股;完整股权表和实体树未公开。
[CO002, CO005, CO009, CO035, CO038, CO039]1.2 创始人、领导层与治理信号
创始人故事对投资判断异常关键。公开叙事反复把陈冕塑造成产品和商业化型创始人:他曾负责 ByteDance 的 Jianying / CapCut 全球商业化,外界围绕他的叙事集中在速度、PMF 纪律和专业创作者工作流。这个背景契合 LiblibAI 的推进路径:专业图像和视频创作者付费购买的是生产基础设施,而不只是试用一个新奇模型。集中度风险同样真实。公开材料显示陈冕持股约 73.95%,处于控股位置;张子杰出现在早期法定代表人和核心产品社区披露中,杨楠则在 2025 年 11 月成为法定代表人。披露的团队底子亮眼,但仍停留在高层概括:成员来自中国及海外顶尖高校,也有 Tencent、Alibaba 和 ByteDance 背景。董事会构成、保护性条款和 Series B+ 后治理安排并未公开,因此本报告应保留创始人依赖这个判断,而不是推断公司已经形成制度化治理。[CO007, CO008, CO009, CO010, CO011, CO012]
1.3 融资时间线与投资方基础
Liblib / Evoken 的融资史是全篇报告的核心记录时间线。公司似乎从 2023 年约 US$15 million 估值下的 US$3.5 million 天使轮起步,经过 2024 年多轮融资验证 AI 图像社区,再到 2025 年初完成数亿元人民币融资。2025 年 10 月 US$130 million Series B 让它成为中国应用层融资的公开标杆,也把 HSG / Sequoia China、CMC Capital、Ant、Lenovo Capital、Shunwei、Source Code、Mingshi、Yingce 和 INCE 带入叙事。2026 年 6 月 Series B+ 是跃迁点:投后估值超过 US$2 billion,融资约 US$300 million,由 Granite Asia、Tencent 和 Shunwei 领投,HT Investment、Times Capital、Gaorong、Ant、HSG / Sequoia China 也被引用为参投方。因此,投资方图谱混合了财务 VC、战略科技资本和连续跟投的老股东;公司披露的资金用途仍集中在研发、全球扩张和产品组合扩建。[CO013, CO014, CO015, CO016, CO017, CO018]
| 投资方 | 轮次 | 角色 |
|---|---|---|
| Granite Asia | 2026 Series B+ 轮 | 共同领投;B+ 财团中的新全球风投机构。 |
| Tencent Holdings | 2026 Series B+ 轮 | 共同领投的战略科技投资方。 |
| Shunwei Capital | 2024 / 2025 / 2026 | 复投方,也是 2026 年 B+ 轮共同领投方。 |
| HT Investment | 2026 Series B+ 轮 | 追加 / 参投方。 |
| Times Capital | 2026 Series B+ 轮 | 追加 / 参投方。 |
| HSG / Sequoia China | 2025 Series B 轮 / 2026 年追加投资 | Series B 共同领投方,存量投资方并在 B+ 轮加码。 |
| CMC Capital | 2025 Series B 轮 | Series B 共同领投方。 |
| Ant Group | 2025 Series B 轮 / 2026 年追加投资 | 后续轮次点名的存量战略 / 财务投资方。 |
| Gaorong Capital | 天使轮 / 2026 年追加投资 | 早期投资方和现有股东。 |
| Source Code Capital | 天使轮 / Series B 轮 | 早期投资方和 Series B 参投方。 |
| Mingshi Venture | 2024 年 / Series B 轮 | 2024 年领投方和 Series B 参投方。 |
| Yingce Capital | 2024 / 2025 | 2024–2025 年投资方和现有股东。 |
| INCE Capital | 2025 Series B 轮 | 组合投资方和追加参投方。 |
投资方角色按公开公告归类;经济持股比例未披露。
[CO014, CO015, CO019, CO020, CO024, CO025]| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2023-05-16 | Beijing Qidian Xingyu Technology Co., Ltd. 成立 | 成立 | 法律主体成立 | Zhang Zijie 法定代表人记录 | 确立后来 Liblib / Evoken 品牌背后的运营主体。 |
| 2023-09 | 天使轮融资 | 融资 | ~US$3.5M,估值 ~US$15M | 投资方:GSR / Jinshajiang、Gaorong、Source Code | 在产品大规模验证前,为 AI 图像社区提供种子资金。 |
| 2024-02 | 深度合成算法备案通过 | 监管 | 备案里程碑 | CAC / LiblibAI | 降低深度合成服务运营早期的监管悬念。 |
| 2024-03 | 生成式 AI 服务备案完成 | 监管 | 来源称其为首个 AI 社区备案 | CAC / LiblibAI | 早期备案受挫后,成为正向合规里程碑。 |
| 2024-07 | 数亿元人民币融资序列 | 融资 | 按简报口径,2024 年多轮融资合计 >US$20M | Mingshi、战略投资方、现有股东 | 验证其在中国 AI 图像工具品类的领先位置。 |
| 2025-02 | 又一轮数亿元人民币融资 | 融资 | 数亿元人民币 | PitchHub 提及 Yingce、Shunwei、Giant Network | 在应用层竞争中延长现金跑道。 |
| 2025-10-23 | Series B 融资公布 | 融资 | US$130M | HSG / Sequoia China、CMC Capital、战略投资方、内部投资方 | 第 1 章来源池中 2025 年中国 AI 应用层最大已披露融资。 |
| 2026-03 | LibTV 上线 | 产品 | AI 视频创作平台 | LiblibAI / Evoken | 从图像社区扩展到专业视频工作流。 |
| 2026-04-13 | CCTV 内容安全曝光 | 负面 | 被指存在色情生成漏洞 | CCTV / Baidu Baike 条目 | 形成合规与审核尽调红旗。 |
| 2026-05 | 披露 ARR 超过 US$300M | 规模 | US$300M+ ARR | Evoken / 媒体报道 | 把产品牵引力转成估值相关的经营指标。 |
| 2026-06-18 | Series B+ 融资 | 融资 | ~US$300M;投后 >US$2B | Granite Asia、Tencent、Shunwei、HT Investment、Times Capital、现有投资方 | 将 Liblib / Evoken 推成数十亿美元级 AI 应用独角兽。 |
| 2026-06 | Evoken 名称作为集团品牌出现 | 治理 | 公司名称 / 集团定位 | Evoken / 媒体报道 | 释放从单一产品转向产品组合控股公司的信号。 |
这是本章确认的时间线;披露精确日期时使用具体日期,仅披露月份的条目保留来源粒度。
[CO002, CO004, CO013, CO014, CO015, CO017]Liblib / Evoken 在约三年内压缩完成公司设立、监管备案、产品扩展和两轮巨额融资。
仅到月份的里程碑保留来源粒度;融资金额按规范简报口径四舍五入。
[CO013, CO018, CO023, CO027, CO036, CO041]1.4 规模、产品组合与负面信号
头部运营指标足以让公司获得独角兽级关注,但必须带着保留来读。当前主口径指标包括:LiblibAI 累计用户超过 30 million,原创模型超过 500,000 个,每日生成图像超过 5 million 张,截至 2026 年 5 月 ARR 超过 US$300 million。更早的标记包括约 25 million 用户、4 million MAU,以及 2024 年收入 RMB 206 million;这些数据有助于解释为什么 2026 年 ARR 跃升是重大加速,而不是轻微重述。产品组合如今覆盖 LiblibAI(社区 / 模型资产)、LibTV(视频制作)、Xingliu(国内设计 Agent 工作流)和 Lovart(海外设计 Agent 实验)。两个负面尽调点应放在第 1 章,而不是埋到后面:CCTV 在色情生成漏洞曝光中点名 LiblibAI;2025 年报道还称公司尚未盈利。两项风险都能与高速增长并存;也都具有实质影响,不能忽视。[CO022, CO028, CO029, CO030, CO031, CO032]
最可复用的封面指标是用户规模、创作者 / 模型存量、日生成量、ARR 和估值。
所有数值沿用规范简报;2024 年美元换算为近似值,不是新的汇率计算。
[CO023, CO028, CO029, CO030, CO031, CO033]1.5 图表
02市场分析
2.1 市场边界与规模区间
不能拿娱乐或设计支出的每一块钱给 Liblib 定价。可投资边界从中国 AIGC 软件和服务开始:这些产品自动化内容创作;随后再收窄到图像生成平台、视频和动画生成器、模型社区,以及创作者或工作室愿意购买的工作流工具。公开分析师估算跨度很大:IMARC 认为 2025 年中国 AIGC 市场为 US$5.16 billion,到 2034 年 CAGR 为 15.96%;本来源池采用的 Grand View 较低口径,则支持约 US$2.36 billion 的当前市场锚。这个区间比单一 TAM 更有用,因为 Liblib 变现的是工作流采用,而不是全部生成式 AI 基础设施;同时,公司还紧邻短剧制作中更大的内容支出池。务实的尽调模型应调和两条路径:自上而下的分析师市场预测,以及来自工作室、代理商和企业创意部门的钱包份额自下而上测算。第一条路径证明中国 AIGC 已经足够大;第二条路径决定有多少预算可由 Liblib 触达,而不是被云基础设施、基础模型订阅或内部创意人力拿走。[CM001, CM002, CM003, CM004, CM005, CM006]
| 来源 / 视角 | 年份或期间 | 测算市场 | 数值(除非注明均为 USD) | CAGR / 增长 | 方法或限制 |
|---|---|---|---|---|---|
| IMARC | 2025 | 中国 AIGC 市场 | 5.16B | 到 2034 年 CAGR 15.96% | 自上而下的分析师估算,覆盖组件、部署和技术 |
| IMARC 预测 | 2034 | 中国 AIGC 市场 | 19.56B | 2026–2034 年 CAGR 15.96% | 预测终点,不代表当前软件支出 |
| Grand View Research | 2024 年 / 当前视角 | 中国生成式 AI 数据手册 | ~2.36B 低口径 | 来源池给出的 ~42% 广义增长口径 | 公开页面展示分段;精确细节受证据限制 |
| DigiTrendz / TNW | 2026 | 中国微短剧内容市场 | 16.5B 等值 | 市场预计超过 RMB120B | 需求池代理,不是软件收入 |
| Wonford | 2025 | 中国短剧出海收入 | 2.38B | 同比 +263% | 海外内容变现视角 |
数值四舍五入至 US$B;AIGC 软件估算与内容市场池是不同视角,不能相加成 TAM。
[CM002, CM003, CM004, CM010, CM011, CM017]相关规模从宽口径短剧和 AIGC 需求,收窄到 Liblib 披露的 ARR 代理指标。
金字塔只把内容支出和软件 / 工作流收入作为逐步收窄的视角混用;各数值不可相加。
[CM010, CM011, CM019, CM022, CM023, CM026]公开估算差异大,因为各自衡量 AIGC 和 AI 赋能内容需求的不同部分。
所有行使用 US$B;因公开来源定义未统一,注释保留期别差异。
[CM004, CM017, CM018, CM019, CM020, CM021]2.2 短剧需求是核心 B 端拉力
最清晰的 B 端驱动来自中国短剧经济。多方来源把 2026 年短剧市场放在 RMB120 billion 以上,约 US$16.5 billion,并把这种形式描述为手机原生、高产量、算法分发的娱乐内容。AI 重要,是因为它同时压缩成本和时间:报道提到真人实拍成本的十分之一、传统拍摄成本的五分之一,制作周期从三个月降到一个月。买方行为因此改变。工作室可以跑更多试播,外包更少低价值资产,并把预算转向提示词、模型工作流、可复用角色、本地化和后期制作。对 Liblib 来说,这不意味着整个短剧市场都会变成收入;真正含义是,一个庞大且经常重复的制作预算,现在有理由购买图像、模型和视频工作流产能。这个区别很重要,因为短剧经济奖励的是可重复吞吐,而不是一次性图像质量。拥有日更剧集流水线的买方更看重可复用角色、一致风格、快速本地化和低摩擦协作;这些需求更接近工作流软件,而不是消费娱乐支出。[CM010, CM011, CM012, CM013, CM014, CM015]
| 指标 | 数值 | 期间 | 来源信号 | 对 Liblib 的含义 |
|---|---|---|---|---|
| 中国微短剧市场 | >RMB120B / ~US$16.5B | 2026 年预测 | DigiTrendz 和 The Next Web | 围绕 AI 视频 / 图像工作流的大型内容支出池 |
| 用户规模 | 660M 用户 | 2026 年引用 | The Next Web | 大众分发支撑工作室需求 |
| AI 原生 Douyin 剧目 | ~50,000 部 | 2026 年 3 月 | The Next Web / DigiTrendz | 高频资产生成带来工具需求 |
| AI 生成内容在前 100 名中的占比 | 38%,上一年为 7% | 2026 年 1 月 | DigiTrendz / TNW | AI 已在头部内容供给中占到实质比例 |
| AI 漫画风市场份额 | RMB16.8B / ~US$2.44B | 2025 | CRI | 支撑数十亿美元级 AI 专项需求切片 |
| 海外短剧出口 | US$2.38B,同比 +263% | 2025 | Wonford | 本地化和全球分发扩大 SAM |
微短剧指标描述的是制作 / 内容需求;只有其中一部分能转化为 AI 创意软件和工作流收入。
[CM010, CM011, CM012, CM013, CM017, CM019]需求路径从低成本生成起步,最终落到专业制片方的付费、经常性工作流使用。
漏斗数值是流程阶段的方向性指数,不是转化率。
[CM012, CM013, CM014, CM015, CM023, CM026]2.3 细分市场、买方与采用路径
采用路径分布在专业创作者、工作室买方、企业营销团队和平台生态之间。终端用户可能是设计师、提示词艺术家、制片人、剪辑师或运营团队,但付款方通常是工作室、代理商、品牌、游戏公司、电商商家或企业部门;这些买方本就为更快的内容产出编列预算。IMARC 的市场细分确认图像生成平台和视频生成器是被识别的 AIGC 解决方案类别;6Wresearch 则显示,中国图像生成市场由 Baidu、Alibaba、Tencent、ByteDance、SenseTime、Meitu 等大型生态塑造。因此,Liblib 的市场定位需要两层过滤:它能否在巨头模型替代品之外占住创作者工作流和资产复用;它又能否把创作者社区转化为短剧、广告、设计和电商工作流里的付费团队。买方旅程还意味着模型层多栖,但资产层粘性更高:团队可能测试多个图像或视频模型,同时留在那个保存已批准模板、生产历史和社区训练模型的平台上。[CM005, CM006, CM007, CM008, CM009, CM023]
| 细分 / 品类 | 纳入支出 | 排除支出 | 买方 / 付费方 | 与 Liblib 的相关性 |
|---|---|---|---|---|
| 文生图和图像平台 | 订阅、点数、模型训练、提示词工作流 | 线下设计人力和非 AI 素材库 | 设计团队、创作者、广告主、电商商家 | LiblibAI 核心平台和模型分享社区 |
| 文生视频 / 3D 和动画 | AI 视频生成、镜头资产、动画工作流 | 传统全案影视制作预算 | 微短剧工作室、影视团队、代理商 | LibTV 向专业制作工作流扩展 |
| 模型市场和创作者资产 | LoRA 模型、模板、工作流、API、资产复用 | 不含创意工作流层的通用云计算 | 创作者、工作室、开发者、品牌团队 | Liblib 的网络效应和转换成本层 |
| 企业内容自动化 | 营销内容、商品图、本地化、虚拟人 | 与创意产出无关的通用办公效率工具 | 营销、电商、游戏、教育部门 | 超出消费者创作者使用场景的 SAM 延展 |
| 基础模型生态 | Baidu、Alibaba、Tencent、ByteDance 模型能力 | 纯硬件资本开支和非创意模型训练 | 云 / 平台团队和企业 | 竞争性输入层和潜在替代品 |
各行把分析师分段映射到与 Liblib 相关的买方工作流;没有来源给出 Liblib 精确的分部收入拆分。
[CM001, CM005, CM006, CM007, CM008, CM009]Liblib 的采用路径取决于用户工作流、预算负责人和采用触发因素能否匹配。
矩阵是证据支持的分群,不是量化份额拆分。
[CM005, CM006, CM007, CM027, CM028, CM042]2.4 增长驱动、约束与塑造市场的监管
增长叙事很强,但并非没有摩擦。政府 AI+ 政策、企业数字化、自动化内容需求和本地制作补贴都会扩大需求。与此同时,监管不是事后补丁;它决定哪些模型可以上线、生成内容如何标识,以及哪些短剧可以分发。CAC 备案记录和 Liblib 自身备案历史表明,合规是市场准入前提。负面情景也发生在市场层面,而不只是公司层面:廉价生成可能把平台灌满同质内容,引发价格竞争,并降低买方为通用产出付费的意愿。对 Liblib 来说,持久机会在于成为一个既省钱又提高产出质量的工作流和资产市场;风险则是,如果买方能在大型模型生态之间多栖,公司会被挤压成商品化工具。尽调重点是区分需求创造和价值捕获。AI 内容的涨潮可以抬高使用指标,但只有留存、付费转化、合规韧性和差异化工作流,才能证明 Liblib 是否能捕获持久利润率。[CM029, CM030, CM031, CM032, CM033, CM034]
| 驱动因素或约束 | 方向 | 时间 | 含义 | 尽调要求 |
|---|---|---|---|---|
| 企业数字化转型和 AI+ 政策 | 驱动因素 | 2026-2034 | 推动 AIGC 从创作者场景扩展到企业工作流 | 核验营销、游戏、电商和影视团队的垂直预算 |
| 微短剧成本坍塌 | 驱动因素 | 2026 | 制作成本降至十分之一到五分之一的说法,为工具预算提供理由 | 审计付费工作室 cohort 和使用强度 |
| 创作者资产网络效应 | 驱动因素 | 当前 | 模型、模板和工作流可抬高转换成本 | 衡量重复模型使用、留存和付费转化 |
| 科技巨头模型生态 | 约束和驱动因素 | 当前 | 更强的基础模型有助于采用,但也可能让界面商品化 | 测试对 Baidu / Alibaba / Tencent / ByteDance API 和定价的依赖 |
| 监管备案、标识与内容审核 | 约束 | 当前 | 合规可能拖慢上线,也可能迫使公司加大审核投入 | 核实备案、标识控制,以及微短剧审核敞口 |
| 内容同质化与供给过剩 | 约束 | 2026 | 廉价 AI 输出可能压低通用生成价格 | 跟踪价格竞争、流失,以及高端工作流差异化 |
驱动因素和约束是市场层面的力量;公司层面的执行证据放在后续章节。
[CM029, CM030, CM031, CM032, CM033, CM034]2.5 图表
03竞争格局
3.1 格局:直接模型社区、全球替代品与中国平台入局者
Liblib 的竞争集合不是一串简单的图像生成器名单。最接近的直接同业是模型分享和创作社区:它们能托管或训练模型,开放 LoRA 式定制,并把创作者留在市场或社区循环里。本章对 Tensor.Art、SeaArt AI、Wujie AI 和 AituBo 有来源支撑;6pen 和 PixAI 仍保留在用户要求覆盖的范围内,但因留存来源文本没有可验证的当前画像,无法完整分析。全球替代品从两个方向包围同一任务:Civitai 强调开放模型发现和社区控制,Midjourney 强调专有模型质量和快速产出。中国平台威胁位于两组之上,因为 Baidu、Alibaba、Tencent 和 ByteDance 可以把图像和视频生成打包进电商、游戏、社交和创作者分发场景;Liblib 必须主动赢得使用,而不是继承使用。[CP008, CP009, CP010, CP011, CP012, CP015]
| 名称 | 来源地 | 上线 / 时间 | 模型 / 规模 | 定位 |
|---|---|---|---|---|
| LiblibAI / Evoken | 中国北京 | Sept. 2023 上线;公司 May 2023 成立 | >30M 累计用户;>500,000 个原创模型;>5M 张图片 / 天 | 中国市场 AI 创意社区、模型市场、专业工作室,视频 / 设计智能体套件继续扩展 |
| SeaArt AI | 全球云平台(保留来源未核实来源地) | 2025-2026 来源覆盖 | 图像 / 视频生成、LoRA、换脸、AI 角色、ComfyUI、超分;资料引用套餐为 US$4.79-US$75/month | 宽泛的一体化创作者套件;功能重叠明显,内容安全需盯紧 |
| Tensor.Art / 回响科技 | 中国 / 香港运营足迹 | 项目 May 2023 启动;Dec. 2023 加入 ComfyUI | >160,000 个模型;July 2024 全球月访问 3.97M;API 平台 2025 上线 | Stable Diffusion 模型托管与工作流社区,走向企业 API |
| Wujie AI / 无界AI | 中国杭州 | 根据 July 2026 清单,信息仍为当前 | 每日积分;¥100 月会员;API 接入;通过 Wujie Bantu 办理版权登记 | 中文提示词 / 参考图社区,加上数字版权登记切口 |
| AituBo AI | 保留来源未核实 | 2026 第三方对比 | 免费 / 新手向图像视频生成、编辑、头像聊天、去背景、超分、换脸 | 低门槛的新手与预算型替代品 |
| Civitai | 美国 / 全球同业(已保留官方页面) | 全球同业;官方页面在 2026 仍为当前信息 | Stable Diffusion 模型社区;免费套餐,付费套餐 US$10/month 起 | 全球模型分享同业,控制权 / 模型所有权叙事更强 |
| Midjourney | 分布式全球实验室 | 官方页面在 2026 仍为当前信息 | 60 人专有模型实验室;订阅驱动的全球替代品 | 靠专有质量和速度替代,而不是模型市场 |
| 6pen | 中国直接同业范围项 | 保留文本无法核实 | 未保留有来源支撑的当前指标 | 具名范围项,评分前需要重新找来源 |
| PixAI | 中国 / 动漫向同业范围项 | 保留文本无法核实 | 未保留有来源支撑的当前指标 | 具名范围项,评分前需要重新找来源 |
各行结合已保留的 CH3 来源文本和后续章节中的 Liblib 背景;6pen 与 PixAI 刻意标为证据缺口,而不是猜测画像。
[CP001, CP002, CP012, CP017, CP018, CP022]序数图:x 轴近似表示中国 / 本地化聚焦度,y 轴近似表示创作工作流能力广度。
序数 1–5 分由保留的来源描述综合得出;不是数值化市场份额或流量估算。
[CP015, CP016, CP017, CP018, CP023, CP025]3.2 能力与规模对比
Liblib 的规模指标让它不只是业余图像工具:2026 年中披露数字包括累计用户超过 30 million、原创模型超过 500,000 个、每日图像超过 5 million 张,ARR 超过 US$300 million。这个规模支撑了专业工作室叙事,但不带来能力排他性。SeaArt 具备文本、图像和视频生成、LoRA 训练、换脸、AI 角色、ComfyUI 和放大能力。Tensor.Art 具备模型托管、在线训练、ComfyUI 工作流和 API 服务。Wujie 增加了国内版权登记路径;AituBo 则攻击初学者和低预算人群。最强结论是:Liblib 的优势是中国本地社区、模型和专业需求密度,而不是掌握独有功能清单。简报引用的流量基准在量化使用前,需要一次新的独立流量面板抓取。这也是为什么定价和套餐机制重要:即便 Liblib 拥有更强本地供给,低价竞争者也能赢得探索性使用,并在创作者围绕其界面形成习惯后再向上销售。[CP002, CP003, CP004, CP018, CP020, CP023]
| 平台 / 同业 | 指标 | 有来源支撑的数值 | 解读 | 限制 |
|---|---|---|---|---|
| LiblibAI | 累计用户 / 模型 / 日生成图片 | >30M 用户;>500,000 个原创模型;>5M 张图片 / 天 | 社区和模型供给深度是 Liblib 最强的规模信号 | 公司相关方和媒体披露指标需要管理层确认 |
| LiblibAI | ARR | May 2026 时 ARR >US$300M | 可比公司篮子转向专业创意 SaaS 和应用层 AI 可比公司 | 收入质量、毛利率和客户集中度仍未公开 |
| Tensor.Art | 模型与流量 | >160,000 个模型;July 2024 全球月访问 3.97M | 直接的模型托管社区可争夺模型供给和创作者注意力 | 流量指标是历史数据,需要 2026 更新 |
| Civitai | 定价 / 模型社区 | 免费套餐;付费 US$10/month 起;资料称有数千个 Stable Diffusion 模型 | 全球模型分享替代品试用门槛低,控制力强 | 简报层面的 7.5M 月访问基准未出现在保留来源文本中 |
| SeaArt AI | 定价与功能宽度 | US$4.79-US$75/month 套餐;图像 / 视频 / LoRA / ComfyUI 栈 | 一体化功能直接重叠,挤压 Liblib 的付费创作者价值主张 | 积分 / 体力机制复杂,授权也不够清晰 |
| Wujie AI | 积分 / 会员 | 每日 30 积分;¥100/month 会员;每次生成 ~2 积分 | 本土 RMB 定价工具争夺中国创作者,版权登记场景尤其明显 | 与 Liblib 对比的质量和流量基准未能独立取得 |
| AituBo AI | 入门价格 | 免费导向的图像 / 视频与编辑套件 | 用户升级到专业工具前,可吸收新手和预算型需求 | 规模、留存和起源未在保留来源中验证 |
| Midjourney | 团队 / 专有模型信号 | 官方页面称其为打造高质量模型的 60 人实验室 | 全球专有替代品抬高创作者对质量的预期 | 财务指标归入后续财务 / 估值来源,此处不引用 |
规模指标只使用公开来源文本;保留来源文本未包含 Civitai 与 Liblib 流量基准的面板数据,因此这里把它保留为缺口。
[CP002, CP003, CP014, CP017, CP020, CP024]| 能力 | LiblibAI | SeaArt AI | Tensor.Art | Wujie AI | AituBo AI | Civitai / Midjourney |
|---|---|---|---|---|---|---|
| LoRA / 模型训练 | 是:LoRA 训练和模型市场 | 是:LoRA 训练 | 是:在线模型训练和托管 | 保留来源未显示为核心;有生成能力和提示词 / 参考图库 | 除创作 / 编辑工具外未核实 | Civitai 模型社区强;Midjourney 不提供模型所有权 |
| 视频生成 | 是:LibTV 和 LiblibAI 2.0 视频能力 | 是:文本 / 图像 / 视频生成,并覆盖 Flow 2.0 | 保留来源未显示为主轴 | 保留来源未显示为主轴 | 是:图像 / 视频生成 | Midjourney / Civitai 对比主要聚焦图像工作流 |
| API / 工作流深度 | 有报道提到 API 和面向 ComfyUI 的工作流 | ComfyUI 工作流;在线套件 | ComfyUI 工作流和 API 平台 | 列出 API 接入 | 未核实 | Civitai 可支持模型工作流;Midjourney 由提示词驱动 |
| 市场 / 社区 | 大规模中国模型分享和创作者社区 | 开放图库和创作者变现 | 模型托管、频道、创作者激励 | 创作者广场和提示词 / 参考图社区 | 资料提到社区功能 | Civitai 有创作者计划和挑战赛;Midjourney 更不像市场 |
| 语言 / 中国适配 | 中文平台,具备备案,并承接国内专业需求 | 全球套件;中国适配未由来源验证 | 根植中国的团队和社区 | 仅中文,RMB 定价 | 新手友好;中国适配未由来源验证 | 全球英语优先替代品 |
| 商业 / 权利清晰度 | 备案支撑合规姿态,但内容安全仍是风险 | 商用与授权清晰度被列为需关注项 | 高级 / 商业条款需要确认 | 版权登记钩子有差异化 | 高级商业条款可能需要订阅 | Midjourney 授权更清晰;Civitai 需要逐模型检查许可 |
矩阵把缺少支撑的单元格标为未核实,不推断功能持平;功能覆盖变化很快,应在产品内重新测试。
[CP004, CP007, CP012, CP013, CP015, CP016]功能广度聚类说明,Liblib 的护城河靠密度和本地化,而不是独占 LoRA 或工作流勾选项。
热力图标签概括公开功能证据;缺乏支持的单元格明确标为弱 / 未验证,而不是假定不存在。
[CP004, CP013, CP018, CP023, CP026, CP028]公开指标强调 Liblib 的模型 / 社区深度,但流量刷新和私有留存仍未回答。
KPI 混合了使用量、供给、定价和团队规模指标;它们是就绪度信号,不是单一综合分。
[CP002, CP003, CP017, CP020, CP024, CP026]3.3 护城河耐久性、切换成本与多栖
当中文社区、模型供给、工作流经验和专业 B 端需求相互强化时,Liblib 的护城河最强。已经依赖 Liblib 托管模型、LoRA 训练、工作流模板和中文审核 / 合规基础设施的设计师或短剧工作室,切换成本高于随手试提示词的用户。但证据同样支持多栖判断。评论把 Civitai 描述为掌控与模型所有权选项,把 Midjourney 描述为速度和质量选项,把 SeaArt / Tensor / Wujie / AituBo 描述为重叠的直接工具。这些替代品会削弱 Liblib 对每个用例变现的能力,尤其当团队可以用一个平台探索,用另一个平台产出精修结果,并把本地 Stable Diffusion 流水线留给敏感资产时。务实的尽调问题不是 Liblib 有没有护城河,而是护城河在功能快速趋同时,是否足够深到保护抽成率和利润率。因此,在 Liblib 证明独家资产、更强的中国创作者流动性,或竞争者难以复制的企业工作流数据之前,多栖应被视为默认用户行为。[CP014, CP015, CP016, CP018, CP023, CP026]
| 护城河 / 风险 | 证据 | 含义 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| 聚焦中国的创作者密度 | Liblib 报告 >30M 用户、>500,000 个模型和 >5M 张图片 / 天 | 创作者需要中文模型、审核和工作流时,网络效应有可能成立 | 高 | 核实活跃 MAU、创作者集中度、流失和模型上传分群 |
| 专业 B 端拉力 | 据报道 ARR >US$300M,来自短剧、影视和广告专业制作方 | Liblib 从纯爱好者流量转向有预算的工作流 | 高 | 获取客户分群收入、头部客户集中度,以及分行业留存 |
| 工作流宽度 | LoRA、市场、API、视频和设计智能体拼出更宽的工作室套件 | 相比单功能工具,可能提高切换成本 | 中 | 用 5 个代表性客户任务,对 Tensor.Art 和 SeaArt 跑工作流迁移测试 |
| 功能持平 / 商品化 | SeaArt、Tensor.Art、Wujie、AituBo、Civitai 和平台巨头覆盖主要功能 | 如果买方多栖或把任务转向更便宜工具,定价权可能被削弱 | 中 | 竞品促销后,对标测试 Liblib 的付费转化和毛利率 |
| 平台巨头分发 | Baidu、Alibaba、Tencent 和 ByteDance 把生成能力嵌入云、电商、游戏、社交和创作者栈 | 对部分企业买家,分发可能比独立创意 UX 更重要 | 中 | 按客户分群测试渠道冲突和合作依赖 |
| 内容安全与授权负担 | Skywork 指出 SeaArt 存在漏洞;后续章节来源还显示 Liblib 有单独的 CCTV 相关安全事件 | 社区开放会给任何市场型护城河带来审核和 IP 风险 | 中 | 审查审核 SLA、下架量、标识合规和企业赔偿条款 |
风险严重性按证据加权,不是概率模型;每行都指向具体尽调路径。
[CP002, CP003, CP004, CP007, CP019, CP021]3.4 负面证据与未决尽调项
负面证据具有实质性。Skywork 对 SeaArt 的评测指出其在露骨内容上的内容安全漏洞;Baidu Baike 则记录了 Liblib 自身另一起与 CCTV 有关的内容安全问题。这些事实不能证明 SeaArt 强于 Liblib,但说明争夺开放社区创作的平台都会继承审核、授权和品牌安全负担。Cybernews 从商业风险角度指出 SeaArt 的定价和授权清晰度问题;Recatools 指出 Wujie 工作流密集且仅中文;Civitai / Midjourney 对比则表明,专业用户理性上可能同时保留多种工具。开源和第三方模型生态还意味着,工作流被复制的速度可能快于社区被复制的速度。在承销 US$2 billion+ 估值溢价前,尽调应刷新流量指标,直接测试创作者迁移摩擦,收集企业合同留存数据,并为 6pen、PixAI 等来源不足的对手取得当前证据。[CP007, CP019, CP021, CP027, CP035, CP036]
3.5 图表
04财务
4.1 融资跑道充足,但资本已成为护城河的一部分
从财务上看,Liblib / Evoken 已不再是缺钱的 AI 艺术社区,而是一家按品类领导者融资的后期应用层公司。本章需要的本地融资事实很清楚:2023 年小额天使轮,2024 年和 2025 年初的一组融资,2025 年 10 月 US$130 million Series B,以及 2026 年 6 月约 US$300 million、投后估值超过 US$2 billion 的 Series B+。重要的承销点不只是时间线,而是它说明了资本充足度。管理层称 B+ 资金将用于研发、全球扩张、AI 创意产品能力和产品组合扩建。这给 Evoken 提供了短期生存缓冲;在这个市场里,算力、模型 API 转售、创作者补贴和付费流量可能很快吞掉自生现金。股权结构同样重要:Granite Asia、Tencent、Shunwei、HSG、Gaorong、Ant、CMC、Source Code、Mingshi、Yingce、INCE 以及 Lenovo 相关资本,带来异常广泛的战略和财务资源;陈冕据称仍持有控股权,也让创始人激励高度集中。[CI001, CI002, CI003, CI004, CI005, CI006]
| 日期 | 轮次 | 金额 | 估值 | 领投或具名投资方 |
|---|---|---|---|---|
| 2023-09 | 天使轮 | US$3.5M | ~US$15M | 投资方:GSR/Jinshajiang、Gaorong、Source Code |
| 2024 | 连续三轮 | >US$20M / 数亿 RMB | 据报道年末 >US$500M | Mingshi、Yingce、Shunwei、战略投资方 |
| 2025-02/03 | A+ / 后续轮次 | 数亿 RMB | 未披露 | 公开资料中的 Yingce、Shunwei 等 |
| 2025-10 | Series B 轮 | US$130M | 未披露 | HongShan/Sequoia China、CMC Capital、战略投资方 |
| 2026-06 | Series B+ 轮 | ~US$300M | 投后 >US$2B | Granite Asia、Tencent、Shunwei;HT Investment 和 Times Capital 跟投 |
| 2026-06 | 近期轮次合计 | 最近两轮合计 >US$500M | n/a | HSG、Gaorong、Ant 等老股东加码 |
轮次时间线只使用公开报道;无法精确到月的日期和未披露估值的单元格保留描述,不作推断。
[CI001, CI002, CI004, CI005, CI006, CI007]| 持有人 / 群体 | 角色 | 已披露持股或经济权益 | 财务意义 | 待确认事项 |
|---|---|---|---|---|
| Chen Mian | 创始人 / CEO | 公开资料披露约 73.95% 持股 | 创始人控制与激励高度一致 | 确认 B+ 轮后完全稀释持股 |
| Granite Asia | B+ 轮联合领投方 | 未披露 | 亚洲成长阶段背书;DBS AI 基金关系拓宽资本入口 | 董事会权利与按比例跟投承诺 |
| Tencent | B+ 轮联合领投方 / 战略方 | 未披露 | 可能带来分发、云与生态价值 | 商业附带协议 |
| Shunwei Capital | B+ 轮联合领投方及早期投资方 | 未披露 | 复投方,连接 Xiaomi / Lei Jun 网络 | 累计持股与清算优先权 |
| 投资方:HSG / Sequoia China, Gaorong, Ant | 既有投资者加码 | 未披露 | 早期机构投资者再次验证 | 跟投规模与优先条款 |
| CMC, Source Code, Mingshi, Yingce, INCE, Lenovo 相关资本 | 早期及 B 轮投资者 | 未披露 | 覆盖媒体、应用与产业资本,融资基础更宽 | 完整优先权层级与投资者同意权 |
仅创始人持股公开量化;投资者持股、优先权和董事会权利需要公司股权结构文件确认。
[CI002, CI005, CI009, CI037, CI038, CI039]B+ 轮把融资转化为研发、产品、全球扩张和资金续航能力,而不是近期分红。
美元金额为公开披露的融资轮金额;募集资金用途条目按定性优先级排序,仅为适配瀑布图展项而数值化呈现。
[CI001, CI002, CI003, CI004, CI045]4.2 ARR 规模异常突出;确认收入和收入结构披露仍偏薄
最强的财务事实,是 2024 年收入到 2026 年 ARR 的跃迁。Tencent News 报道 2024 年全年收入为 RMB206 million;多篇 2026 年报道则称,截至 2026 年 5 月 ARR 为 US$300 million,且同比增长超过 30 倍,或超过 3,000%。因此,公开叙事的重点已经不再是需求是否存在,而是收入组合有多持久。收入似乎来自四条相互连接的流:C 端创作者订阅和积分、专业创作者算力消耗、B 端 API 或定制化企业服务,以及 LibTV 的生产工作流变现。LibTV 尤其重要,因为它把公司从图像生成爱好和设计工作流,推入短剧、电影、广告和品牌视频预算。首月单日收入据称超过 US$1 million,5 月收入超过上线首月 13 倍,服务团队接近 1,000 个,这些都说明付费意愿真实。但按收入流的确认口径、留存、退款、积分沉淀和企业合同条款仍未公开。[CI011, CI012, CI013, CI014, CI015, CI016]
| 指标 | 报告值 | 时间 | 财务解读 |
|---|---|---|---|
| 2024 收入 | RMB206M (~US$28.5M) | FY2024 | 2025-26 产品扩张前基数较小 |
| ARR | US$300M | May 2026 | 当前收入 run-rate 的主要锚点 |
| ARR / 集团增长 | >30x 或 >3,000% YoY | May 2026 | 低 2025 基数上的超高速增长 |
| LibTV 首月变现 | 单日收入 >US$1M | Mar 2026 上线后的首月 | AI 视频工作流高付费意愿的证明 |
| LibTV May 爬坡 | 上线月收入的 >13x | May 2026 | 显示 PMF 很快跑通,但需要分群留存验证 |
| 收入预算来源 | 短剧、影视、广告制作方 | 2026 报道 | B 端专业预算改善收入质量 |
ARR 不等于确认收入;公开披露没有拆分订阅续费、积分消耗、API 和企业收入。
[CI011, CI012, CI013, CI014, CI015, CI016]| 收入流 | 机制 | 单位或定价信号 | 质量判断 | 尽调问题 |
|---|---|---|---|---|
| C 端创作者订阅 | 免费增值用户为计算额度和高级功能升级 | 月度或年度会员 / 积分 | 漏斗大,但消费者留存未知 | 按创作者分群看同期群留存和付费转化 |
| 积分 / 算力消耗 | 用户为生成、LoRA 训练和更高负载付费 | 使用积分 / 生成资产 | 直接绑定模型 API 与云成本 | 按生成模态拆毛利率 |
| B 端 API / 企业 | API 服务、定制化、私有部署式项目 | 合同或 API 消耗 | 预算质量更高,但销售效率未验证 | 合同 ACV、回本周期、续约和服务人力 |
| LibTV 视频工作流 | 订阅加按次付费的视频生成 | 据报道有 RMB 订阅档位和按分钟生成定价 | 短剧工作流中的强 PMF 信号 | 竞品价格正常化后的留存 |
| 国际设计智能体 / Lovart | 海外 AI 设计智能体变现 | 外部报道披露 ARR 贡献 | 分散对中国图像社区的依赖 | 分区域 ARR、渠道 CAC 和合规成本 |
收入流由公开产品和融资报道综合而来;实际成交价格和收入确认政策未公开。
[CI018, CI019, CI020, CI021, CI046, CI047]创作者流量靠订阅、点数、API / 企业项目和视频工作流使用转成收入,随后要面对模型供应商侧 COGS。
流向为定性呈现,因为收入结构和 COGS 未公开披露。
[CI017, CI018, CI019, CI020, CI026, CI027]4.3 ARR 倍数不高,但利润率问题才是支点
按头部数字计算,相比全球 AI 创意同业,Evoken 并不显贵:US$2 billion 估值对应 US$300 million ARR,约为 6.7 倍 ARR。这个倍数低于 Canva 约 10.5 倍至 16 倍 ARR 的参考区间,也低于 Sacra 对 Gamma 约 20.6 倍的样本。Midjourney 据称 US$500 million 收入和 US$10 billion 估值,则隐含更高的创意 AI 标杆。这个折价是理性的,因为 Liblib 尚未证明自己掌握毛利结构。China Biz Insider 的负面解读处于核心位置:Evoken 不训练自有基础模型,每一次生成任务都依赖第三方模型提供商,因此公司的经济模型是零售定价与批发模型 API 或算力成本之间的价差。最好的上行情景是,社区资产、工作流和 B 端生产集成创造足够锁定,让公司谈下批量折扣并守住价差。下行情景则是,一方模型厂商降价、打包工作流,或迫使 Liblib 补贴使用以防守份额。[CI026, CI027, CI028, CI029, CI030, CI031]
| 公司 / 参考 | ARR 或收入 | 估值 | 隐含倍数 | 对 Liblib 的相关性 |
|---|---|---|---|---|
| Evoken / Liblib | US$300M ARR | >US$2B | ~6.7x ARR | 当前投资测算锚点 |
| Midjourney | ~US$500M 2025 年收入 | US$10B | 收入约 20.0x | 创意 AI 同类公司,自有产品拉力强 |
| Canva 低参考值 | US$4B ARR | US$42B | ARR 约 10.5x | 已成规模且盈利的设计平台 |
| Canva Sacra 二级市场 | US$4B ARR | US$65B | ARR 约 16.3x | AI 优先设计套件的上行情景 |
| Gamma 示例,来自 Sacra | US$102M ARR | US$2.1B | ARR 20.6x | AI 原生应用溢价基准 |
| Runway / ElevenLabs 同类组 | 引用来源未披露 | 引用来源未披露 | 来源将其列为全球 ARR 同类组,但公开数据不足以计算倍数 |
倍数为近似值,使用公开 ARR/收入快照;空值表示引用来源提到了该同类组,但缺少 ARR 和估值两项输入。
[CI030, CI031, CI032, CI033, CI034, CI035]公开 ARR 支撑的是相对创意 AI 可比公司的低倍数;上行空间取决于能否证明毛利可持续。
Evoken 的低 / 高档是投资测算情景;Canva 和 Gamma/Suno 区间来自已引用的可比公司报告。
[CI030, CI033, CI034, CI042, CI043]最高增长的收入流也最暴露于批发模型定价和补贴决策。
矩阵是尽调框架;具体贡献毛利未公开。
[CI018, CI019, CI020, CI027, CI028, CI029]4.4 财务结论:真实 PMF,未验证的贡献利润率
财务结论偏建设性,但有条件。Liblib 罕见地同时拥有公开 ARR 规模、近期增长、多产品扩张和蓝筹投资方联盟;如果专业内容预算证据成立,这些就不是虚荣信号。B+ 轮应覆盖短期产品和全球扩张需求;相对较低的 ARR 倍数也留下上行空间,前提是 Evoken 证明自己正在成为创作者操作系统,而不是薄利模型转售商。未决尽调项也足以决定投资判断。买方需要获得按产品划分的实际毛利、批发模型 API 价格表、云承诺、队列留存、企业收入占比、积分负债、月度烧钱、投后现金和跑道。监管连续性也属于财务章节,因为 CAC 备案状态和内容合规失败会改变可用性、付费转化和企业采购。在这些私有输入被验证前,本章应把收入质量承销为强,把利润率质量承销为未验证,把融资依赖承销为已改善但仍具战略重要性。[CI022, CI023, CI024, CI025, CI043, CI044]
4.5 图表
05产品与技术
5.1 产品组合已从图像社区扩展为创作者操作系统
应把 Liblib / Evoken 评估为一家多产品创意工作流公司,而不只是一个 AI 图像网站。旗舰产品 LiblibAI 界面仍是核心流量和资产层:公开报道提到累计用户超过 30 million,原创模型超过 500,000 个,每日生成图像超过 5 million 张。LiblibAI 2.0 随后通过加入集成式图像和视频生成、头部视频模型、特效模板,以及更广泛的开放 + 闭源模型菜单,把这个社区重塑为专业创作工作室。相邻产品把同一资产底座延伸到新预算:LibTV 面向视频和短剧团队,Xingliu 面向国内设计 Agent 工作流,Lovart 则把设计 Agent 论点带到国际市场。因此,产品组合逻辑是顺序展开的:创作者社区创造资产,工作室工作流提高频次,Agent / 视频产品变现专业制作场景。这个逻辑对尽调很重要,因为只有这些工作流界面转化为反复付费任务,而不是被促销算力或暂时便宜的上游模型接入补贴的一次性实验,产品承诺才持久。[CE001, CE002, CE003, CE004, CE005, CE006]
| 产品 | 上线 / 状态 | 核心功能 | 规模证据 | 尽调缺口 |
|---|---|---|---|---|
| LiblibAI | 2023 年上线;核心产品 | 图像社区、模型市场、创作工作室 | >30M 用户;50 万+ 原创模型;公开报道 >5M 张图像/日 | 核实活跃创作者、付费转化和内容审核事故整改情况 |
| LiblibAI 2.0 | 2025 年末升级 | 整合图像 / 视频、模型、特效和模板的专业工作室 | 上线报道称 20M+ 创作者;报道称 500+ 视觉特效 | 验证免费算力促销后的留存使用 |
| LibTV | 2026 年 3 月上线 | 面向短剧、影视、广告团队的 AI 视频创作 | 首月单日收入 >$1M;报道称近 1,000 个团队 | 审查 Seedance/API 合同与贡献毛利 |
| Xingliu 星流 | 国内设计智能体 | 面向中国用户的智能体设计工作流 | 报道称 >10M 用户 | 拆分 MAU、留存和与 LiblibAI 用户的重叠 |
| Lovart | 国际设计智能体 | 面向海外的 AI 设计智能体 | China Biz Insider 报道约 $80M ARR | 验证地域分布、支付、流失率和模型供应商组合 |
规模指标来自 2026 年中公开第三方报道;私有留存、用户重叠和付费转化分母未披露。
[CE001, CE002, CE003, CE004, CE005, CE006]产品组合覆盖社区、创作台、视频和 Agent 能力,其中自有基础模型能力明确偏弱。
能力标签是基于公开来源的定性判断;自有基础模型为弱的单元格反映反向报道,而不是技术基准测试。
[CE006, CE007, CE010, CE011, CE012, CE034]5.2 技术中心是围绕扩散、LoRA 和 ComfyUI 式图构建的工作流深度
最强的公开技术证据显示,Liblib 的应用栈围绕成熟扩散模型和工作流编排组装,而不是围绕自有基础模型。在模型层,留存技术来源覆盖 Stable Diffusion、SDXL、Stable Diffusion 3.5、FLUX.1、ControlNet、image-to-image 和 LoRA。在工作流层,ComfyUI 很关键,因为它把生成变成显式节点图:checkpoint 加载、prompt 编码、采样、VAE 解码、后处理和自定义节点都可以替换、版本化,并通过 API 提交。这种架构与 LiblibAI 作为模型市场和创作工作室的公开定位吻合;创作者可以使用专门的中文风格模型资产,把基础 checkpoint 与 LoRA adapter 组合起来,并打包可复用工作流。证据支持工作流成熟度,但不能证明公司拥有底层基础模型或差异化推理基础设施。这些细节决定实际技术杠杆。[CE012, CE014, CE015, CE016, CE018, CE019]
| 能力 | 证据是否支持? | 证据基础 | 技术说明 |
|---|---|---|---|
| Stable Diffusion 生态 | 是 | 资料称 LiblibAI 是 Stable Diffusion 枢纽;DataCamp 解释 SD 基础模型 | 市场可托管微调后的 SD 风格检查点和适配器 |
| SDXL / SD3.5 / FLUX 工作流 | 生态支持层面为是 | Tech-Insider 在 ComfyUI 工作流中覆盖 SDXL、SD3.5 Large 和 FLUX.1 | 证据支持工具链兼容,不支持 Evoken 拥有模型 |
| ControlNet | 是 | GPTProto 和 Tech-Insider 在生成工作流中提到 ControlNet | 姿态、深度、边缘控制的生产级输出离不开它 |
| LoRA 微调 / 训练 | 是 | Apatero、RunComfy、Stable Diffusion Art、Hubpy 和 GPTProto 提到 LoRA 训练 / 使用 | 支撑角色 / 风格一致性和社区模型市场 |
| ComfyUI 节点工作流 | 是 | Runflow 和 Tech-Insider 描述工作流 JSON、节点、API 和 DAG 执行 | 可作为 Liblib 工作流深度和 API 集成的类比 |
| 视频模型聚合 | 是,但有依赖 | AIbase、Yicai 和 China Biz Insider 提到 LibTV;China Biz Insider 提到通过 Volcano Engine 使用 Seedance | 拉宽能力边界,但暴露模型访问和批发成本风险 |
行项目把 Liblib 直接证据与技术生态证据合并使用;能力获得支持并不意味着 Evoken 拥有底层基础模型。
[CE014, CE015, CE016, CE018, CE019, CE020]| 工作流步骤 | 公开规格 | 重要性 | 注意事项 |
|---|---|---|---|
| 数据集收集 | Apatero 指南中,角色 LoRA 需要 15–50 张高质量图像 | 样本足以支撑角色 / 风格反复使用 | Liblib 专属训练限制未公开 |
| 图像打标 | 每张图都应配标注文本 | 标注把视觉概念绑定到提示词 | 标注质量控制未披露 |
| 训练运行 | Apatero 指南中,合适硬件上为 1,000–3,000 步;耗时 1–4 小时 | 给创作者训练提供粗略算力 / 时间预期 | Liblib 实际云硬件和队列优先级未知 |
| ComfyUI 扩展 | Lora-Training-in-Comfy 增加训练节点和参数 | 展示训练如何嵌入节点工作流 | 扩展支持是生态证据,不是 Liblib 官方文档 |
| 模型可用性 | RunComfy 将模型直接保存到 ComfyUI LoRA 文件夹 | 快速训练-测试迭代支撑市场供给 | Liblib 审核 / 批准流程未披露 |
| 适配器使用 | Stable Diffusion Art 将 LoRA 描述为与基础检查点搭配使用的小型适配器 | 不用存储庞大检查点,也能启用多种风格 | 质量取决于基础模型和训练数据权利 |
规格是 Liblib 宣传 LoRA 工作流的技术文档代理;Liblib 准确训练配额、价格和审核检查需要私下尽调。
[CE024, CE025, CE026, CE027, CE028, CE029]Liblib 的公开架构更适合理解为叠在第三方和开源模型生态之上的分层工作流基础设施。
分层来自公开产品和技术证据的综合归纳;并不主张 Evoken 拥有每个基础设施组件。
[CE005, CE008, CE018, CE019, CE020, CE030]生产任务从提示词和参考素材出发,经过模型选择、节点执行、安全检查,最后交付资产。
该流程是基于公开证据的通用工作流;Liblib 内部队列、审核和存储实现未披露。
[CE015, CE016, CE021, CE022, CE023, CE032]LoRA 训练生成可复用适配器,可测试、发布,并与基础 checkpoint 组合。
训练规格来自公开 ComfyUI/LoRA 指南,不是 Liblib 私有配额。
[CE024, CE025, CE026, CE027, CE028, CE029]5.3 开发者界面存在,但集成证据薄于产品牵引
开发者信号来源表明,LiblibAI 可以被视为可编程生成入口,但公开证据还不足以等同于一个成熟企业 API 计划;后者应有透明定价、SLA、SDK 和安全文档。GPTProto 描述了一个 LiblibAI API 生态,开放文生图、图生图、ControlNet、ADetailer、LoRA 训练、ComfyUI 在线和复杂节点工作流。Runflow 的 ComfyUI 指南展示了这类后端通常遵循的模式:客户端向 /prompt 提交工作流 JSON,通过 WebSocket 或历史端点观察执行,上传输入并取回生成文件。这对需要可重复生产流水线的工作室有用,但尽调缺口真实存在。公开来源没有展示 Evoken 的 API 速率限制、客户认证模型、数据留存政策、可用性历史或模型提供商透传条款;这些私有控制项决定开发者接入是一条可扩展产品线,还是高级用户便利功能。[CE016, CE017, CE021, CE022, CE023, CE043]
| 功能 | 公开证据呈现方式 | 主要使用场景 | 风险或尽调问题 |
|---|---|---|---|
| 工作流提交 | ComfyUI 风格 /prompt 接受可执行的工作流 JSON | 程序化图像 / 视频生成任务 | 确认 Liblib 端点形态、认证和速率限制 |
| 执行跟踪 | Runflow 指南中的 WebSocket 和 /history 模式 | 队列监控和生产任务状态 | 索取可用性、重试和可观测性记录 |
| 输入 / 输出处理 | ComfyUI API 指南中的 upload/image 和 /view 模式 | 参考图、蒙版和生成资产 | 检查数据留存和版权控制 |
| ComfyUI 在线 / 节点工作流 | GPTProto 称 Liblib API 包含 ComfyUI 在线工作流 | 高级工作室和集成场景 | 验证工作流是否可导出 / 可版本化 |
| 区域访问 | GPTProto 称完整访问可能需要中国手机号注册 | 合规和身份验证 | 评估全球开发者采用阻力 |
开发者功能根据 Liblib 开发者信号和 ComfyUI API 模式推断;企业 API 合同和 SLA 未公开。
[CE016, CE017, CE021, CE022, CE023, CE043]5.4 差异化来自本地化工作流密度;风险是第三方模型依赖
建设性技术论点是,Liblib 已积累一个本地化中国创作者生态:创作者上传、训练和分享 LoRA 模型与工作流,普通用户则无需拥有 GPU 或掌握本地工具,就能复用这些资源。这可能比单一提示词界面更可防守,因为社区资产、风格库、工作流模板和生产习惯都会制造切换成本。负面论点同样重要。China Biz Insider 报道称,Evoken 不开发自有基础模型,LibTV 使用第三方模型 API,包括通过 Volcano Engine 接入 Seedance 2.0。如果模型所有者降低直连价格、限制访问或打包一方工作流应用,Evoken 的毛利和差异化可能迅速被压缩。合规是另一项产品风险:CAC 备案背景和已报道的 CCTV 内容安全事件说明,审核和标识不是事后事项。承销结论是:Liblib 具有牵引力很强的工作流应用护城河,但仍需要对模型合同、安全整改、API 控制和按产品线划分的贡献利润率做私有尽调。[CE013, CE032, CE033, CE034, CE035, CE036]
| 组件 | 供应商 / 来源 | 依赖类型 | 风险 |
|---|---|---|---|
| 基础图像 / 视频模型 | Stability 生态、Black Forest Labs FLUX、ByteDance Seedance/Volcano Engine | 第三方模型 / API 访问 | 访问限制或直接产品捆绑可能削弱 Liblib 护城河 |
| LibTV 视频生成 | China Biz Insider 报道通过 Volcano Engine 接入 ByteDance Seedance 2.0 | 第三方视频模型 API | 批发定价变化会挤压利润率 |
| 云 GPU / 推理容量 | 未披露;生态需要 GPU 推理 | 基础设施和算力采购 | 公开来源无法证明成本、可用性或容量优势 |
| ComfyUI / 开放工作流生态 | 开源和第三方节点生态 | 工作流引擎和开发者工具 | 默认 ComfyUI 缺少内置认证;必须验证生产级加固 |
| 用户训练的 LoRA / 模型资产 | 创作者社区和市场 | 用户生成模型供给 | IP、审核和质量控制需要证据 |
| AIGC 合规控制 | CAC 备案环境和平台审核 | 监管和信任层 | CCTV 报道的漏洞显示,对抗性提示词下控制会失效 |
本表有意强调依赖;这不是完整供应商清单,因为 Evoken 未公开披露模型供应商合同或云支出。
[CE008, CE012, CE023, CE032, CE033, CE034]5.5 图表
06客户
6.1 客户基础:大众创作者社区输入 B 端工作流
Liblib 的客户基础应按双边市场系统承销,而不是按纯 SaaS 账户名单承销。在 C 端,公开报道提到截至 2026 年中累计用户超过 30 million、2025 年末总用户 25 million 且 MAU 约 4 million、AI 创作者超过 20 million、原创模型超过 500,000 个、每日生成图像超过 5 million 张。这创造了一个庞大的供给侧,包含提示词用户、LoRA 训练者、设计师、插画师、自媒体创作者和专业视觉工作者。在 B 端,经过验证的公开证据更窄,但战略上重要:Jiemian 称 LiblibAI 为 Kingsoft Office / WPS、Wondershare、G-bits Games、Tmall Campus 和 Tsinghua University 提供专业 AI 图像场景解决方案。因此,最强的细分视角是:C 端创作者贡献流动性和数据,专业创作者贡献订阅和积分,B 端工作室或企业贡献 API、定制化服务和更高价值的工作流支出。[CU001, CU002, CU003, CU004, CU005, CU006]
| 指标 | 值 | 日期 / 版本 | 客户意义 | 置信度 |
|---|---|---|---|---|
| 累计用户 | >30M | 2026 年中 | 大规模漏斗顶部创作者 / 社区触达 | 高 |
| 注册 / 总用户 | 25M | 2025 年末 | 2026 年规模跃升前的早期基线 | 中 |
| 月活跃用户 | ~4M | 2025 年末 | 活跃受众代理,不是留存队列 | 中 |
| AI 创作者 | >20M 创作者;更广生态 >30M 用户 | 2025-2026 | 模型、提示词、工作流和内容供给 | 中 |
| 原创模型 | >500,000 | 2026 年中 | 资产深度和创作者流动性 | 中 |
| 生成图像 | >5M/日;Jiemian 来源称累计 >500M | 2025-2026 | 重复使用强度代理 | 中 |
数值为公开报道指标,时间口径混杂;MAU 和注册用户数据为 2025 年末,累计用户和模型指标为 2026 年中。
[CU001, CU002, CU003, CU004, CU005, CU006]| 客群 | 规模 / 证据 | 需求 | 变现 | 主要缺口 |
|---|---|---|---|---|
| 泛创作者与准专业创作者 | >30M 累计用户;每日免费生成可用 | 低摩擦图像生成、灵感获取、提示词迭代 | 免费层转向积分 / 订阅 | 付费转化和流失率未披露 |
| 模型训练者 / 工作流搭建者 | 报道称 >500,000 个原创模型 / 工作流 | LoRA 训练、模型分享、工作流复用 | 订阅、积分、创作者激励和市场效应 | 创作者供给集中度和分成经济性未公开 |
| 专业设计师 / 机构 | “每三名中国设计师就有一名使用”的说法;LibTV 品牌客户信号 | 稳定资产、视频 / 图像工作流、团队协作 | 专业版方案、积分和 LibTV 工作流支出 | 席位数和 NRR 不可得 |
| 企业 / API / 定制解决方案买家 | Jiemian 点名五家 B 端客户;报道提到 API / 定制需求 | 可靠 AI 图像场景、集成、合规和支持 | API、定制服务、可能的私有化部署 | 生产状态、SLA 和合同金额未披露 |
| 短剧 / 影视 / 广告团队 | 300+ 家影视公司接入 LibTV;服务近 1,000 个团队 | 剧本到分镜再到视频的流水线,降低制作成本 | 订阅叠加按量付费,或团队工作流支出 | 留存和输出质量阈值未知 |
客群规模并不互斥;各行把公开规模事实与基于已披露定价和产品界面推断的变现方式合并。
[CU007, CU015, CU017, CU028, CU029, CU036]Liblib 把社区发现转成模型使用、付费算力和企业工作流扩张。
旅程节点综合公开产品和变现证据;转化率未披露。
[CU007, CU028, CU029, CU034, CU035, CU045]6.2 具名客户真实存在,但 LibTV 和垂直预算是更强采用信号
具名客户名单重要,因为它满足了客户证明标准,但单靠名单不能证明经常性生产部署或具体结果。更有说服力的 B 端证据来自游戏、短剧、电影和广告周边的垂直预算。LibTV 据称从上线到首月单日收入超过 US$1 million,5 月收入超过首月收入 13 倍,一个月内接入 300 多家短片和电影公司,服务近 1,000 家代理商、制作公司和品牌客户。游戏带来第二个大型预算池:miHoYo 的 RMB100 billion AI 承诺、NetEase 和 Tencent 的 AI 生产流水线,以及中国游戏开发者广泛采用生成式 AI,都支持美术、资产、NPC 和 UGC 工作流正在成为采购优先事项这个论点。这些事实意味着真实需求存在,但合同规模、生产状态和逐客户结果仍属私有信息。[CU009, CU010, CU011, CU012, CU013, CU015]
| 客户 | 行业 | 公开使用场景推断 | 引用质量 |
|---|---|---|---|
| Kingsoft Office / WPS (金山办公) | 生产力软件 | 办公 / 生产力视觉工作流中的 AI 图像场景解决方案 | Jiemian 点名;生产范围和结果未披露 |
| Wondershare (万兴科技) | 创意软件 / 视频 | AI 图像或创意资产工作流,与 Wondershare 创意工具组合匹配 | Jiemian 点名;生产范围和结果未披露 |
| G-bits Games (吉比特) | 游戏 | 游戏美术、资产或营销图工作流;行业来源佐证游戏 AI 需求 | Jiemian 点名;使用场景按行业推断 |
| Tmall Campus (天猫校园) | 电商 / 校园商业 | 商业场景下的商品、活动或校园市场图像 | Jiemian 点名;合同状态未披露 |
| Tsinghua University (清华大学) | 教育 / 研究 | AI 创作或教育 / 研究社区视觉工作流 | Jiemian 点名;付款方和部署状态未披露 |
枚举表只是公开具名 B 端客户样本,不是完整客户名单;来源仅点名客户时,行业使用场景采取保守推断。
[CU008, CU009, CU010, CU011, CU012, CU013]| 垂直领域 | 需求驱动 | Liblib / Evoken 产品匹配度 | 证据信号 | 风险 |
|---|---|---|---|---|
| 游戏 | 高资产成本、UGC、NPC 和剧情制作需求 | LiblibAI 图像资产、模型工作流、API;潜在游戏美术管线 | miHoYo RMB100B AI 投入;NetEase / Tencent AI 部署 | 大型工作室可能自建工具 |
| 短剧 / 影视 | 微短剧生产承受数量和成本压力 | LibTV 无限画布视频工作流、分镜和团队生产 | 已接入 300+ 家公司;服务近 1,000 个团队 | 内容过剩和劳工反弹 |
| 电商 | 需要高频产出商品图、活动图和本地化图片 | AI 图像场景方案和模板工作流 | Tmall Campus 被点名为 B 端客户 | 效果和转化提升未公开 |
| 广告 / 品牌视频 | 需要快速出概念、变体和活动素材 | 图像 / 视频生成、特效模板和 API / 定制服务 | 广告专业制作人被提为核心收入来源 | 品牌安全以及肖像 / IP 筛查 |
| 教育 / 专业设计 | 培训、试验和设计工作流 | LiblibAI 社区、模型和创作者激励 | Tsinghua 被点名;“三分之一设计师”说法 | 付费方身份和席位留存未公开 |
垂直领域映射把已点名客户证据、LibTV 牵引力和独立行业需求来源合在一起;表格不意味着每个垂直领域的变现程度相同。
[CU019, CU020, CU021, CU022, CU023, CU024]最强证据来自创作者规模和 LibTV 团队采用;具名客户还需要更深部署证明。
序数评级总结公开证据质量;不代表经审计的客户评分。
[CU008, CU017, CU020, CU023, CU024, CU026]6.3 获客路径:免费增值激活,转向积分、订阅和企业 / API 支出
可观察的市场进入动作始于低摩擦社区创作,而不是自上而下的企业销售。免费每日生成、创作者激励和模型浏览把用户带入生态;高频使用、优先 GPU 访问、训练和专业功能再转化为订阅或积分;API 和定制化企业需求则把 B 端客户推向更高价值合同。这解释了为什么同一家公司既能展示创作者社区指标,又能展示 US$300 million ARR:创作者基础提供内容、工作流、模型和分发,专业生产者和企业则提供付费意愿。扩张依赖跨产品循环。创作者或工作室可以从 LiblibAI 上的图像模型开始,迁移到 LibTV 的视频工作流,随后购买 API 或定制服务。尽调问题在于,这个循环有多少是经常性且可防守的,多少是补贴驱动的;尤其 API 设置摩擦和第三方模型替代品可能鼓励多栖。[CU019, CU028, CU029, CU030, CU031, CU032]
| 销售动作 / 层级 | 用户或买方 | 公开证据 | 变现逻辑 | 尽调要问 |
|---|---|---|---|---|
| 免费社区入口 | 轻量创作者和模型浏览用户 | 免费每日算力、模型浏览和创作者激励 | 冷启动流动性,养成使用习惯 | 按 cohort 拆分的免费转付费率 |
| 订阅 | 专业创作者和设计师 | 高用量和优先 GPU 访问需要订阅或点数 | 经常性套餐收入和可预测算力分配 | 分套餐毛利率和流失 |
| 点数 / 按需算力 | 重度创作者和工作室 | Tencent News 称采用订阅加按需付费模式 | 生成和训练按用量变现 | 点数沉淀、退款和模型批发成本 |
| 企业 / API / 定制服务 | 公司、工作室和软件平台 | 企业客户为 API 和定制需求付费;API 能力已有公开讨论 | 更高 ACV、集成和工作流锁定 | SLA、私有化部署条款和采购周期 |
| LibTV 团队工作流 | 影视、短剧、代理商和品牌团队 | 单日收入 >US$1M;300+ 家公司和近 1,000 个团队 | 团队订阅叠加生产管线用量 | 上线补贴后的留存和团队扩张 |
定价层级由公开报道和评测重构;精确价目表、企业折扣和私有化部署条款未披露。
[CU016, CU017, CU018, CU028, CU029, CU030]从免费创作者入口到企业 / 定制服务扩张的方向性 GTM 漏斗。
数值是方向性指数,不是报告的转化率百分比。
[CU017, CU018, CU028, CU030, CU033, CU044]6.4 耐久性仍未验证:留存、集中度和声誉风险是缺口
客户章节的建设性结论是需求存在;谨慎点在于公开留存证据不存在。留存来源没有披露 NRR、GRR、流失率、续约率、合同期限、头部客户占比、企业销售管线转化或满意度。每日图像生成、创作者激励、模型供给和 LibTV 团队采用都是有用的重复使用代理指标,但它们不是队列留存。需求侧风险也不是理论问题。Tech Times 报道了工业级 AI 短剧热潮:每天 470 部 AI 制作作品、内容过剩、真实工人被替代、写作项目取消,以及未经授权肖像问题。这会产生两条可能的需求路径:更多工作室购买 AI 工具来竞争产量,或者品牌和机构因为 AI 内容变得声誉敏感而放慢采用。因此,投资尽调应取得按细分市场划分的队列留存、企业收入集中度、合同续约数据、API / 私有化部署条款和品牌安全采购反对意见,然后再把 Liblib 客户基础视为持久。务实测试是:免费算力、创作者奖励和发布期宣传正常化后,使用是否仍能留下。没有这些证据,ARR 应被视为强需求证明,但尚不能证明低流失的客户耐久性。[CU024, CU025, CU026, CU027, CU037, CU038]
0–100 的示意性队列视角,用来框定创作者和收入扩张循环缺失的留存数据。
留存百分比是来自公开重复使用代理指标的示意占位;实际队列未披露,仍是尽调缺口。
[CU001, CU002, CU003, CU004, CU031, CU032]6.5 图表
07风险
7.1 中国 AIGC 监管已是操作系统,不是上线清单
Liblib 的第一层风险是监管连续性。中国如今不是靠单一法律监管面向公众的生成式 AI,而是靠一组工具叠加:2023 年 1 月《深度合成管理规定》、2023 年 8 月《生成式人工智能服务管理暂行办法》、通过 CAC 和地方网信部门管理的备案与登记名册,以及 2025 年 9 月 AIGC 标识办法和 GB 45438-2025。对 Liblib 来说,这套监管栈直接映射到图像、视频、LoRA 训练、下载、用户上传和分发工作流。有利事实是,36Kr 报道 Liblib 在 2024 年 2 月通过深度合成算法备案,并在 2024 年 3 月成为首个完成生成式 AI 服务备案的 AI 社区平台。负面承销点是,备案必要但不充分:标识文件、元数据、用户声明、事件响应、未成年人保护和周期性名册可见性,都必须跟上 LibTV、Xingliu、Lovart 以及跨境工作流的产品扩张。[CR001, CR002, CR003, CR004, CR005, CR006]
| 法规或案件 | 日期 | 主管机关或法院 | 与 Liblib 的相关性 |
|---|---|---|---|
| 《深度合成规定》 | Jan 2023 起生效 | CAC 及有关主管机关 | 适用于合成图像 / 视频服务;要求透明度、标识、用户控制和深度合成服务备案。 |
| 《生成式 AI 服务暂行办法》 | Aug 15 2023 起生效 | CAC 等七部门 | 面向公众的生成式 AI 服务核心制度;支撑备案、数据合法性、内容安全和用户责任。 |
| Liblib 深度合成算法备案 | Feb 2024 | CAC 备案流程 | 36Kr 报道,Liblib 通过第四批深度合成服务算法备案。 |
| Liblib 生成式 AI 服务备案 | Mar 2024 | CAC / 地方网信备案流程 | 36Kr 报道,Liblib 成为中国首个完成《暂行办法》备案的 AI 社区平台。 |
| AIGC 标识办法与 GB 45438-2025 | Sept 1 2025 起生效 | 监管机构:CAC、MIIT、MPS、NRTA、TC260 / SAMR | 要求显式和隐式标识、元数据或水印控制、用户提示以及应用商店材料核验。 |
| CAC 生成式 AI 备案登记公告 | Nov 11 2025 | 国家网信办 | 体现持续备案预期:截至 Nov 1 2025,已有 611 项服务备案、306 个应用 / 功能登记。 |
| 综合性 AI 法草案 | Dec 2025 发布官方草案;能否出台尚不明确 | Deep Lex 跟踪的国家立法进程 | 未来横向 AI 义务可能超出行业 AIGC 规则,带来不确定性。 |
| 北京互联网法院 AI 图像版权裁判脉络 | 2023 年判决和 2025 年后续案件 | 北京互联网法院 | 有证据证明人类创作投入时,AI 图像可以产生版权;流程记录薄弱会导致主张失败。 |
| Ultraman v Acgnai LoRA 纠纷 | 2024-2025 | 杭州互联网法院和杭州市中级人民法院 | AIGC 平台因稳定产出侵权 LoRA 结果承担帮助侵权责任;赔偿 RMB 30,000 并停止侵害。 |
| 广州互联网法院 Ultraman AIGC 案 | Feb 2024 | 广州互联网法院 | 更早释放 AIGC 服务提供者责任信号;强调技术措施、投诉举报、风险提示和标识。 |
截至 runDate,表格摘录与图像、视频、LoRA 和社区平台风险最相关的中国 AIGC 法规与判例。
[CR001, CR002, CR003, CR004, CR005, CR006]| 义务 | Liblib 状态证据 | 剩余风险敞口 | 尽调要问 |
|---|---|---|---|
| 深度合成算法备案 | 36Kr 报道,Feb 2024 备案已通过。 | 新功能和模型接入可能改变备案范围。 | 获取备案证书、算法名称、范围和更新历史。 |
| 生成式 AI 服务备案 | 36Kr 报道,Mar 2024《暂行办法》备案已完成。 | 公开登记展示和地方注册可能需要更新。 | 确认当前 CAC 登记条目和产品映射。 |
| 显式和隐式标识 | 制度覆盖生成图像、音频、视频、虚拟场景、下载和导出。 | 具体产品的标识落地无法公开审计。 | 审查 LiblibAI、LibTV、Xingliu 和 Lovart 的水印 / 元数据测试。 |
| 内容安全审核 | CCTV / Baidu Baike 报道过涉黄生成绕过和公司整改。 | 面对混淆提示词时,控制能否长期有效仍未知。 | 委托第三方红队测试,并审查事故复盘。 |
| IP 投诉和下架控制 | Ultraman 系列案件要求有效预防和及时响应。 | 公开证据未披露 SLA、积压量或重复侵权者指标。 | 检查通知日志、下架 SLA、申诉流程和模型审核工具。 |
| 跨境合规 | Linklaters 提示全球 AIGC 标识规则分化、跨境成本上升。 | Lovart / 全球扩张可能面对中国以外的标识和 IP 制度。 | 梳理各司法辖区发布闸门、数据流和模型提供方承诺。 |
状态仅反映公开证据;关闭尽调前需要私有证书、产品测试和合规日志。
[CR004, CR005, CR009, CR010, CR011, CR012]备案、标识、审核、IP 控制和登记更新共同影响产品可用性和企业采购。
依赖关系图综合法律义务和运营控制;它不是官方监管流程图。
[CR006, CR008, CR009, CR010, CR011, CR014]7.2 最大法律尾部风险不是罚款,而是创作者工作流中断
最难承销的风险,是用户生成模型、生成产出和内容审核的组合。CCTV 在 2026 年曝光 LiblibAI 可被提示词绕过过滤器生成色情或不当内容,这一点直接相关,因为中国 AIGC 监管强调执法,而且越来越以标识为中心。公司据称进行了技术整改,但投资者需要证据证明修复能在对抗性提示词下保持稳定,而不只是被动应对。版权风险同样具有工作流特征。Ultraman v Acgnai 并没有把所有训练行为都定为违法;杭州法院对输入阶段训练相对宽容,把责任重点放在稳定侵权产出、被推广的 LoRA 模型、平台获利、侵权明显性和未采取预防措施上。这正是为什么 Liblib 的模型分享和 LoRA 市场需要强 IP 审查、投诉渠道、模型屏蔽、标识和下架运营。赔偿金额只有 RMB 30,000,但禁令、下架或企业采购受阻会重要得多。[CR012, CR013, CR014, CR015, CR016, CR017]
| 先例 | 裁判要点或启示 | 对 Liblib 的影响 | 关键缓释 |
|---|---|---|---|
| Li v Liu,北京互联网法院 | 只要人类提示词和参数选择体现智力成果,AI 生成图像可以受版权保护。 | 用户保留流程证据时,产出可能受保护,进而影响市场所有权和纠纷。 | 提示词历史、生成记录、产出溯源和用户披露。 |
| 2025 年北京互联网法院猫咪吊坠案 | 创作者缺少原创流程记录、依赖事后模拟,主张因此失败。 | Liblib 应支持创作者追溯流程,而不是依赖重建的提示词叙述。 | 可导出生成日志、版本历史和审计轨迹。 |
| Ultraman v Acgnai,杭州 | 训练环节可能较宽松,但稳定侵权输出和平台推广会触发帮助侵权责任。 | LoRA 市场推广、复用或变现知名 IP 模型时风险暴露。 | IP 检测、模型下架、通知—删除、风险提示和封面 / 示例审核。 |
| 广州互联网法院 Ultraman 案 | AIGC 服务提供者被责令停止侵权生成,并赔偿 RMB 10,000。 | 即便是服务层提供者,也可能负有防止高度相似输出的义务。 | 关键词 / 模型过滤、投诉举报、用户警示和生成内容标识。 |
| EU IP Helpdesk 对 Ultraman 的分析 | 盈利性、架构和防止侵权能力越强,注意义务越高。 | Liblib 通过订阅、点数和企业用途变现后,法院可能期待更高控制。 | 商业风险分层、前置预防和快速整改证据。 |
这些案件不是普通法意义上的正式约束性先例,但对平台注意义务的承销判断释放了很强的司法信号。
[CR015, CR016, CR017, CR018, CR019, CR020]一次审核、IP 或模型访问事件,可能从产品中断一路传导到收入质量和估值倍数压缩。
因果图是基于所引法律、内容和业务证据搭建的投资风险模型。
[CR012, CR020, CR021, CR027, CR031, CR035]7.3 只有工作流锁定跑赢模型商品化,聚合器经济学才可投资
Evoken 的业务风险不是需求:公开来源提到截至 2026 年 5 月用户超过 30 million、原创模型 500,000 个、ARR 为 US$300 million。问题在于这些用户和工作流能否创造足够的利润率控制。China Biz Insider 的负面框架处于核心位置:Evoken 不拥有专有基础模型,依赖第三方 API、模型提供商以及云或算力经济学。这让公司暴露于重新定价、接入降级、一方打包和直接替代;一旦模型厂商进入工作流层,风险会放大。LibTV 在 Jimeng 价格变化后增长,说明 Evoken 能快速套利用户痛点,但也让排队时间和 API 成本表成为论点变量。尚未盈利状态和一轮大额 B+ 融资意味着,资本仍是竞争装置的一部分。核心尽调要求是按模型家族划分的产品级毛利,而不只是 ARR,因为模型转售商可能在摧毁贡献利润率的同时做大收入。[CR026, CR027, CR029, CR030, CR031, CR032]
| 风险 | 可能性 | 影响 | 缓释措施 |
|---|---|---|---|
| CCTV 审核绕过报道后的内容安全执法 | 中高 | 高 | 红队提示词测试、未成年人控制、审计日志、应用商店证据和重复整改核验。 |
| 用户 LoRA 模型和侵权输出引发的版权责任 | 中高 | 高 | IP 审查、举报渠道、模型屏蔽、下架 SLA、水印 / 标识和重复侵权者控制。 |
| 模型提供方重新定价或 API 限制 | 中 | 高 | 多提供方采购、批量合同、优雅降级、自有工作流资产和客户切换成本数据。 |
| 算力转售和补贴压缩毛利率 | 中高 | 高 | 产品级毛利率披露、批发成本表、用量上限和按客户分群定价测试。 |
| 产品扩张时标识或备案监管遗漏 | 中 | 高 | 合规负责人、发布检查清单、标识测试证据和持续 CAC 备案登记监测。 |
| 围绕 Chen Mian 的创始人 / 关键人集中风险 | 中 | 中高 | 董事会控制、继任计划、下放产品领导权和保留事项权利。 |
| 地缘政治 / 出口或跨境标识分歧 | 中 | 中 | 分司法辖区发布闸门、模型提供方法律承诺和数据 / 本地化审查。 |
| 短剧中的劳工替代反弹和内容过剩 | 中 | 中 | 创作者受益叙事、企业治理、质量控制和客户 ROI 证明。 |
可能性和影响是基于已引用监管、法律和商业证据得出的承销判断;私有事故和毛利数据可能改变排序。
[CR012, CR020, CR022, CR026, CR027, CR030]| 风险领域 | 证据 | 投资影响 | 监测指标 |
|---|---|---|---|
| 聚合器毛利率 | 第三方模型 API 和算力转售决定价差经济。 | 若没有产品级贡献毛利,ARR 质量会打折。 | 按产品、模型家族和使用 cohort 拆分毛利率。 |
| 模型访问和替代 | 模型原厂可以降价、捆绑应用或限制 API。 | 工作流锁定必须抵消商品化模型访问。 | API 条款、排队时间、提供方集中度和价格变化后的流失。 |
| 资本强度 | US$300M B+ 支撑 R&D、全球扩张和产品组合搭建。 | 融资降低 runway 风险,但可能掩盖补贴依赖。 | 月度 burn、云承诺、预付模型点数和回本周期。 |
| 尚未盈利 | AIbase 报道,截至 Series B 阶段公司尚未盈利。 | 增长可能靠风险资本供血,而非自我造血。 | EBITDA、现金 runway、CAC 回本和按用量单位经济。 |
| 企业采购风险 | 已点名客户增加证明力,也抬高审计预期。 | 合规失误可能拖慢 B 端采用。 | 安全 / 法务问卷、采购被拒次数和续约数据。 |
| 创始人集中 | 创始人驱动速度是产品扩张的核心。 | 执行上行伴随继任和控制风险。 | 董事会纪要、授权地图、关键人保险和继任计划。 |
运营行结合公开证据和尽调假设;需要私有财务和治理数据才能量化剩余风险敞口。
[CR026, CR027, CR029, CR030, CR031, CR032]内容安全、IP 责任、模型访问和毛利压缩处在最高残余风险单元格。
定性投资测算矩阵;评级需在完成私有事故、毛利和治理尽调后更新。
[CR042, CR043, CR045, CR046, CR047]7.4 治理尽调应聚焦集中度、控制项和可量化触发线
人和治理风险中等但真实。陈冕此前 CapCut 商业化经验是一项资产,因为公司依赖快速判断 PMF;但同一证据也让投资论点以创始人为中心。来源包报告的 73.95% 创始人持股和之后的法定代表人变更,不应被当作不当行为;它们应成为董事会构成、保留事项、关联方控制、信息权和继任规划的尽调提示。务实投资流程应把每项主要风险转化为触发线:再次受到 CAC 或应用商店行动、AIGC 标识缺陷未解决、另一次 CCTV 级审核失败、IP 通知积压未解决、LoRA 工作流被禁令限制、关键模型 API 流失或重新定价,或有证据显示规模化后产品级毛利仍结构性为负。只有当合规、IP 运营和模型采购纪律像产品发布一样可重复,Liblib 才有吸引力。证据还留下几个定量问题:当前合规人员规模、通知量、红队失败率和 API 提供商集中度都未公开,因此每一项都应成为交割前交付物,而不是投后改进计划。[CR038, CR039, CR040, CR041, CR043, CR044]
7.5 图表
08估值
8.1 头部估值:大额融资,但 ARR 倍数不高
Liblib / Evoken 2026 年 6 月融资表面上很大,但用公开 ARR 证据归一化后,并不明显虚高。多份独立报道共同指向:Series B+ 融资约 US$300 million,投后估值超过 US$2 billion,2026 年 5 月 ARR 约 US$300 million。这让头部轮次约为 6.7x ARR,且尚未调整资产负债表现金或未公开的优先条款。对于一家成立两年的 AI 应用集团,若其拥有超过 30 million 用户、500,000 个原创模型,且收入背后有专业内容预算支撑,这个价格意外克制。因此,承销问题不是投资者是否支付了典型 AI 泡沫倍数,而是报告 ARR 在扣除第三方模型和算力成本后,是否经常性、能留存且毛利为正。这个框架也让估值更容易与传统 SaaS 对比:如果把超过 US$2 billion 的标记视为企业价值下限,那么除非真实投后估值远高于披露值,否则 US$2 billion 以上每增加一美元,只会温和推高倍数。[CV001, CV002, CV003, CV004, CV005, CV008]
| 轮次 | 日期 | 金额 | 投后估值 | 隐含 ARR 倍数 | 证据说明 |
|---|---|---|---|---|---|
| Series B | Oct 2025 | US$130M | 本报告所用来源未披露 | n/a | 2025 年已披露中国 AI 应用最大融资;盈利情况尚未公开 |
| Series B+ | June 18 2026 | ~US$300M | >US$2.0B | 基于 US$300M ARR 约 6.7x | Granite Asia、Tencent 和 Shunwei 共同领投 |
| 两轮合计 | Oct 2025-Jun 2026 | >US$500M | >US$2.0B 当前估值 | n/a | 约八个月内募集的资本支撑 R&D 和扩张 |
| 当前 ARR 基数 | May 2026 | n/a | 估值分母 | US$300M ARR | 报道称 ARR 同比增长超过 30x |
ARR 倍数用 US$2.0B 除以 US$300M;实际投后估值仅披露为高于 US$2B,因此该倍数是下限 / 近似值。
[CV001, CV002, CV003, CV004, CV008, CV009]| 细分板块 | 2026 倍数区间 | 来源依据 | 对 Liblib 的含义 |
|---|---|---|---|
| 公有云 / 广义 SaaS 指数 | 平均收入约 6.2x | ScaleXP 引用 BVP Nasdaq Emerging Cloud Index | 尽管有私有公司流动性折价,Liblib 仍接近公有云平均水平 |
| 传统 / 遗留 SaaS | ARR 约 4x-7x | Acquiry 和 We Are Founders 的重置期区间 | Liblib 仅略高于常规 SaaS |
| 风险资本支持的私有 SaaS | 中位数约 5.3x;上四分位 8x-10x | We Are Founders 基准表 | Liblib 高于中位数,但不到私有 SaaS 上四分位 |
| AI / 垂直 SaaS | 约 9x-12x;溢价 15x+ | We Are Founders 对 AI 溢价的讨论 | 除非证明工作流锁定,Liblib 估值低于垂直 AI 溢价 |
| 高增长 AI 原生 SaaS | ARR 约 10x-20x | Acquiry 2025 / 2026 年初交易区间 | 相比头部 AI 原生应用,Liblib 折价较大 |
| AI 原生 VC 轮次 | 收入中位数约 21.2x | SaaSRise AI 软件估值报告 | Liblib 明显低于 AI 原生风险投资中位数 |
区间混合使用公私市场的 ARR 和收入倍数;Liblib 的收入确认和毛利率仍为私有信息,因此只能方向性参考。
[CV015, CV016, CV017, CV018, CV019, CV020]KPI 组合显示规模和增长强劲,但盈利能力和留存仍未解答。
公开报道称估值高于 US$2B,因此这里的估值是下限;增长按报道中的 >3,000% 口径呈现。
[CV001, CV002, CV003, CV004, CV008, CV010]8.2 基准让 6.7x 显得便宜,但不是每一美元 AI 收入都应享受溢价
外部基准组合支持中性偏建设性的估值立场。2026 年 SaaS 来源显示,公有云平均值接近中个位数,传统 SaaS 约为 4x 至 7x ARR,而更高增长的 AI 原生或垂直 AI 软件高得多。在这个背景下,Liblib 的 6.7x ARR 倍数低于 AI 原生头部四分位,更接近重置周期后的公开 SaaS。创意 AI 可比公司进一步强化这个点:Canva 的 US$4 billion ARR 支撑 US$42 billion 至 US$65 billion 估值区间,Midjourney 似乎接近 20x 收入倍数,Sacra 的 Gamma 数据点高于 20x ARR。折价合乎逻辑,不是惩罚性定价。Liblib 增长异常,但公开证据尚未证明专有模型经济学、企业留存,或可与全球最佳 AI 创意平台相比的持久工作流护城河。[CV015, CV016, CV017, CV018, CV019, CV020]
| 公司 | ARR / 收入参考 | 估值参考 | 隐含倍数 | 相关性 / 局限 |
|---|---|---|---|---|
| Liblib / Evoken | May 2026 ARR ~US$300M | >US$2.0B 投后 | ~6.7x | 目标公司;中国业务占比高,聚合器经济性披露不充分 |
| Canva | 2025 年底 ARR US$4.0B | US$42B–US$65B | ~10.5x-16.3x | 规模化设计平台,套件更深,披露的 B2B 基础更强 |
| Midjourney | 2025 年收入 ~US$500M | ~US$10B | ~20x | AI 图像参照标的;私募估算和产品模式不同 |
| Runway ML | US$70M-US$80M ARR | US$5.3B | ~68x-76x | AI 视频可比公司,但倍数高得多,模型栈也不同 |
| ElevenLabs | ~US$330M ARR | US$11B | ~33.3x | AI 原生语音可比公司,有企业客户验证 |
| Gamma | 2025 年 10 月 ARR US$102M | US$2.1B | 20.6x | AI 演示 / 工作流可比公司;ARR 基数更小,但倍数高 |
该列举是估值可比标的集合,不是完整市场清单;各私营公司的 ARR/收入定义和优先条款不一。
[CV021, CV022, CV023, CV024, CV025, CV026]Liblib 估算的 6.7x 倍数明显低于所引 AI 创意可比公司,接近估值重置期 SaaS 参照。
Runway、ElevenLabs 和 Midjourney 的倍数采用所引私募估算;Canva 高端采用 Sacra 二级市场估值。
[CV022, CV023, CV024, CV025, CV026, CV027]8.3 情景分析:工作流锁定能抵消聚合器风险时,估值才站得住
情景法指向一个有条件的「合理」价格。熊市情景假设 ARR 卡在已披露的 US$300M 附近,零售端使用仍对补贴敏感,模型原厂压缩转售价差;按 4 倍 ARR 算,隐含价值约 US$1.2B。基准情景假设前瞻 ARR 升至约 US$420M,但毛利率、NRR 和 API 成本传导仍未披露,投资人只给 5 倍 ARR,估值落在约 US$2.1B,足以支撑本轮。牛市情景需要证明 LibTV、LiblibAI、Xingliu 和 Lovart 正在变成一个有工作流锁定的创作者操作系统;若前瞻 ARR 达到 US$500M、市场愿意给 8 倍,公司可支撑约 US$4B。只有私下尽调验证收入质量后,这种不对称才有吸引力。[CV028, CV029, CV030, CV031, CV032, CV033]
| 情景 | ARR 假设 | 倍数假设 | 隐含价值 | 概率信号 | 下行 / 上行触发因素 |
|---|---|---|---|---|---|
| 悲观 | 当前 ARR US$300M | 4.0x | US$1.2B | ARR 停滞,或被证明靠补贴拉动 | API 成本上升,第一方平台降价,留存走弱 |
| 基准 | 远期 ARR US$420M | 5.0x | US$2.1B | 增长延续,但利润率仍不透明 | 非公开毛利率和 NRR 合格,但不是一流 |
| 乐观 | 远期 ARR US$500M | 8.0x | US$4.0B | 工作流锁定开始显性化 | 经审计留存、企业占比和产品级利润率验证创作者 OS 论点 |
情景输出是根据引用的 ARR、基准倍数和聚合器风险折价推导出的承销估算,不是公司指引。
[CV031, CV032, CV033, CV034, CV035, CV048]| 驱动因素 / 风险 | 证据信号 | 估值影响 | 尽调测试 |
|---|---|---|---|
| ARR 规模 | 2026 年 5 月 ARR US$300M | 支撑当前 >US$2B 估值 | 对齐 ARR、开票额、收入和队列 |
| 增长速度 | 披露同比增长 >30x 或 >3,000% | 支撑高增长溢价倍数 | 拆分自然留存、发布峰值和补贴 |
| 工作流锁定 | 专业内容团队使用 LibTV 和 LiblibAI | 可能把倍数推向垂直 AI 区间 | 验证企业重复使用和团队级切换成本 |
| 聚合器经济性 | 没有自有基础模型,依赖第三方 API | 把倍数压向公开 / 传统 SaaS | 按产品审计毛利率和 API 成本合同 |
| 中国 AI 泡沫 | 2026 年上半年中国新增 67 家独角兽,且有 2027–2028 年负面出清警告 | 抬高降轮和退出窗口风险 | 将烧钱、资金续航和收入质量与同批公司对比 |
| 监管 / 内容安全 | 网信办备案制度和 CCTV 相关内容安全风险 | 增加合规折价 | 核验备案、标识控制和整改证据 |
本表把公开证据映射到尽调动作;缺失的私有指标刻意按估值折价处理。
[CV003, CV011, CV012, CV028, CV029, CV036]公开 ARR 支撑当前估值;利润率和监管尽调决定倍数是扩张还是压缩。
该流程是基于所引证据的投资承销逻辑图,并非因果模型。
[CV004, CV020, CV028, CV035, CV049, CV050]情景估值从 US$1.2B 熊市情形到 US$4.0B 牛市情形,覆盖当前估值。
数值以十亿美元计并已四舍五入;区间反映 ARR 与倍数假设的敏感性。
[CV031, CV032, CV033, CV048, CV049]8.4 最终判断:估值有支撑,但要扣除中国 AI 泡沫和合规折价
2026 年中国独角兽环境既提供支撑,也提示谨慎。Nation Press 报道,2026 年上半年中国新增 67 家独角兽;Hurun 全球榜单显示,AI 是私募市场价值创造的主引擎。这意味着 Liblib 并非孤立的融资异象,而是处在资本向 AI 应用、机器人、基础设施和工作流自动化大规模再配置的大潮里。反向因素同样关键:China Biz Insider 对具身智能的现实核查警告称,许多现金跑道短的 AI 独角兽可能在 2027–2028 年面临整合、降价融资或失败。Liblib 已披露 ARR,股东名单也强于许多概念阶段同行;但监管 / 内容安全风险敞口,以及缺乏公开利润率指标,都应明确进入估值折价。基于公开记录,超过 US$2B 的估值有支撑,但它只是一个合理且对证据敏感的入场价,不是明显便宜货。因此,投委会的正确姿态不是把本轮斥为 AI 亢奋,而是要求证明收入质量强于折价所暗示的水平:经审计的 ARR 桥、客户 cohort、贡献毛利和合规控制,才能把结论从合理推向有吸引力。[CV036, CV037, CV038, CV039, CV040, CV041]
| 投资方 | B+ 轮角色 / 历史 | 信号质量 | 估值相关性 |
|---|---|---|---|
| Granite Asia | B+ 轮共同领投 | 高:总部在新加坡的成长型投资机构,积极配置 AI | 增加跨境成长资本可信度 |
| Tencent Holdings | B+ 轮共同领投 | 高:战略型中国平台投资方 | 潜在分发、生态以及云 / 支付信号 |
| Shunwei Capital | B+ 轮共同领投及既有投资方 | 高:雷军关联的中国创投平台 | 创始人—市场匹配和中国消费 / 应用信号 |
| HSG / Sequoia China | 既有投资方加码或持续持有 | 高:中国顶级创投品牌 | 跟投降低逆向选择担忧 |
| Gaorong Capital | 既有投资方加码 | 中高:中国早期创投支持方 | 支撑早期轮次以来的连续性 |
| Ant Group | 既有投资方加码 | 战略相关性高 | 潜在金融科技 / 支付 / 云生态邻近性 |
| HT Investment / Times Capital | 跟投参与方 | 中:后期轮支持 | 扩大融资财团,但公开差异化较弱 |
角色来自 2026 年融资报道;确切持股、清算优先权和治理条款未公开。
[CV005, CV006, CV007, CV041, CV042, CV043]8.5 展示材料
免责声明
本报告基于截至 2026-07-15 的公开及公司披露信息,用于尽调筛选。财务和经营指标大多是未经审计的公司说法。不构成投资建议。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Evoken is the parent company of LiblibAI, a major Chinese AI image creation and sharing platform. | 高 | SO001, SO006 |
| CO002 | Evoken was formerly known as Qidian Xingyu Technology and adopted the Evoken corporate name in June 2026. | 高 | SO001, SO006 |
| CO003 | The company is based in Beijing, China. | 高 | SO001, SO018 |
| CO004 | Beijing Qidian Xingyu Technology Co., Ltd. was founded on 2023-05-16. | 中 | SO018, SO007 |
| CO005 | LiblibAI is positioned as an AI-native creative platform for AI content creation, sharing, model discovery, and cloud generation. | 中 | SO007, SO024, SO020 |
| CO006 | LiblibAI started from AI image generation and model sharing before expanding into a broader AI creative studio and product group. | 中 | SO013, SO015, SO024 |
| CO007 | Founder Chen Mian previously oversaw global commercialization for ByteDance's Jianying / CapCut product family. | 高 | SO001, SO006, SO012 |
| CO008 | Chen Mian was born in 1992 and is described as a product-oriented founder with strong commercialization focus. | 中 | SO006, SO012 |
| CO009 | Chen Mian is reported to hold roughly 73.95% of the company and remains the controlling founder. | 中 | SO012, SO006 |
| CO010 | Zhang Zijie is publicly associated with the company as an early legal representative and co-founder / core product-community figure. | 中 | SO018, SO012, SO007 |
| CO011 | Yang Nan became the legal representative on 2025-11-22 while Chen Mian remained the founder and largest shareholder. | 中 | SO012 |
| CO012 | The core team is reported to include graduates from Tsinghua, Peking University, Carnegie Mellon, and alumni of Tencent, Alibaba, and ByteDance. | 中 | SO012, SO022, SO005 |
| CO013 | The company's 2023 angel financing was about US$3.5 million at about a US$15 million valuation. | 中 | SO014, SO016, SO011 |
| CO014 | The 2023 angel investor group included GSR Ventures / Jinshajiang, Gaorong Capital, and Source Code Capital. | 中 | SO007, SO012, SO023 |
| CO015 | In 2024, LiblibAI completed several rounds totaling multiple hundreds of millions of RMB or more than US$20 million, with Mingshi, Yingce, Shunwei, and strategic investors appearing in reports. | 中 | SO014, SO022, SO023, SO010 |
| CO016 | By year-end 2024, reports placed LiblibAI's valuation above US$500 million. | 中 | SO014, SO016 |
| CO017 | In February and March 2025, LiblibAI added further several-hundred-million-RMB financings and model integrations. | 中 | SO007, SO012, SO021 |
| CO018 | In October 2025, LiblibAI completed a US$130 million Series B round. | 高 | SO001, SO013, SO015, SO016 |
| CO019 | The October 2025 Series B was led by HSG / Sequoia China, CMC Capital, and an undisclosed strategic investor. | 中 | SO013, SO015, SO016 |
| CO020 | Existing investors including Shunwei, Source Code, Mingshi, Yingce, INCE Capital, Ant Group, and Lenovo Capital participated or increased in the Series B syndicate. | 中 | SO015, SO016, SO007 |
| CO021 | The October 2025 Series B was described as the largest disclosed China AI-application financing of 2025. | 中 | SO013, SO015, SO016 |
| CO022 | LiblibAI was not yet fully profitable at the time of the Series B financing. | 中 | SO014, SO016 |
| CO023 | On 2026-06-18, Evoken completed a nearly US$300 million / US$300 million Series B+ financing at a post-money valuation above US$2 billion. | 高 | SO001, SO002, SO003, SO006 |
| CO024 | The Series B+ was jointly led by Granite Asia, Tencent Holdings, and Shunwei Capital. | 高 | SO001, SO002, SO006, SO011 |
| CO025 | HT Investment and Times Capital joined the Series B+ while HSG / Sequoia China, Gaorong, Ant Group, and other existing investors followed on. | 中 | SO002, SO006, SO011 |
| CO026 | The company disclosed more than US$500 million raised across the two large rounds within eight months. | 中 | SO002, SO003, SO006 |
| CO027 | Evoken said Series B+ proceeds would support R&D, global expansion, AI creative product capabilities, and portfolio build-out. | 中 | SO001 |
| CO028 | LiblibAI had more than 30 million cumulative users by mid-2026. | 高 | SO001, SO002, SO003, SO011 |
| CO029 | LiblibAI had accumulated more than 500,000 original models by mid-2026. | 中 | SO002, SO003, SO011, SO006 |
| CO030 | LiblibAI was reported to generate more than 5 million images per day. | 中 | SO003, SO011, SO012, SO005 |
| CO031 | Evoken's ARR exceeded US$300 million as of May 2026. | 高 | SO001, SO002, SO003, SO004 |
| CO032 | Evoken's group revenue grew more than 3000% year over year in May 2026. | 中 | SO002, SO004, SO011 |
| CO033 | Tencent News reported 2024 full-year revenue of RMB 206 million, approximately US$28.5 million. | 中 | SO021 |
| CO034 | By October 2025, reports put LiblibAI at about 25 million total users and about 4 million monthly active users. | 中 | SO021, SO012, SO017 |
| CO035 | LiblibAI, LibTV, and Xingliu form the core domestic product lines disclosed around Evoken's 2026 financing. | 中 | SO001, SO006, SO011 |
| CO036 | LibTV was launched in March 2026 as an AI video-creation platform for professional production workflows. | 中 | SO001, SO011, SO012 |
| CO037 | LibTV recorded more than US$1 million in single-day revenue in its first month and served nearly 1,000 professional teams. | 中 | SO002, SO004, SO011 |
| CO038 | Xingliu is a domestic AI design Agent reported to have more than 10 million users. | 中 | SO001, SO002, SO006, SO011 |
| CO039 | Lovart was introduced as an overseas design-agent product / beta associated with the LiblibAI platform in 2025. | 中 | SO012, SO005 |
| CO040 | LiblibAI integrates open and closed model capabilities and offers LoRA / custom model workflows through a community platform. | 中 | SO015, SO021, SO024 |
| CO041 | LiblibAI passed CAC deep-synthesis algorithm filing in February 2024 and became a generative-AI-services filed AI community in March 2024. | 高 | SO007, SO010, SO022, SO019 |
| CO042 | As of 2025-11-01, CAC said 611 generative-AI services had completed filing and 306 generative-AI applications or functions had completed registration. | 中 | SO019 |
| CO043 | CCTV exposed in April 2026 that LiblibAI could bypass moderation under specific prompts to generate inappropriate pornographic content. | 中 | SO012 |
| CO044 | LiblibAI responded after the CCTV exposé that it initiated internal self-inspection, technical rectification, and moderation-strategy upgrades. | 中 | SO012 |
| CO045 | LiblibAI was briefly forced offline in its early stage after failing to complete required large-model filing. | 中 | SO014, SO016 |
| CO046 | Multiple reports describe LiblibAI as an application-layer company that packages third-party and community model capabilities rather than competing primarily as a foundation-model lab. | 中 | SO003, SO009, SO021, SO024 |
| CO047 | Hurun reported 1,603 global unicorns in 2026, with China second globally at 381 unicorns and AI among the leading sectors. | 中 | SO025 |
| CO048 | LiblibAI's official website title identifies it as a Chinese leading AI creation platform. | 中 | SO020 |
| CM001 | Liblib's relevant market boundary includes China AIGC solutions and services for automated content creation, image generation, video and animation generation, model communities, and workflow tooling. | 中 | SM001, SM024 |
| CM002 | IMARC reported China AIGC market size of US$5.16082 billion in 2025 and forecast US$19.55829 billion by 2034 at a 15.96% CAGR. | 中 | SM001 |
| CM003 | Grand View Research's China generative-AI databook is used as the lower public sizing lens around US$2.36 billion and exposes software/services segmentation, with software at 63.9% of 2024 revenue. | 中 | SM002 |
| CM004 | The public AIGC sizing range used in this chapter is approximately US$2.36 billion to US$5.16 billion for the current China generative-AI market lens. | 高 | SM001, SM002 |
| CM005 | IMARC's technology segmentation explicitly includes text-to-image models, text-to-video/3D, GANs, transformer models, text-to-speech, and speech-to-text. | 中 | SM001 |
| CM006 | IMARC's component segmentation includes image generation platforms and video and animation generators, directly matching Liblib's image and LibTV video workflows. | 中 | SM001 |
| CM007 | 6Wresearch describes China AI image generation as led by global technology leaders and domestic pioneers, with Baidu, Alibaba, and Tencent holding sizable share. | 中 | SM003 |
| CM008 | 6Wresearch identifies Baidu ERNIE-ViLG as a text-to-image model specializing in Chinese cultural nuances and art styles. | 中 | SM003 |
| CM009 | 6Wresearch identifies Alibaba Tongyi Wanxiang as an AI painting model for e-commerce and cloud ecosystems and Tencent Hunyuan as a multimodal model for social and gaming platforms. | 中 | SM003 |
| CM010 | DigiTrendz reported that China's micro-drama industry is projected to exceed RMB120 billion, or about US$16.5 billion, in 2026. | 中 | SM004 |
| CM011 | The Next Web independently reported the same micro-drama market projection above RMB120 billion and described it as a US$16.5 billion industry serving 660 million users. | 中 | SM005 |
| CM012 | Multiple CH2 sources support using the 2026 micro-drama market as Liblib's broad B-side demand pool rather than as direct software revenue. | 高 | SM004, SM005, SM006 |
| CM013 | The Next Web reported that about 50,000 AI-native titles were added to Douyin in March 2026, illustrating industrial-scale content supply. | 中 | SM005 |
| CM014 | DigiTrendz reported that AI-driven workflows cut production time from three months to one month and reduce costs to about one-fifth of traditional shoots. | 中 | SM004 |
| CM015 | The Next Web reported that AI-native micro-drama production can run at roughly one-tenth the cost of live-action production. | 中 | SM005 |
| CM016 | CRI reported that an AI-generated version of a comparable micro-drama can cost just over RMB100,000 versus several hundred thousand yuan for live action. | 中 | SM008 |
| CM017 | Wonford reported China short-drama export revenue of US$2.38 billion in 2025, up 263% year over year. | 中 | SM007 |
| CM018 | DigiTrendz reported overseas micro-drama revenue of US$1.525 billion in the first eight months of 2025, up 195% year over year. | 中 | SM004 |
| CM019 | CRI reported AI-generated comic-style micro-dramas represented an estimated RMB16.8 billion, or about US$2.44 billion, of market share in 2025. | 中 | SM008 |
| CM020 | AInChina estimated China's AI short-drama market expanded from a US$100 million niche to a US$650 million industry in Q1 2026. | 中 | SM009 |
| CM021 | The chapter's SAM range treats AI-generated micro-drama demand as at least hundreds of millions of dollars in Q1 2026 and potentially several billions of dollars when comic-style micro-dramas are included. | 中 | SM008, SM009 |
| CM022 | Yicai reported LiblibAI had more than 30 million cumulative users by June 2026 and was one of China's largest AI image-resource libraries and creator communities. | 中 | SM010 |
| CM023 | AIbase reported Evoken's ARR exceeded US$300 million as of May 2026, with core revenue from professional content producers in short dramas, films, and advertising. | 中 | SM011 |
| CM024 | KuCoin reported LiblibAI had more than 30 million users, over 500,000 original models, and more than 5 million images generated daily. | 中 | SM012 |
| CM025 | BestHub reported Yanyu's products draw organic traffic and the broader creator ecosystem exceeds 30 million users. | 中 | SM013 |
| CM026 | Liblib's current SOM is best proxied by its disclosed US$300 million ARR, not by the full RMB120 billion micro-drama content market. | 中 | SM011, SM013 |
| CM027 | Liblib's buyer map centers on professional content producers in short dramas, film, advertising, design, e-commerce, and gaming, while users include creators and designers operating AI image/video workflows. | 中 | SM001, SM011, SM012, SM015 |
| CM028 | Tencent News connected booming AI drama demand to demand for related models and tools, naming micro-drama as a pull factor for creative AI tooling. | 中 | SM015 |
| CM029 | IMARC cited strong government support, enterprise digital transformation, expanding AI infrastructure, and automated content-creation demand as China AIGC growth drivers. | 中 | SM001 |
| CM030 | The Next Web described local-government subsidies and production hubs alongside NRTA review as an industrial policy framework for AI entertainment. | 中 | SM005 |
| CM031 | CAC's filing registry listed 611 generative-AI services cumulatively by November 2025, showing compliance filing is a market-access feature in China. | 中 | SM022 |
| CM032 | 36Kr PitchHub records that Liblib completed a deep-synthesis service algorithm filing in February 2024 and a generative-AI service filing in March 2024. | 中 | SM016 |
| CM033 | IMARC identified data governance, content filtering, watermarking, high computing cost, reliability, and enterprise adoption barriers as China AIGC market challenges. | 中 | SM001 |
| CM034 | The Next Web warned that when AI production costs fall by 90%, output volume rises by an order of magnitude and much of it is mediocre. | 中 | SM005 |
| CM035 | Startup Fortune similarly described a flood of similar stories, visual styles, and emotional beats at the lower end of the AI micro-drama market. | 中 | SM006 |
| CM036 | Wonford identified copyright ownership, personality rights, content homogenization, and AI-plus-content talent gaps as challenges for AI short dramas. | 中 | SM007 |
| CM037 | Baidu Baike reports a CCTV adverse incident in which LiblibAI was named for prompts that could bypass moderation, underscoring content-safety risk as a market shaper. | 中 | SM017 |
| CM038 | AIbase reported LiblibAI had earlier been forced offline for incomplete large-model filing, demonstrating that regulatory readiness can interrupt market participation. | 中 | SM019 |
| CM039 | The application-layer investment thesis is supported by AIbase's report that AI investment was shifting from foundation models toward application-layer companies. | 中 | SM018 |
| CM040 | INCE Capital and 36Kr reported the October 2025 Series B financing, supporting the view that capital was available for validated AI application-layer market leaders. | 中 | SM020, SM021 |
| CM041 | Tencent News reported LiblibAI 2024 revenue of RMB206 million, providing a historical baseline below the much larger May 2026 ARR claim. | 中 | SM023 |
| CM042 | Hubpy describes LiblibAI as a 2026 guide-worthy platform for AI model and creator workflows, supporting the workflow-tooling portion of the market boundary. | 低 | SM024 |
| CM043 | Hurun's unicorn context supports classifying Evoken as a China AI unicorn within a broader private-market cohort, but it does not isolate Liblib's market share. | 中 | SM025 |
| CM044 | Because public sources do not disclose Liblib's exact paid-customer count, vertical revenue mix, or Chinese creative-professional penetration, the precise SAM-to-SOM bridge remains only partially evidenced. | 低 | |
| CP001 | Evoken / LiblibAI was reported in June 2026 as valued above US$2 billion after a US$300 million Series B+ financing. | 高 | SP010, SP011, SP012, SP014 |
| CP002 | LiblibAI was reported in mid-2026 at more than 30 million cumulative users, more than 500,000 original models, and more than 5 million images generated per day. | 高 | SP010, SP011, SP012 |
| CP003 | LiblibAI was reported to have exceeded US$300 million ARR as of May 2026, with core revenue tied to professional content producers in short drama, film, and advertising. | 高 | SP010, SP011, SP013 |
| CP004 | LiblibAI 2.0 is positioned as a professional creation studio that aggregates multiple models and supports video generation, effects templates, LoRA training, a marketplace, and API-oriented workflows. | 中 | SP020, SP018, SP011 |
| CP005 | LibTV launched in March 2026 as an AI video-creation platform, extending Liblib beyond still image generation into agentic video workflows. | 中 | SP011, SP010 |
| CP006 | Xingliu and Lovart broaden Evoken from a model-sharing community into a multi-product creative platform spanning domestic and international AI design agents. | 中 | SP010, SP011 |
| CP007 | LiblibAI completed Chinese deep-synthesis / generative-AI service filing milestones in early 2024, and CAC maintains the public registry context for such filings. | 高 | SP015, SP017 |
| CP008 | The China AI image-generator market includes top ecosystem leaders Baidu, Alibaba, and Tencent with a sizable combined market share. | 中 | SP024 |
| CP009 | Alibaba uses Tongyi Wanxiang to provide professional-grade image generation for e-commerce and cloud-computing ecosystems. | 中 | SP024 |
| CP010 | Tencent integrates Hunyuan multimodal models into social and gaming platforms for asset creation and digital avatars. | 中 | SP024 |
| CP011 | ByteDance offers AIGC image and video creation tools such as Lumi to support its creator ecosystem. | 中 | SP024 |
| CP012 | Civitai is an AI art platform centered on discovering, creating, and sharing Stable Diffusion media and models. | 中 | SP003, SP002 |
| CP013 | Civitai exposes a broader community layer including models, images, videos, 3D models, comics, articles, challenges, collections, leaderboards, and creator programs. | 中 | SP002 |
| CP014 | Civitai provides a free plan and paid access starting at US$10 per month, with paid tiers adding Buzz allowances, queue depth, support, and creator controls. | 中 | SP002 |
| CP015 | Independent Civitai-versus-Midjourney reviews position Civitai as stronger for control, custom models, model ownership, and long-term cost. | 中 | SP001, SP002 |
| CP016 | Independent Civitai-versus-Midjourney reviews position Midjourney as stronger for beginners, speed, out-of-box quality, and simpler commercial licensing. | 中 | SP001, SP002 |
| CP017 | Midjourney describes itself as a 60-person lab known for building high-quality proprietary AI models. | 中 | SP004 |
| CP018 | SeaArt AI combines text/image/video generation, LoRA training, face swap, AI characters, ComfyUI workflows, upscaling, and filters in one cloud suite. | 中 | SP005, SP006 |
| CP019 | Skywork reported that SeaArt AI has ethical and NSFW-filter loophole concerns, including provocative or explicit content in an open-library context. | 中 | SP005 |
| CP020 | SeaArt AI uses a hybrid stamina / credits model with plans observed around US$4.79 to US$75 per month and potentially confusing credit consumption. | 中 | SP005, SP006 |
| CP021 | Cybernews characterized SeaArt as a capable all-in-one suite but flagged licensing clarity and public pricing transparency as professional watch-outs. | 中 | SP006 |
| CP022 | Tensor.Art was incubated by Huixiang Technology, started in May 2023, and is built around Stable Diffusion image generation and model-hosting community workflows. | 中 | SP009 |
| CP023 | Tensor.Art supports text-to-image, image-to-image, style transfer, repair/enhancement, upscaling, conditional control generation, ComfyUI workflows, online training, model hosting, batch generation, and API services. | 中 | SP009, SP008 |
| CP024 | Tensor.Art was reported to have more than 160,000 models and 3.97 million global monthly visits in July 2024, with further traffic growth noted by September 2025. | 中 | SP009 |
| CP025 | Wujie AI is built by Hangzhou Chaojiepoint and differentiates through Wujie Bantu, which can register and auction AI-generated work as verified digital copyright. | 中 | SP007 |
| CP026 | Wujie AI offers text-to-image and image-to-image generation, prompt search, reference libraries, pose recognition, a creator plaza, API access, and RMB-priced membership tiers. | 中 | SP007 |
| CP027 | Recatools described Wujie AI as Chinese-only, information-dense, and more compelling for users who value domestic copyright registration than for broad ASEAN users. | 中 | SP007 |
| CP028 | AituBo is positioned in a third-party comparison as a free, beginner-friendly AI image and video generation platform with editing, avatar chat, background removal, upscaling, and face-swap tools. | 中 | SP008 |
| CP029 | Sharewalker described Liblib as a model-sharing-centered platform with more than 100,000 model resources across anime games, photography illustration, and brand design. | 中 | SP008 |
| CP030 | Sharewalker described Tensor.Art as the more complete all-in-one creation option among Liblib, Tensor.Art, SeaArt, and AituBo. | 中 | SP008 |
| CP031 | Sharewalker described SeaArt as attractive for users exploring diverse AI applications such as virtual characters and model training. | 中 | SP008 |
| CP032 | Sharewalker described AituBo as suitable for limited-budget beginners needing rapid creation. | 中 | SP008 |
| CP033 | 6pen is in the requested direct-competitor scope, but the retained CH3 source texts did not provide a verifiable current profile, scale metric, or feature matrix entry. | 低 | |
| CP034 | PixAI is in the requested direct-competitor scope, but the retained CH3 source texts did not provide a verifiable current profile, scale metric, or feature matrix entry. | 低 | |
| CP035 | The direct peer set creates a competitive threat because SeaArt, Tensor.Art, Wujie, and AituBo each replicate portions of Liblib’s image/video generation, LoRA or model, workflow, marketplace, or low-cost beginner proposition. | 中 | SP005, SP006, SP007, SP008, SP009 |
| CP036 | Liblib’s strongest defensibility signal is the combination of China-focused community scale, a large model marketplace, professional creator revenue, and a regulatory filing record. | 高 | SP002, SP010, SP011, SP015, SP017, SP020 |
| CP037 | Liblib remains exposed to commoditization because direct peers and Chinese tech giants can offer overlapping generation, LoRA, workflow, marketplace, and ecosystem-distribution capabilities. | 中 | SP005, SP007, SP008, SP009, SP024 |
| CP038 | Midjourney and Civitai illustrate a global substitute pattern in which proprietary quality/speed and open model ownership/control can bracket Liblib’s China-localized community model. | 中 | SP001, SP002, SP003, SP004 |
| CP039 | China’s AIGC market expansion and micro-drama/video demand support multiple well-funded competitors rather than a winner-take-all creative-tool market. | 中 | SP022, SP023, SP025, SP026 |
| CP040 | Civitai and Midjourney official pages were retained as JavaScript-only official surfaces, so third-party review evidence carries more of the comparative feature burden in this chapter. | 中 | SP003, SP004, SP001, SP002 |
| CP041 | The retained source pool did not independently verify the brief-level traffic comparison of Civitai at roughly 7.5 million monthly visits versus Liblib at roughly 2.2 million with about 91% China traffic; the metric should be refreshed from a traffic panel before use in underwriting. | 低 | |
| CI001 | Evoken completed a June 2026 Series B+ round of about US$300 million at a post-money valuation above US$2 billion. | 高 | SI010, SI012, SI014 |
| CI002 | Granite Asia, Tencent, and Shunwei Capital jointly led the June 2026 Series B+ round. | 高 | SI010, SI011 |
| CI003 | The company said the B+ proceeds would fund R&D, global expansion, AI creative product capability, and product-portfolio build-out. | 高 | SI010, SI011 |
| CI004 | Evoken's latest two disclosed rounds in the eight months to June 2026 totaled more than US$500 million in financing. | 中 | SI011, SI012 |
| CI005 | LiblibAI announced a US$130 million Series B round in October 2025 led by Sequoia China or HongShan, CMC Capital, and a strategic investor. | 中 | SI017, SI019, SI020 |
| CI006 | During 2024, LiblibAI completed three rounds totaling hundreds of millions of renminbi, including financing led by Mingshi Capital. | 中 | SI025, SI026 |
| CI007 | LiblibAI's early angel financing was reported at US$3.5 million on an approximately US$15 million valuation. | 中 | SI018, SI015 |
| CI008 | Public profiles list February or March 2025 financing rounds of several hundred million renminbi before the October 2025 Series B. | 中 | SI015, SI018 |
| CI009 | Founder Chen Mian remains the largest disclosed shareholder, with public profiles indicating roughly a 73.95% stake. | 中 | SI016, SI014 |
| CI010 | Chen Mian previously oversaw commercialization of ByteDance's Jianying and CapCut products before founding Evoken. | 中 | SI010, SI014 |
| CI011 | Evoken reported annual recurring revenue of US$300 million as of May 2026. | 高 | SI010, SI012, SI014 |
| CI012 | Reported ARR growth was described as more than 30 times year over year or more than 3,000% year over year. | 高 | SI010, SI012 |
| CI013 | Tencent News reported LiblibAI's 2024 full-year revenue as RMB206 million, approximately US$28.5 million at contemporaneous exchange rates. | 中 | SI024, SI014 |
| CI014 | Yicai and KuCoin described the current revenue base as mainly coming from professional content producers in short-drama, film, and advertising budgets. | 高 | SI010, SI012 |
| CI015 | LibTV reported more than US$1 million of single-day revenue in its first month and May revenue more than 13 times the first-month figure. | 中 | SI011, SI013 |
| CI016 | LibTV was reported to serve nearly 1,000 teams and to have been integrated by more than 300 short-film and film companies within a month of launch. | 中 | SI011, SI013 |
| CI017 | LiblibAI was reported to have more than 30 million users, more than 500,000 original models, and more than 5 million generated images per day by mid-2026. | 高 | SI010, SI011, SI012 |
| CI018 | Public reporting describes the C-side model as freemium access plus subscriptions and on-demand compute or credit payments for professional creators. | 中 | SI024, SI025 |
| CI019 | Public reporting describes the B-side model as API services, enterprise customization, and private-deployment or workflow services for vertical customers. | 中 | SI024, SI019 |
| CI020 | LibTV's reported model combines annual subscription tiers and pay-per-use economics for video generation workloads. | 中 | SI001, SI013, SI027 |
| CI021 | Liblib's official site identifies the product as a leading AI creation platform for the Chinese market. | 高 | SI023, SI010 |
| CI022 | The CAC announced that, by November 1, 2025, 611 generative-AI services and 306 generative-AI applications or functions had completed filing or registration. | 高 | SI022, SI015 |
| CI023 | 36Kr's LiblibAI profile says the company passed a deep-synthesis algorithm filing in February 2024 and became the first AI community to complete the generative-AI service filing in March 2024. | 中 | SI015, SI026 |
| CI024 | AIbase reported in October 2025 that LiblibAI had not yet fully turned profitable. | 中 | SI018, SI020 |
| CI025 | Reporting on the Series B framed capital as a larger barrier in AI applications because compute and traffic acquisition costs make scale expensive. | 中 | SI018, SI021 |
| CI026 | China Biz Insider's adverse analysis states that Evoken does not train proprietary foundation models and operates as a third-party model aggregator. | 中 | SI001, SI012 |
| CI027 | China Biz Insider states that each token consumed on Liblib or LibTV draws on model-provider data centers, leaving Evoken to earn a spread between wholesale API cost and retail subscription revenue. | 中 | SI001, SI024 |
| CI028 | China Biz Insider links LibTV's surge partly to competitor pricing changes and warns that below-market consumer pricing may require supplier discounts or direct subsidy absorption. | 中 | SI001, SI013 |
| CI029 | China Biz Insider reports investor concern that foundation-model providers could reduce API prices or release native applications that directly substitute for aggregators like Evoken. | 中 | SI001, SI012 |
| CI030 | A US$2 billion valuation on US$300 million ARR implies an approximate 6.7 times ARR multiple. | 高 | SI001, SI010, SI012 |
| CI031 | ElectroIQ and GetLatka both describe Midjourney revenue reaching roughly US$500 million in 2025. | 中 | SI006, SI007 |
| CI032 | GetLatka lists Midjourney's most recent disclosed valuation at US$10 billion. | 中 | SI006, SI007 |
| CI033 | Sacra and TechCrunch report Canva at about US$4 billion of ARR with a US$42 billion valuation reference. | 高 | SI003, SI005 |
| CI034 | Sacra's February 2026 research cites Canva at US$4 billion ARR and a US$65 billion late-2025 secondary valuation, and cites Gamma at US$102 million ARR valued at US$2.1 billion. | 高 | SI004, SI003 |
| CI035 | Sacra's Canva dataset says Canva has operated profitably for seven years, a contrast with Liblib's still-private and not-yet-profitable profile. | 高 | SI003, SI004, SI018 |
| CI036 | ARR Club tracks verified revenue and growth intelligence for AI products and lists Canva among leading ARR examples. | 中 | SI002, SI003 |
| CI037 | DBS and Granite Asia closed a US$110 million AI-focused IPO fund in February 2026 for Asian startup exposure. | 中 | SI008, SI009 |
| CI038 | Migrant Times reports that Granite Asia has US$10 billion in assets under management and co-managed capital. | 中 | SI008, SI009 |
| CI039 | CB Insights profiles Granite Asia within the Singapore technology venture ecosystem. | 中 | SI009, SI008 |
| CI040 | The Series B+ included follow-on participation from existing investors including HSG or Sequoia China, Gaorong, and Ant Group. | 中 | SI010, SI011, SI012 |
| CI041 | The investor roster spans financial VCs, Chinese strategic platforms, and prior AI-application backers, giving the company both capital access and potential distribution help. | 中 | SI010, SI019, SI008 |
| CI042 | Liblib's roughly 6.7 times ARR multiple is below Canva's cited 10.5 times to 16 times range and below Gamma's cited 20.6 times multiple. | 高 | SI001, SI003, SI004 |
| CI043 | No cited public source discloses Liblib's exact gross margin, wholesale model-API cost, or contribution margin by product line. | 低 | |
| CI044 | No cited public source discloses Liblib's post-B+ cash balance, monthly burn rate, debt obligations, or runway in months. | 低 | |
| CI045 | The June 2026 financing materially reduces near-term financing risk but raises the execution hurdle for global expansion and portfolio build-out. | 中 | SI010, SI011, SI018 |
| CI046 | Liblib's product matrix can reuse creator traffic, models, and workflows across LiblibAI, Xingliu, and LibTV, which may lower effective customer-acquisition cost if retention holds. | 中 | SI011, SI013, SI014 |
| CI047 | Revenue quality is stronger when tied to professional production budgets than when tied only to hobbyist image generation, but public reporting does not split revenue by stream. | 中 | SI010, SI013, SI024 |
| CI048 | China Biz Insider reports an April 2026 CCTV Finance content-compliance incident at LiblibAI, creating a non-financial risk to monetization continuity. | 中 | SI001, SI016 |
| CE001 | LiblibAI is reported as one of China’s largest AI image-resource libraries and creator communities, with more than 30 million cumulative users. | 高 | SE009, SE010, SE011 |
| CE002 | Public 2026 coverage reports that LiblibAI has accumulated more than 500,000 original models. | 中 | SE010, SE011 |
| CE003 | KuCoin reports that LiblibAI generates more than 5 million images per day. | 中 | SE011, SE010 |
| CE004 | LiblibAI 2.0 is described as a shift from a simple collection of models and tools to a professional AI studio for creators. | 中 | SE008 |
| CE005 | LiblibAI 2.0 integrates image generation, video generation, open-source models, closed-source models, and more than 500 visual effects. | 中 | SE008 |
| CE006 | Evoken’s 2026 product group includes the LiblibAI image community, LibTV video creation platform, and Xingliu design agent. | 高 | SE009, SE010 |
| CE007 | Yicai reports that LibTV was launched in March 2026 as an AI video-creation platform. | 中 | SE009, SE010 |
| CE008 | China Biz Insider reports that LibTV aggregates mainstream image and video generation models, including ByteDance Seedance 2.0 through Volcano Engine. | 中 | SE021 |
| CE009 | LibTV is reported to have exceeded US$1 million of single-day revenue in its first month and to have served nearly 1,000 teams. | 中 | SE010, SE021, SE012 |
| CE010 | Xingliu is reported as an AI design agent with more than 10 million users. | 中 | SE009, SE010 |
| CE011 | China Biz Insider reports that Lovart is an overseas-facing AI design agent that reached about US$80 million ARR five months after launch. | 中 | SE021 |
| CE012 | Evoken is reported not to train its own proprietary foundation models, making it an application-layer aggregator rather than a foundation-model company. | 中 | SE021, SE011 |
| CE013 | The favorable product thesis is that creator relationships, industry data, assets, and production workflows are harder to replicate than the replaceable underlying model. | 中 | SE012, SE021 |
| CE014 | LiblibAI is described by developer-oriented coverage as a central hub for Stable Diffusion users rather than only a gallery. | 中 | SE004, SE018 |
| CE015 | Public developer coverage says the LiblibAI API surface supports text-to-image, image-to-image, ADetailer, ControlNet, and LoRA training workflows. | 中 | SE004 |
| CE016 | Developer coverage describes LiblibAI API integration as including ComfyUI online and programmatic triggering of complex node-based workflows. | 中 | SE004 |
| CE017 | GPTProto reports that non-China users face a Chinese mobile ID registration barrier for full LiblibAI API access. | 中 | SE004 |
| CE018 | ComfyUI is described as a node-based DAG editor where checkpoint loading, prompt encoding, sampling, VAE decoding, and post-processing are explicit swappable nodes. | 中 | SE002 |
| CE019 | Tech-Insider reports that 2025–2026 model families such as FLUX.1 and Stable Diffusion 3.5 publish ComfyUI workflows as a direct way to run them. | 中 | SE002 |
| CE020 | A public ComfyUI tutorial covers SDXL, Stable Diffusion 3.5 Large, FLUX.1, LoRA, ControlNet, image-to-image, and an API server in one workflow stack. | 中 | SE002 |
| CE021 | Runflow describes the ComfyUI API as HTTP and WebSocket infrastructure for submitting generation workflows, uploading inputs, tracking execution, and retrieving outputs. | 中 | SE003 |
| CE022 | Runflow identifies /prompt, /history, /view, image upload, queue, and WebSocket status events as core ComfyUI integration primitives. | 中 | SE003 |
| CE023 | Runflow warns that a default ComfyUI server lacks a built-in authentication flag, so production deployments must add auth, queueing, scaling, and observability controls. | 中 | SE003 |
| CE024 | Apatero states that a character LoRA training workflow typically uses 15 to 50 high-quality images, captions, and 1,000 to 3,000 training steps. | 中 | SE001 |
| CE025 | Apatero states that dedicated LoRA training encodes character knowledge into model weights and can outperform quick reference-based consistency methods for repeated use. | 中 | SE001 |
| CE026 | RunComfy describes Lora-Training-in-Comfy as a ComfyUI extension that lets artists train LoRA models directly within ComfyUI. | 中 | SE005 |
| CE027 | RunComfy states that LoRA training in ComfyUI involves image-caption data preparation, path configuration, queued training, and immediate testing from the LoRA folder. | 中 | SE005 |
| CE028 | Stable Diffusion Art explains that LoRA adapters are commonly much smaller than full checkpoint models and usually require a base checkpoint. | 中 | SE006 |
| CE029 | Stable Diffusion Art explains that LoRA fine-tunes cross-attention layers by decomposing large matrices into low-rank matrices. | 中 | SE006 |
| CE030 | DataCamp describes Stable Diffusion as an open-source diffusion model for text-to-image generation, image modification, and image enhancement. | 中 | SE007 |
| CE031 | DataCamp describes Stable Diffusion 3 as using a Multimodal Diffusion Transformer architecture with separate image and language weights. | 中 | SE007 |
| CE032 | CAC filing evidence and Baidu Baike background support that LiblibAI operates in a Chinese regulated AIGC environment requiring filings and controls. | 高 | SE015, SE014 |
| CE033 | Baidu Baike records a CCTV-identified content-safety incident in which LiblibAI could be manipulated by specific prompts to generate inappropriate content, followed by technical rectification. | 中 | SE014 |
| CE034 | BestHub describes LiblibAI’s core loop as creators uploading, training, and sharing LoRA models and workflows while ordinary users reuse those resources. | 中 | SE012 |
| CE035 | The creator flywheel is that more users create richer models and assets, which then attract more creators and regular users. | 中 | SE012, SE018 |
| CE036 | 6Wresearch’s China AI-image-generator landscape includes large technology competitors such as Baidu, Alibaba, Tencent, and ByteDance-linked platforms. | 中 | SE019 |
| CE037 | Sharewalker compares Liblib.Art with Tensor.Art, SeaArt.AI, and AituBo.AI as peer AI-art generation platforms. | 中 | SE020 |
| CE038 | Skywork’s SeaArt review provides evidence that competing image platforms also offer broad AI-art generation capabilities. | 中 | SE025 |
| CE039 | Analyst-market sources indicate that China’s generative-AI market is a large and growing demand pool for AIGC applications. | 高 | SE022, SE023 |
| CE040 | The Next Web reports that China’s micro-drama sector is becoming a mass market for AI video, supporting demand for LibTV-like production workflows. | 高 | SE024, SE021 |
| CE041 | China Biz Insider frames the key technology risk as Evoken earning a spread between retail pricing and wholesale third-party model or compute costs. | 中 | SE021 |
| CE042 | If model owners lower prices, broaden direct APIs, or release native applications, Evoken’s workflow features could face substitution pressure. | 中 | SE021, SE019 |
| CE043 | Public evidence does not disclose Evoken’s model-provider contracts, wholesale API prices, uptime architecture, security controls, or content-filter remediation details. | 低 | |
| CE044 | The official Liblib site positions the product as a leading Chinese AI creation platform, which is directionally corroborated by independent guide coverage. | 高 | SE016, SE018 |
| CU001 | LiblibAI had more than 30 million cumulative users by mid-2026. | 高 | SU009, SU010, SU011, SU013 |
| CU002 | Late-2025 reporting put Liblib at about 4 million monthly active users and 25 million total users. | 中 | SU016 |
| CU003 | Liblib's creator base exceeded 20 million AI creators by 2025-2026, with source language ranging from 20 million creators to more than 30 million users in the broader ecosystem. | 中 | SU001, SU018, SU012 |
| CU004 | The platform had more than 500,000 original models and generated more than 5 million images per day by mid-2026. | 中 | SU010, SU011 |
| CU005 | Chinese reporting says one in three Chinese designers uses LiblibAI for creation, positioning designers as a core professional creator segment. | 中 | SU013 |
| CU006 | Jiemian reported more than 500,000 user-trained original AI models/workflows and more than 500 million cumulative images, showing depth beyond a logo-only customer story. | 中 | SU001 |
| CU007 | Liblib operates a two-sided creative platform: individual creators supply models, workflows and usage, while enterprises buy stable services, APIs or customized AI-image scenarios. | 中 | SU001, SU016, SU017 |
| CU008 | Jiemian named Kingsoft Office/WPS, Wondershare, G-bits Games, Tmall Campus and Tsinghua University as LiblibAI B-side customers. | 中 | SU001 |
| CU009 | Kingsoft Office/WPS is a named B-side customer proof point for office-product and productivity-design workflows, but the public source does not disclose deployment scale or outcomes. | 中 | SU001 |
| CU010 | Wondershare is a named B-side customer proof point for creative-software and video/design workflows, with deployment depth not disclosed in public reporting. | 中 | SU001 |
| CU011 | G-bits Games is a named B-side customer proof point for game-art or game-marketing imagery, consistent with broader gaming-sector demand for AI production tools. | 中 | SU001, SU008 |
| CU012 | Tmall Campus is a named B-side customer proof point for e-commerce or campus-market visual content, but public sources do not quantify production usage. | 中 | SU001 |
| CU013 | Tsinghua University is a named B-side customer proof point for education or research-community usage, but public sources do not say whether it is a paid enterprise deployment. | 中 | SU001 |
| CU014 | Jiemian characterizes the B-side offering as professional AI-image scenario solutions rather than mere logo affiliation. | 中 | SU001 |
| CU015 | LibTV extended Liblib's customer surface from images into video workflows for creators, studios, brands and film teams. | 中 | SU010, SU012, SU013 |
| CU016 | LibTV reportedly exceeded US$1 million of single-day revenue in its first month, and May revenue was more than 13 times first-month revenue. | 中 | SU010, SU012 |
| CU017 | Within a month of launch, more than 300 short-film and film companies integrated LibTV, and the platform served nearly 1,000 agencies, production houses and brand clients. | 中 | SU010, SU012 |
| CU018 | LibTV attracted 100,000 creators on launch day and used a RMB10 million creator cash pool, suggesting a subsidy-assisted launch loop. | 中 | SU012, SU010 |
| CU019 | Professional content producers in short dramas, films and advertising are cited as the core source of Evoken's current revenue. | 高 | SU009, SU010, SU012 |
| CU020 | miHoYo's RMB100 billion three-year AI commitment is a demand-side signal that large Chinese game studios are treating AI as production infrastructure. | 中 | SU003, SU004 |
| CU021 | miHoYo has already deployed production AI systems in active titles, including the Pamu Helper for Honkai: Star Rail and an AI NPC system for Starry Valley. | 中 | SU003 |
| CU022 | NetEase, Tencent and other Chinese game companies are integrating generative AI into asset creation, NPCs, customer service and user-generated environments. | 中 | SU005, SU008, SU002 |
| CU023 | WebProNews cites China gaming revenue of US$50.7 billion in 2025, 722 million players, and more than 60% of Chinese developers using generative AI, supporting gaming as a large vertical demand pool. | 中 | SU008 |
| CU024 | China's micro-drama market is projected to exceed RMB120 billion in 2026, giving LibTV a large production-workflow market to sell into. | 高 | SU023, SU024, SU025 |
| CU025 | The overseas micro-drama market generated an estimated US$229 million of IAP revenue in May 2026, with AI-native app VibeShort entering the top revenue ranks. | 中 | SU006 |
| CU026 | Tech Times reports 470 AI-produced short-drama titles per day in January 2026 and roughly 50,000 AI-native Douyin episodes in March, creating both demand for production tools and a content-oversupply risk. | 中 | SU007 |
| CU027 | Tech Times frames labor displacement as current rather than hypothetical, citing halted live-action production, canceled writing projects and the phrase that jobs were being eliminated. | 中 | SU007 |
| CU028 | The public GTM model runs from free basic use to subscriptions and pay-as-needed credits, with enterprise customers paying for API services and customized needs. | 中 | SU016, SU017 |
| CU029 | Hubpy describes LiblibAI's free tier as limited daily generations, with higher-volume usage and priority GPU access requiring subscriptions or credits. | 中 | SU017 |
| CU030 | Public developer commentary says LiblibAI API availability exists, but international setup and phone verification can create developer friction. | 中 | SU019 |
| CU031 | Evoken's ARR exceeded US$300 million by May 2026, implying that customer willingness to pay extends beyond free creator usage. | 高 | SU009, SU010, SU012, SU013 |
| CU032 | Tencent News reported Liblib's 2024 full-year revenue at RMB206 million, providing an earlier monetization baseline before the 2026 ARR step-up. | 中 | SU016 |
| CU033 | BestHub reports that 80% of daily new users came from organic traffic, a customer-acquisition signal that lowers dependence on paid acquisition if durable. | 中 | SU012 |
| CU034 | Tencent News says Evoken's product matrix lowers switching cost for designers across image, video and design-agent workflows, supporting cross-sell potential. | 中 | SU013 |
| CU035 | KuCoin reports that new products can draw on the existing community for creators, models, assets and distribution channels rather than starting user acquisition from zero. | 中 | SU011 |
| CU036 | The most evidence-backed segment split is C-side individual creators and prosumers versus B-side enterprises, studios and API/custom-solution buyers. | 中 | SU001, SU016, SU017, SU012 |
| CU037 | No retained public source discloses NRR, GRR, churn, contract length, renewal rate or customer-concentration metrics for Liblib or LibTV. | 中 | SU001, SU012, SU016 |
| CU038 | Durability is indirectly supported by daily image generation, creator incentives, model/workflow supply and LibTV team adoption, but these are retention proxies rather than cohort-retention data. | 中 | SU001, SU010, SU012, SU018 |
| CU039 | Public customer evidence supports gaming, short-drama/film, e-commerce, advertising/brand video, education and professional design as customer verticals. | 中 | SU001, SU008, SU006, SU012, SU013 |
| CU040 | Public reporting does not identify top customers by revenue share, so customer concentration remains a material private-data diligence item. | 中 | SU001, SU012, SU016 |
| CU041 | Public sources verify API and customized enterprise services, but not private-deployment terms, service-level agreements or on-premise security posture. | 中 | SU016, SU019 |
| CU042 | Independent market reports support continued expansion of China's generative-AI and AI-image markets, giving Liblib a favorable macro adoption backdrop. | 高 | SU020, SU021, SU022 |
| CU043 | The same market reports identify large technology platforms as AI-image competitors, so customer acquisition and retention cannot be assumed from category growth alone. | 中 | SU022, SU020 |
| CU044 | LiblibAI 2.0 promotions and more than RMB10 million in creator incentives show that subsidies and free compute are part of acquisition and activation, not just paid conversion. | 中 | SU018, SU012 |
| CU045 | Liblib's model-sharing and LoRA-training loop creates self-expanding supply: more creators add models and assets, which can attract more creators and enterprise use cases. | 中 | SU011, SU017, SU001 |
| CU046 | The AI short-drama boom carries reputational risk because real-person likeness reuse and actor-committee restrictions could increase buyer caution around AI-generated creative assets. | 中 | SU007 |
| CU047 | LibTV's agency, production-house and brand-client adoption is stronger evidence of B-side willingness to pay than a named-logo list alone, because it includes revenue and team-count signals. | 中 | SU010, SU012 |
| CU048 | The named enterprise-customer list is sample proof rather than an exhaustive customer roster; the public article gives examples but not a complete account list. | 中 | SU001 |
| CR001 | China's public-facing generative AI services have been governed by the Interim Measures for the Management of Generative AI Services since August 15, 2023. | 高 | SR005, SR006 |
| CR002 | The Deep Synthesis Provisions became effective in January 2023 and regulate AI-generated text, images, voice, and video. | 高 | SR006, SR008 |
| CR003 | The 2025 Measures for Labeling of AI-Generated Synthetic Content apply to service providers generating or synthesizing text, images, audio, video, virtual scenes, or other information. | 高 | SR001, SR002 |
| CR004 | The AIGC labeling regime requires both explicit labels perceivable by users and implicit labels embedded in file metadata or technical measures. | 高 | SR001, SR002, SR007 |
| CR005 | The labeling Measures and mandatory national standard GB 45438-2025 took effect on September 1, 2025. | 高 | SR001, SR002, SR003, SR007 |
| CR006 | App distribution platforms must check whether applications provide generative AI services and review materials related to generated-content labeling. | 高 | SR001, SR003 |
| CR007 | China's AI regime remains multi-instrument rather than a single enacted comprehensive AI statute, with a draft comprehensive AI law first officially published in December 2025 and uncertain enactment timing. | 高 | SR006, SR007 |
| CR008 | As of November 1, 2025, CAC reported 611 generative-AI services filed and 306 applications or functions registered, evidencing an ongoing registry regime. | 高 | SR023, SR006 |
| CR009 | 36Kr reports that LiblibAI passed CAC's fourth batch of deep-synthesis service algorithm filing in February 2024. | 中 | SR022, SR020 |
| CR010 | 36Kr reports that LiblibAI became China's first AI-community platform to complete filing under the Interim Measures for Generative AI Services in March 2024. | 高 | SR022, SR023 |
| CR011 | Filing evidence lowers Liblib's launch-continuity risk but does not eliminate later labeling, content-safety, or registry-update obligations. | 中 | SR001, SR003, SR022, SR023 |
| CR012 | Baidu Baike records a 2026 CCTV exposure alleging that LiblibAI could bypass moderation with specific prompts to generate inappropriate pornographic content. | 中 | SR020, SR027 |
| CR013 | LiblibAI reportedly started internal self-inspection and technical rectification after the CCTV exposure, including fixes to identified problems and optimization of review strategy and model capabilities. | 中 | SR020, SR027 |
| CR014 | The 2025 CAC enforcement agenda targeted AI-service abuses including non-compliant AI products, weak security measures, AI-generated rumors, false information, and minors' rights violations. | 高 | SR002, SR007 |
| CR015 | The Beijing Internet Court recognized copyright in an AI-generated image where the human user selected many prompts, arranged them, and adjusted parameters to align output with the user's conception. | 高 | SR004, SR009 |
| CR016 | Chinese AI-image copyright protection is fact-specific and depends on evidence of human creative effort, prompt process, selection, and modification rather than mere machine output. | 高 | SR004, SR009, SR010 |
| CR017 | China IP Law Update reports a later Beijing Internet Court case denying an AI-image copyright claim where the plaintiff lacked original generation-process records and relied on after-the-fact recreation. | 高 | SR010, SR004 |
| CR018 | In Ultraman v Acgnai, users uploaded Ultraman images to train LoRA models that other users could apply to generate substantially similar Ultraman-style images. | 高 | SR011, SR012, SR013 |
| CR019 | The Hangzhou courts treated training-stage use more leniently as potentially fair use while applying stricter scrutiny to output generation and dissemination. | 高 | SR011, SR012, SR016 |
| CR020 | The Hangzhou Internet Court held the AI platform contributorily liable because it should have known users were infringing and failed to take necessary preventive measures. | 高 | SR012, SR013, SR014 |
| CR021 | The Ultraman LoRA case resulted in an order to cease infringing activity and pay RMB 30,000 in economic losses and reasonable expenses. | 高 | SR013, SR014, SR016 |
| CR022 | The Ultraman rulings indicate that a commercial AI platform's duty of care rises with profit model, promoted infringing materials, identifiability of stable outputs, and availability of preventive controls. | 高 | SR012, SR014, SR016 |
| CR023 | For a LoRA model-sharing platform like Liblib, the direct analog risk is user training or publishing of models that reliably reproduce protected characters or art styles. | 中 | SR012, SR013, SR031 |
| CR024 | Recommended mitigation from the AIGC infringement cases includes complaint channels, user risk notices, IP review mechanisms, prompt/model blocking, generated-content labeling, and timely takedown. | 高 | SR014, SR015, SR016 |
| CR025 | LiblibAI is reported to operate a large creator community with users uploading, training, and sharing LoRA models and workflows. | 中 | SR027, SR031, SR021 |
| CR026 | Yicai reports that Evoken reached US$300 million ARR as of May 2026, more than 30 times the prior-year level. | 高 | SR017, SR019, SR027 |
| CR027 | AIbase reported in October 2025 that LiblibAI had not yet turned a profit despite accelerating development. | 中 | SR024, SR025 |
| CR028 | AIbase reported that LiblibAI was once forced offline because it had not completed large-model filing in its early stage. | 中 | SR024, SR022 |
| CR029 | Evoken's June 2026 B+ round of nearly US$300 million was intended to support R&D, global expansion, AI creative product capability, and product-portfolio build-out. | 高 | SR017, SR018, SR019 |
| CR030 | China Biz Insider reports that Evoken does not train proprietary foundation models and builds LiblibAI, Lovart, and LibTV on third-party model APIs. | 高 | SR027, SR017 |
| CR031 | Evoken's model-resale economics depend on the spread between wholesale model or compute costs and retail subscriptions or credits. | 中 | SR027, SR024 |
| CR032 | Aggregator exposure can become acute if first-party model providers lower prices, bundle workflows, restrict APIs, or ship native applications that substitute for Liblib's interface. | 中 | SR027, SR029 |
| CR033 | LibTV's reported growth partly benefited from ByteDance Jimeng pricing changes, making model-provider pricing and queue-time dynamics a monitorable dependency. | 中 | SR027, SR021 |
| CR034 | LiblibAI's platform scale exceeded 30 million cumulative users and 500,000 original models by mid-2026. | 高 | SR017, SR018, SR019, SR027 |
| CR035 | Large community scale increases moderation, IP-review, labeling, and takedown operating load because more creators, models, and outputs expand the surface area for violations. | 中 | SR020, SR027, SR031 |
| CR036 | Enterprise and professional customers cited for Liblib include Kingsoft Office, G-bits Games, Tmall Campus, Wondershare, and Tsinghua University. | 中 | SR028, SR020 |
| CR037 | B-side customer proof raises the cost of compliance failures because regulated enterprise buyers are more likely to require stable legality, labeling, and auditability. | 中 | SR028, SR001, SR014 |
| CR038 | Chen Mian is reported as founder and CEO with prior ByteDance CapCut commercialization experience, making his product judgment central to the investment thesis. | 高 | SR017, SR021, SR020 |
| CR039 | The master source pack reports Chen Mian holds about 73.95% of the company, creating founder-control concentration that investors should diligence through governance documents. | 中 | SR020, SR021 |
| CR040 | The master source pack reports the legal representative changed to Yang Nan on November 22, 2025, while Chen Mian remained founder and controlling shareholder. | 中 | SR020 |
| CR041 | The legal-representative change is a governance diligence item rather than proof of misconduct because public sources do not connect it to enforcement, litigation, or loss of founder control. | 中 | SR020, SR021 |
| CR042 | China's short-drama AI boom has produced labor-displacement and content-glut concerns that could provoke customer, creator, or regulator backlash. | 中 | SR029, SR030 |
| CR043 | ByteDance Seedance and other third-party video models create geopolitical and export-access exposure because Liblib's product quality can depend on model availability, compliance filters, and cross-border launch decisions. | 中 | SR027, SR029 |
| CR044 | Linklaters warns that divergent AIGC labeling rules across jurisdictions may raise technical development and compliance costs for businesses engaged in cross-border activity. | 高 | SR007, SR003 |
| CR045 | The highest residual risks for Liblib are content-safety enforcement, copyright liability from user LoRA models, model-provider repricing or substitution, and unverified gross margin. | 中 | SR012, SR013, SR020, SR024, SR027 |
| CR046 | Mitigations should be monitored through repeat CAC filing visibility, labeling implementation, moderation red-team pass rates, IP takedown response times, wholesale model-cost schedules, and product-level gross margin. | 中 | SR001, SR014, SR016, SR023, SR027 |
| CR047 | A thesis-break event would be a material CAC enforcement action, repeated pornography or minors-safety failure, injunction against core LoRA workflows, API access restriction, or evidence that gross margin remains structurally negative at scale. | 中 | SR002, SR013, SR014, SR024, SR027 |
| CR048 | The 2026 draft and standards environment makes risk management a continuous operating function rather than a one-time pre-launch checklist. | 中 | SR006, SR007, SR023 |
| CV001 | Evoken completed a Series B+ round of approximately US$300 million in June 2026. | 高 | SV012, SV013, SV014, SV016 |
| CV002 | The June 2026 Series B+ round valued Evoken at more than US$2 billion post-money. | 高 | SV012, SV013, SV014, SV016 |
| CV003 | The reported May 2026 ARR base for Evoken was about US$300 million. | 高 | SV012, SV013, SV014, SV015 |
| CV004 | A US$2.0 billion valuation on US$300 million ARR implies an ARR multiple of approximately 6.7x. | 高 | SV012, SV022, SV003 |
| CV005 | The round was co-led by Granite Asia, Tencent Holdings and Shunwei Capital. | 高 | SV012, SV013, SV016, SV009 |
| CV006 | HT Investment and Times Capital participated as follow-on investors in the Series B+ round. | 中 | SV012, SV016, SV009 |
| CV007 | Existing shareholders including HSG or Sequoia China, Gaorong and Ant Group increased or continued their investment in the B+ financing. | 高 | SV012, SV013, SV009 |
| CV008 | Evoken raised more than US$500 million across its October 2025 Series B and June 2026 Series B+ rounds within roughly eight months. | 高 | SV012, SV014, SV016 |
| CV009 | The company announced intended use of B+ proceeds for R&D, global expansion, AI creative product capability and portfolio build-out. | 中 | SV012 |
| CV010 | LiblibAI had more than 30 million cumulative users and more than 500,000 original models by mid-2026. | 高 | SV012, SV013, SV014 |
| CV011 | Evoken group revenue grew more than 3,000% year over year by May 2026 according to 2026 reports. | 中 | SV013, SV015, SV016 |
| CV012 | Core revenue was reported to come mainly from professional content producers in short dramas, film and advertising. | 高 | SV012, SV031 |
| CV013 | LibTV reportedly generated more than US$1 million of single-day revenue in its first month. | 中 | SV013, SV015, SV022 |
| CV014 | LibTV served nearly 1,000 teams and had more than 300 short-film or film companies integrate within about a month of launch. | 中 | SV015, SV022, SV031 |
| CV015 | ScaleXP cites a 2026 public-cloud revenue multiple around 6.2x for the BVP Nasdaq Emerging Cloud Index. | 中 | SV003 |
| CV016 | Acquiry’s 2026 private-market ranges put high-growth AI-native SaaS at 10x to 20x ARR and traditional SaaS growth bands mostly lower. | 中 | SV004 |
| CV017 | SaaSRise reports AI-native VC deal multiples materially above legacy SaaS, with AI-native companies at a median 21.2x revenue versus 5.5x for legacy SaaS. | 中 | SV001 |
| CV018 | SaaSRise benchmarks report AI-native companies growing faster than legacy SaaS but with gross margins of 55% to 70% due to inference costs. | 中 | SV002 |
| CV019 | We Are Founders argues that AI wrappers without proprietary models or embedded workflows are seeing multiples compress while vertical AI can trade around 9x to 12x ARR. | 中 | SV005 |
| CV020 | Liblib’s roughly 6.7x ARR multiple sits below top-quartile AI-native software ranges and closer to public or legacy SaaS benchmarks. | 中 | SV003, SV004, SV005, SV022 |
| CV021 | Canva reached about US$4 billion ARR by end-2025 and was valued at US$42 billion to US$65 billion in cited sources. | 高 | SV024, SV025, SV026 |
| CV022 | Canva’s cited valuation range implies an approximate 10.5x to 16.3x ARR multiple on US$4 billion ARR. | 高 | SV024, SV025, SV026 |
| CV023 | Midjourney’s cited US$10 billion valuation and roughly US$500 million 2025 revenue imply about a 20x revenue multiple. | 中 | SV027, SV028 |
| CV024 | SaaSRise lists Runway ML at US$70 million to US$80 million ARR and a US$5.3 billion valuation in February 2026. | 中 | SV001 |
| CV025 | SaaSRise lists ElevenLabs at US$330 million ARR and an US$11 billion valuation in February 2026. | 中 | SV001 |
| CV026 | Sacra’s Canva update lists Gamma at US$102 million ARR and a US$2.1 billion valuation, or a 20.6x multiple. | 中 | SV025 |
| CV027 | Evoken’s multiple is materially lower than cited creative-AI comparables such as Midjourney, Runway, ElevenLabs and Gamma. | 中 | SV001, SV022, SV025, SV027 |
| CV028 | China Biz Insider’s adverse valuation read says Evoken does not train proprietary foundation models and depends on third-party model APIs. | 中 | SV022 |
| CV029 | The aggregator model exposes Evoken to wholesale model pricing, API access and first-party substitution risk. | 中 | SV022, SV005, SV002 |
| CV030 | China Biz Insider calculates the US$2 billion on US$300 million ARR headline as roughly 6.7x revenue and calls it reasonable only if Evoken becomes a creator operating system. | 中 | SV022 |
| CV031 | A base valuation case can support about US$2.1 billion if Evoken sustains roughly US$420 million forward ARR at a 5x ARR multiple after discounting for aggregator risk. | 中 | SV003, SV004, SV022 |
| CV032 | A bull case can support roughly US$4.0 billion if forward ARR reaches US$500 million and workflow lock-in justifies about an 8x ARR multiple. | 中 | SV004, SV005, SV015 |
| CV033 | A bear case implies roughly US$1.2 billion if ARR stalls near US$300 million and model-provider economics compress the multiple to about 4x. | 中 | SV003, SV022, SV008 |
| CV034 | The public evidence does not disclose gross margin, NRR, CAC payback, revenue cohort retention or product-level contribution margin. | 低 | |
| CV035 | The most important private diligence item is whether LibTV and LiblibAI have durable contribution margin after model-API, cloud and creator-subsidy costs. | 中 | SV002, SV022 |
| CV036 | China created 67 new unicorns in the first half of 2026, with AI and robotics dominating the cohort. | 中 | SV011 |
| CV037 | Hurun’s Global Unicorn Index 2026 found China had 381 unicorns and 80 newly minted unicorns in the year, with AI a major value driver. | 中 | SV021 |
| CV038 | China Biz Insider’s adverse embodied-AI note warns that many China AI unicorns face 2027-2028 consolidation, down-round or failure risk because cash runways are short. | 中 | SV008 |
| CV039 | Tech Funding News framed June 2026 AI unicorn formation as capital moving toward applications, infrastructure and autonomous systems beyond foundation models. | 中 | SV009 |
| CV040 | TechCrunch’s 2026 unicorn tracker shows AI-related startups were a large share of newly minted unicorns in the year. | 中 | SV010 |
| CV041 | Granite Asia is an active Singapore-based investor with hundreds of investments and appears as an investor associated with LiblibAI in June 2026 data. | 高 | SV006, SV029 |
| CV042 | Tracxn access for Granite Asia was rate-limited, so it does not independently validate investor details beyond documenting an access gap. | 低 | SV007 |
| CV043 | DBS and Granite Asia launched a US$110 million AI fund in early 2026, reinforcing Granite Asia’s current AI allocation signal. | 中 | SV030 |
| CV044 | CAC filing registry evidence confirms the continuing public filing regime for generative-AI services and applications in China. | 高 | SV020, SV032 |
| CV045 | LiblibAI had previously completed China deep-synthesis and generative-AI service filings according to the shared diligence brief and CAC registry context. | 高 | SV020, SV017 |
| CV046 | Baidu Baike records a 2026 CCTV content-safety exposure naming LiblibAI and subsequent technical rectification, making regulatory risk valuation-relevant. | 中 | SV017, SV032 |
| CV047 | The valuation stance is fair-to-modestly-attractive rather than stretched because the headline multiple is modest but margin and aggregator-risk disclosures are still missing. | 中 | SV003, SV004, SV022, SV008 |
| CV048 | A thesis-break trigger would be evidence that May 2026 ARR was subsidy-led, non-recurring, or concentrated in low-margin usage rather than durable professional workflows. | 中 | SV015, SV022, SV002 |
| CV049 | A valuation upside trigger would be audited retention, gross margin and product-level ARR evidence showing Evoken is a workflow system of record rather than a model reseller. | 中 | SV002, SV003, SV022 |
| CV050 | The reasoned investment view is that the greater-than-US$2 billion valuation is justified on public ARR and growth evidence, but only conditionally until margin and retention diligence are completed. | 中 | SV012, SV015, SV022, SV004, SV008 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Yicai Global | LiblibAI Parent Evoken Valued at Over USD2 Billion After New Funding Round | Evoken, the parent company of a major Chinese artificial intelligence image creation and sharing platform LiblibAI, has completed a USD300 million Series B+ funding round, valuing the company at more than USD2 billion. |
| SO002 | AIbase | 演语科技(Evoken)完成近3亿美元B+轮融资,ARR达3亿美元,AI应用层进入商业化阶段 | As of May 2026, the company's Annual Recurring Revenue (ARR) has exceeded $300 million. |
| SO003 | KuCoin News | AI Application Layer Unicorn LiblibAI Secures $300M B+ Round Valued at Over $2.0B | According to the company, LiblibAI has accumulated over 30 million users, more than 500,000 original models, and over 100 million professional images and video assets, generating more than 5 million images daily. |
| SO004 | BestHub | How LiblibAI's $300M ARR fueled a near $300M funding round valuing it at over $2B | Yanyu’s annual recurring revenue (ARR) has reached $300 million, and its group revenue grew over 3000% year-over-year by May 2026. |
| SO005 | Baidu Baike (EN) | LiblibAI | LiblibAI (LiblibAI) is an AI image generation platform founded in March 2023, focusing on the creation and sharing of AI painting original models. |
| SO006 | Tencent News | 问AI · 创始人陈冕的字节经历如何助力公司破圈? | 6 月 18 日,演语科技(Evoken)官宣近日完成近 3 亿美元 B+ 轮融资,估值超 20 亿美元. |
| SO007 | 36Kr PitchHub | LiblibAI project profile | LLiblibAI 于 2023 年 5 月创立,是一家 AI 原生应用公司,致力于 AI 内容的创作和分享. |
| SO008 | X-Techcon | 演语科技完成近3亿美元B+轮融资 | 截至2026年5月,演语科技年度经常性收入(ARR)达到3亿美元. |
| SO009 | Firecat Daily AI | LiblibAI secures nearly $300M B+ round | LiblibAI, a unicorn in the AI application layer and parent company of Yanyu Technology, has completed a B+ round of financing of nearly $300 million. |
| SO010 | Sohu | LiblibAI哩布哩布AI一年内完成三轮融资 | 「LiblibAI哩布哩布AI」成立于2023年5月,作为一家AI原生应用公司. |
| SO011 | Sohu | Liblib母公司演语科技完成近3亿美元B+轮融资 | 截至目前,LiblibAI累计用户超过3000万,沉淀超过50万个原创模型和过亿张专业图片、视频素材,日均生成图片超过500万次. |
| SO012 | Baidu Baike | LiblibAI | 2026年4月13日,央视曝光多款AI应用存在涉黄生成漏洞,其中提及“哩布哩布AI”应用在特定提示词下可绕过审核机制生成不当内容. |
| SO013 | AIbase | LiblibAI completes $130 million Series B financing | LiblibAI recently announced the successful completion of a $130 million Series B funding round. |
| SO014 | AIbase | LiblibAI Series B financing and competitive pressure | Although the company has not yet turned a profit, it is clearly accelerating its development to cope with the competitive market environment. |
| SO015 | INCE Capital | INCE Capital’s Portfolio Company LiblibAI Announces $130 Million Series B Financing | LiblibAI, an AI application company invested in by INCE Capital, announced the closing of a $130 million Series B financing recently. |
| SO016 | 36Kr | 暗涌Waves独家:LiblibAI完成1.3亿美元B轮融资 | LiblibAI已于近期完成1.3亿美元B轮融资,由红杉中国、CMC资本及一大厂战投联合领投. |
| SO017 | Sina Finance | LiblibAI完成1.3亿美元B轮融资 | 据接近Liblib的人士透露,目前Liblib的月活跃用户为400万,总用户数2500万. |
| SO018 | Baidu Baike | 北京奇点星宇科技有限公司 | 北京奇点星宇科技有限公司于2023年05月16日成立。法定代表人张子捷. |
| SO019 | Cyberspace Administration of China | 国家互联网信息办公室关于发布生成式人工智能服务已备案信息的公告 | 截至11月1日,累计有611款生成式人工智能服务完成备案,306款生成式人工智能应用或功能完成登记. |
| SO020 | Liblib (company) | LiblibAI-哩布哩布AI - 中国领先的AI创作平台 | LiblibAI-哩布哩布AI - 中国领先的AI创作平台 |
| SO021 | Tencent News | LiblibAI如何避免大模型技术吞噬其生存空间? | 据接近Liblib的人士透露,目前Liblib的月活跃用户为400万,总用户数2500万. |
| SO022 | TMTPost | AI图像生成平台LiblibAI宣布一年内完成三轮融资 | LiblibAI 成立于 2023 年 5 月,是国内最早诞生的一批AI图像生成平台,其核心成员毕业于清华大学、北京大学、卡内基梅隆大学等国内外名校. |
| SO023 | Sina Finance | AI图像生成平台LiblibAI宣布在过去一年内连续完成四轮融资 | 天使轮由源码资本、高榕创投和金沙江创投投资;第二轮由战略投资方领投;最新一轮由明势资本领投. |
| SO024 | Hubpy | LiblibAI Guide 2026 | LiblibAI (哩布哩布) is China's largest AI image generation platform and model sharing community. |
| SO025 | Hurun Report | Global Unicorn Index 2026 | Hurun Research Institute found 1603 unicorns in the world, based in 52 countries and 299 cities. |
| SM001 | IMARC Group | China Generative AI (AIGC) Market Size, Share, Trends and Forecast by Component, Deployment Mode, Technology, and Region, 2026-2034 | The China generative AI (AIGC) market size reached USD 5,160.82 Million in 2025 and is projected to reach USD 19,558.29 Million by 2034. |
| SM002 | Grand View Research | China Generative AI Market Horizon Databook | Software was the largest segment with a revenue share of 63.9% in 2024. |
| SM003 | 6Wresearch | Prominent Companies in China AI Image Generator Market | The industry is highly concentrated among top ecosystem leaders, with Baidu, Alibaba, and Tencent collectively holding a sizable market share. |
| SM004 | DigiTrendz | China’s $16.5B Micro-Drama Industry Adopts AI Video First | China’s micro-drama industry is projected to exceed 120 billion yuan ($16.5 billion) in 2026. |
| SM005 | The Next Web | China’s micro-drama industry is becoming the first mass market for AI video | More than 50,000 AI-native titles hit Douyin in March 2026 alone, at one-tenth the cost of live-action production. |
| SM006 | Startup Fortune | China’s AI micro-drama boom is a business model worth watching | Lower costs create their own problem: when the price of making content falls sharply, the market gets flooded with similar stories. |
| SM007 | Wonford | AI Short Dramas Boom: Technology Reshapes a 100 Billion Yuan Market as 2026’s Tech New Frontier | China’s short drama export revenue reached 2.38 billion US dollars, a year-on-year increase of 263%. |
| SM008 | CRI (China Radio Int'l) | AI-generated micro dramas surge in China as production costs fall | AI-generated comic-style micro dramas accounted for an estimated 16.8 billion yuan in market share in 2025. |
| SM009 | AInChina | China AI Drama Revolution: ByteDance’s $650M Empire in 2026 | The AI short drama market exploded from a $100 million niche to a $650 million industry during Q1 2026. |
| SM010 | Yicai Global | LiblibAI Parent Evoken Valued at Over USD2 Billion After New Funding Round | Evoken's LiblibAI platform has grown into one of China's largest AI image-resource libraries and creator communities, with more than 30 million cumulative users. |
| SM011 | AIbase | AI Application Layer Unicorn LiblibAI Secures $300M in B+ Round Valued at Over $2B | As of May 2026, the company's Annual Recurring Revenue (ARR) has exceeded $300 million. |
| SM012 | KuCoin News | AI Application Layer Unicorn LiblibAI Secures $300M in B+ Round Valued at Over $2B | LiblibAI has amassed over 30 million users and more than 500,000 original models, generating over 5 million images daily. |
| SM013 | BestHub | How LiblibAI’s $300M ARR fueled a near-$300M funding round valuing it at over $2B | Yanyu’s annual recurring revenue (ARR) has reached $300 million, and its group revenue grew over 3000% year-over-year by May 2026. |
| SM014 | Baidu Baike (EN) | LiblibAI | |
| SM015 | Tencent News | 演语科技完成近3亿美元B+轮融资,估值超20亿美元 | AI 剧火爆拉动了对相关模型和工具的需求。 |
| SM016 | 36Kr PitchHub | LiblibAI company profile | Feb 2024 passed the deep-synthesis algorithm filing and Mar 2024 completed generative-AI service filing. |
| SM017 | Baidu Baike | LiblibAI | CCTV reported that LiblibAI could be induced by certain prompts to generate inappropriate content and the company undertook rectification. |
| SM018 | AIbase | LiblibAI Completes $130M Series B Financing | The AI investment boom is shifting from the underlying models to the application layer. |
| SM019 | AIbase | LiblibAI Series B financing and early filing challenges | Due to not completing the large model filing, LiblibAI was once forced to be removed and faced a shortage of funds. |
| SM020 | INCE Capital | INCE Capital portfolio news: LiblibAI financing | |
| SM021 | 36Kr | LiblibAI completes Series B financing | |
| SM022 | Cyberspace Administration of China | Generative artificial intelligence service filing information announcement | By November 2025, 611 generative-AI services had completed filing cumulatively. |
| SM023 | Tencent News | LiblibAI October 2025 financing and operating metrics | 2024 full-year revenue was RMB 206 million. |
| SM024 | Hubpy | LiblibAI Guide 2026 | |
| SM025 | Hurun Report | Hurun Global Unicorn Index 2026 | |
| SP001 | GeniusFirms | Civitai vs. Midjourney: which is better? | Midjourney is better for beginners and fast, high-quality image generation; Civitai is better for advanced users needing control, custom models, and lower long-term cost. |
| SP002 | Creati.ai | Civitai vs Midjourney Comprehensive Comparison: AI Art Platforms | Civitai combines AI art generation, model exploration, community publishing, creator programs, and clear pricing into a platform that is easy to assess. |
| SP003 | Civitai | Civitai homepage | Civitai is an AI art platform built around discovering, creating, and sharing AI-generated media. |
| SP004 | Midjourney | Midjourney homepage | We're a lab of 60 people known for building the most beautiful AI models in the world. |
| SP005 | Skywork | SeaArt AI review 2025: hands-on testing | SeaArt AI is a controversial tool due to its ethical policies... NSFW filters, but there are various loopholes to get around them. |
| SP006 | Cybernews | SeaArt AI Review | SeaArt.AI is worth a serious look... with strong anime and photoreal styles, built-in face swap, LoRA training, and even video generation. |
| SP007 | Recatools | Wujie AI (无界AI) | Wujie AI's real hook isn't the image generator itself — it's the pairing with Wujie Bantu... register and auction AI-generated work as verified digital copyright. |
| SP008 | Sharewalker | Detailed review of four major AI art generation platforms: Liblib.Art, Tensor.Art, SeaArt.AI, Aitubo.AI | liblib.art, tensor.art, seaart.ai, and aitubo.ai each have their own advantages, meeting diverse needs from model sharing to image/video generation and virtual character creation. |
| SP009 | Baidu Baike | Tensor.Art | 平台提供文生图、图像风格转换、在线模型训练及ComfyUI工作流等功能模块。 |
| SP010 | Yicai Global | LiblibAI Parent Evoken Valued at Over USD2 Billion After New Funding Round | Evoken... valued at over USD2 billion after new funding round. |
| SP011 | AIbase | LiblibAI secures $300M B+ round and exceeds $300M ARR | As of May 2026, the company's Annual Recurring Revenue (ARR) has exceeded $300 million. |
| SP012 | KuCoin News | AI application-layer unicorn LiblibAI secures $300M in B+ round, valued at over $2B | LiblibAI secures $300M in B+ round, valued at over $2B. |
| SP013 | BestHub | How LiblibAI's $300M ARR fueled a near $300M funding round valuing it at over $2B | $300M ARR fueled a near $300M funding round valuing it at over $2B. |
| SP014 | Tencent News | Evoken / LiblibAI financing coverage | LiblibAI's parent adopted the Evoken corporate identity during the June 2026 financing cycle. |
| SP015 | 36Kr PitchHub | LiblibAI project profile | LiblibAI launched in September 2023 and completed the required Chinese generative-AI service filing in early 2024. |
| SP016 | Baidu Baike | LiblibAI | CCTV flagged 哩布哩布AI as able to bypass moderation under specific prompts to generate inappropriate content, followed by technical rectification. |
| SP017 | Cyberspace Administration of China | Generative AI service filing registry notice | CAC maintains public filing registries for generative-AI services. |
| SP018 | Liblib (company) | Liblib Art official site | Liblib's official surface was retrieved as a JavaScript-only site during this run. |
| SP019 | Tencent News | LiblibAI Series B coverage and operating metrics | 2024 full-year revenue was RMB 206 million, with approximately 25 million registered users and 100,000+ original models reported in late 2025 coverage. |
| SP020 | Hubpy | LiblibAI guide 2026 | LiblibAI aggregates open and closed models and supports LoRA training, model marketplace, and API-oriented workflows. |
| SP021 | Hurun Report | Hurun unicorn context | Hurun Report provides analyst-market-data context for private unicorn scale. |
| SP022 | IMARC Group | China Generative AI Market | China's generative AI market remains a multi-billion-dollar market with sustained growth expectations. |
| SP023 | Grand View Research | China Generative AI Market Outlook | Grand View Research provides a China generative-AI market outlook for sizing and growth context. |
| SP024 | 6Wresearch | Prominent companies in China AI image generator market | The industry is highly concentrated among top ecosystem leaders, with Baidu, Alibaba, and Tencent collectively holding a sizable market share. |
| SP025 | The Next Web | China micro-drama AI state funding | China's micro-drama sector is adopting AI with state support and industrial-scale production incentives. |
| SP026 | DigiTrendz | China's $16.5B micro-drama industry adopts AI video first | China's micro-drama industry is projected to exceed RMB 120 billion in 2026 and is adopting AI video-first workflows. |
| SI001 | China Biz Insider | Evoken hits $2B valuation as AI aggregator model faces its next test | Evoken develops no proprietary foundation models...built on third-party model APIs. |
| SI002 | ARR Club | ARR Club | Track Verified Revenue Data & Growth Strategies of Top AI Products. |
| SI003 | Sacra | Canva company profile | Revenue $4.00B; Valuation $42.00B; Growth Rate (y/y) 44%. |
| SI004 | Sacra | Canva at $4B ARR, growing 43% YoY | Sacra estimates Canva hit $4B in ARR at the end of 2025...valued at $65B...for a 16x revenue multiple. |
| SI005 | TechCrunch | Canva gets to $4B in revenue as LLM referral traffic rises | Canva had more than 265 million monthly active users and over 31 million paid users in 2025...annual recurring revenue to $4 billion. |
| SI006 | GetLatka | Midjourney revenue and company profile | Midjourney's most recent disclosed valuation is $10B. |
| SI007 | ElectroIQ | Midjourney Statistics | Revenues of Midjourney...went up to US$300 million and almost US$500 million in 2024 and 2025, respectively. |
| SI008 | Migrant Times | DBS, Granite Asia sign $110M AI fund deal to back Asian startups | Granite Asia has $10 billion in assets under management and co-managed capital. |
| SI009 | CB Insights | Granite Asia investor profile | Compare Granite Asia to Competitors. |
| SI010 | Yicai Global | LiblibAI Parent Evoken Valued at Over USD2 Billion After New Funding Round | As of May, Evoken's annual recurring revenue had reached USD300 million, more than 30 times the level of a year earlier. |
| SI011 | AIbase | LiblibAI secures $300M in B+ round at over $2B valuation | LibTV...daily revenue exceed one million US dollars in its first month and serves nearly 1,000 teams. |
| SI012 | KuCoin News | AI application-layer unicorn LiblibAI secures $300M in B+ round valued at over $2B | As of May 2026, Yanyu Technology disclosed that its ARR exceeded $300 million, with the group's total revenue growing by more than 3,000% year-over-year. |
| SI013 | BestHub | How LiblibAI's $300M ARR fueled a near $300M funding round valuing it at over $2B | LibTV's mixed subscription-plus-pay-per-use model earned $1 million in single-day revenue during its first month. |
| SI014 | Tencent News | Founder Chen Mian's ByteDance experience helps Evoken break out | The company's ARR surpassed $300 million and has approached some overseas leading vertical applications. |
| SI015 | 36Kr PitchHub | LiblibAI company profile | 2024 年 2 月,LiblibAI 通过了国家互联网信息办公室第四批深度合成服务算法备案;同年3月,成为国内首家通过国家《生成式人工智能服务管理暂行办法》备案的 AI 社区。 |
| SI016 | Baidu Baike | LiblibAI | Founder Chen Mian holds roughly 73.95% of the company and remains the largest shareholder. |
| SI017 | AIbase | LiblibAI completes $130M Series B financing | LiblibAI recently announced the successful completion of a $130 million Series B funding round. |
| SI018 | AIbase | LiblibAI completes $130M financing but is not yet profitable | Although the company has not yet turned a profit, it is clearly accelerating its development to cope with the competitive market environment. |
| SI019 | INCE Capital | INCE Capital's Portfolio Company LiblibAI Announces $130 Million Series B Financing | This round of financing was jointly led by Sequoia Capital China, CMC Capital, and a strategic investor. |
| SI020 | 36Kr | LiblibAI $130M Series B financing report | LiblibAI completed 1.3亿美元 B-round financing led by Sequoia China and CMC Capital. |
| SI021 | Sina Finance | LiblibAI Series B financing report | LiblibAI has completed a $130 million Series B round led by Sequoia China and CMC Capital. |
| SI022 | Cyberspace Administration of China | Notice publishing generative-AI service filing information | 截至11月1日,累计有611款生成式人工智能服务完成备案,306款生成式人工智能应用或功能完成登记。 |
| SI023 | Liblib (company) | LiblibAI-哩布哩布AI - 中国领先的AI创作平台 | LiblibAI-哩布哩布AI - 中国领先的AI创作平台 |
| SI024 | Tencent News | How LiblibAI avoids being swallowed by large-model technology | 2.0版本后,平台采用订阅制+按需付费的混合模式...企业客户为API服务和定制化需求付费。 |
| SI025 | TMTPost | LiblibAI completes three rounds of financing in one year | AI 图像生成平台LiblibAI...在一年内已完成三轮融资,总金额达数亿元人民币。 |
| SI026 | Sina Finance | LiblibAI completed three rounds totaling hundreds of millions of RMB | AI 图像生成平台「LiblibAI 哩布哩布 AI」日前宣布在一年内已完成三轮融资,总金额达数亿元人民币。 |
| SI027 | Hubpy | LiblibAI guide 2026 | LiblibAI 2.0 upgraded from aggregated tools into a professional AI creative studio. |
| SE001 | Apatero | ComfyUI LoRA Training Guide: Create Consistent Characters from Scratch | Training a character LoRA requires 15-50 high-quality images, captioning each image, and running training for 1000-3000 steps. |
| SE002 | Tech-Insider | ComfyUI tutorial: SDXL, FLUX workflow 13 steps 2026 | ComfyUI is a node-based DAG editor where every step of a diffusion pipeline is an explicit, swappable node. |
| SE003 | Runflow | ComfyUI API Developer Guide | The ComfyUI API is the HTTP and WebSocket interface exposed by the ComfyUI server that lets external applications submit generation workflows. |
| SE004 | GPTProto | What Exactly Is the LiblibAI API? | LiblibAI functions as a central hub for Stable Diffusion enthusiasts. It is not just a gallery; it is a full-throttle production environment. |
| SE005 | RunComfy | Lora-Training-in-Comfy | ComfyUI Nodes | Lora-Training-in-Comfy simplifies the creation of LoRA models within ComfyUI. |
| SE006 | Stable Diffusion Art | LoRA models and how to use them in Stable Diffusion | LoRA models are small Stable Diffusion models that apply tiny changes to standard checkpoint models. |
| SE007 | DataCamp | How to Run Stable Diffusion | Stable Diffusion is an open-source deep learning model designed to generate high-quality, detailed images from text descriptions. |
| SE008 | AIbase | LiblibAI 2.0 launches as a professional AI creation studio | LiblibAI 2.0 is no longer a simple collection of models and tools, but a true AI studio for creators. |
| SE009 | Yicai Global | LiblibAI Parent Evoken Valued at Over USD2 Billion After New Funding Round | Evoken’s LiblibAI platform has grown into one of China’s largest AI image-resource libraries and creator communities, with more than 30 million cumulative users. |
| SE010 | AIbase | 演语科技(Evoken)完成近3亿美元B+轮融资,ARR达3亿美元,AI应用层进入商业化阶段 | LiblibAI has over 30 million users and has accumulated 500,000 original models. |
| SE011 | KuCoin News | AI Application Layer Unicorn LiblibAI Secures $300M B+ Round Valued at Over $2.0B | LiblibAI has amassed over 30 million users and more than 500,000 original models, generating over 5 million images daily. |
| SE012 | BestHub | How LiblibAI’s $300M ARR fueled a near-$300M funding round valuing it at over $2B | Creators can upload, train, and share LoRA models and workflows on the platform, while regular users can directly utilize these resources. |
| SE013 | Tencent News | Founder Chen Mian’s ByteDance experience helps Evoken break out | |
| SE014 | Baidu Baike | LiblibAI | Baidu Baike records the CCTV content-safety incident and subsequent technical rectification. |
| SE015 | Cyberspace Administration of China | 国家互联网信息办公室关于发布生成式人工智能服务已备案信息的公告 | CAC publishes public generative-AI service filing information. |
| SE016 | Liblib (company) | LiblibAI-哩布哩布AI - 中国领先的AI创作平台 | LiblibAI-哩布哩布AI - 中国领先的AI创作平台 |
| SE017 | Tencent News | LiblibAI October 2025 financing and operating metrics | |
| SE018 | Hubpy | LiblibAI Guide 2026 | LiblibAI is China’s largest AI image generation platform and model sharing community. |
| SE019 | 6Wresearch | Prominent Companies in China AI Image Generator Market | |
| SE020 | Sharewalker | Detailed review of four major AI art generation platforms: Liblib.Art, Tensor.Art, SeaArt.AI, AituBo.AI | |
| SE021 | China Biz Insider | Evoken hits $2B valuation as AI aggregator model faces its next test | Unlike Aishi, Evoken develops no proprietary foundation models. Its entire product suite is built on third-party model APIs. |
| SE022 | IMARC Group | China Generative AI (AIGC) Market Size, Share, Trends and Forecast | |
| SE023 | Grand View Research | China Generative AI Market Horizon Databook | |
| SE024 | The Next Web | China’s micro-drama industry is becoming the first mass market for AI video | |
| SE025 | Skywork | SeaArt AI review 2025: hands-on testing | |
| SU001 | Jiemian News | AI image generation platform LiblibAI announces four rounds of financing within one year | 为金山办公、万兴科技、吉比特游戏、天猫校园、清华大学等B端客户提供了专业的AI图像场景解决方案。 |
| SU002 | Outlook Respawn | Chinese gaming sector backs generative AI push | Chinese video game companies are emerging as financial backers of the country's generative artificial intelligence sector. |
| SU003 | China Biz Insider | miHoYo commits RMB 100 billion to AI, repatriates Silicon Valley LLM team | The maker of Genshin Impact is staking up to RMB 100 billion (US$13.9 billion) on artificial intelligence over three years. |
| SU004 | Esports.gg | miHoYo reportedly investing $14 billion into AI over next three years | HoYoverse is reportedly planning to invest up to 100 billion yuan (around $14 billion) over the next three years into large-scale AI development. |
| SU005 | AsiaICT | The production capacity barrier of the gaming industry is being dismantled by AI technology | Tencent, NetEase, miHoYo... All major game companies are placing their bets on AI. |
| SU006 | China Biz Insider | China's overseas short-drama market hits $229M in May as AI-generated content surges | The overseas micro-drama market generated an estimated $229 million in in-app purchase revenue across both iOS and Android platforms in May 2026. |
| SU007 | Tech Times | China's AI short drama boom hit industrial scale, faces stolen jobs concern | The displacement of production workers is not a forecast — it is a current condition. |
| SU008 | WebProNews | China’s Gaming Giants Fuel AI Revolution in Game Creation | Chinese game developers, facing soaring production costs and flat consumer spending, are pouring resources into generative artificial intelligence. |
| SU009 | Yicai Global | LiblibAI Parent Evoken Valued at Over USD2 Billion After New Funding Round | Evoken's LiblibAI platform has grown into one of China's largest AI image-resource libraries and creator communities, with more than 30 million cumulative users. |
| SU010 | AIbase | AI application-layer unicorn LiblibAI secures $300M B+ round | LiblibAI has over 30 million users and has accumulated 500,000 original models; LibTV saw daily revenue exceed one million US dollars in its first month. |
| SU011 | KuCoin News | AI application-layer unicorn LiblibAI secures $300M in B+ round, valued at over $2.0B | LiblibAI has accumulated over 30 million users, more than 500,000 original models, and over 100 million professional images and video assets. |
| SU012 | BestHub | How LiblibAI's $300M ARR fueled a near $300M funding round valuing it at over $2B | Within a month of launch, over 300 short-film and film companies integrated LibTV, and the platform now serves nearly a thousand agencies, production houses and brand clients. |
| SU013 | Tencent News | AI application company Evoken completes new financing at unicorn valuation | 明星产品 Liblib AI 累计用户超过 3000 万,目前全公司的 ARR(年化收入)突破 3 亿美元。 |
| SU014 | 36Kr PitchHub | LiblibAI project profile | LiblibAI became China's first AI-community platform to complete generative-AI service filing under the interim measures. |
| SU015 | Baidu Baike | LiblibAI | CCTV named 哩布哩布AI in coverage of AI apps that could bypass moderation under specific prompts; the company undertook technical rectification. |
| SU016 | Tencent News | LiblibAI is typical of China's AI-application startup wave | 目前Liblib的月活跃用户为400万,总用户数2500万。 |
| SU017 | Hubpy | LiblibAI Guide 2026 | LiblibAI offers a free tier with limited daily generations. Higher-volume usage and priority GPU access require a subscription or credits. |
| SU018 | AIbase | LiblibAI 2.0 upgraded to AI professional creation studio | LiblibAI has quickly become the core platform for the creator community, now gathering over 20 million creators. |
| SU019 | GPTProto | LiblibAI API: What exactly is the LiblibAI API? | API Availability: Both offer APIs, but LiblibAI's is much harder to set up for international developers. |
| SU020 | IMARC Group | China Generative AI Market Size, Share, Trends and Forecast | China generative AI market report covers size, share, trends and forecasts for the market. |
| SU021 | Grand View Research | China Generative AI Market Size & Outlook | China generative AI market size and outlook by Grand View Research. |
| SU022 | 6Wresearch | Prominent companies in China AI image generator market | Prominent companies in China AI image generator market include major Chinese technology platforms. |
| SU023 | DigiTrendz | China's $16.5B micro-drama industry adopts AI video first | China's micro-drama market is projected to exceed RMB 120 billion in 2026 as AI video is adopted first. |
| SU024 | The Next Web | China micro-drama AI state funding | China's booming micro-drama sector is adopting AI and attracting policy support. |
| SU025 | Startup Fortune | China's AI micro-drama boom is a business model worth watching | China's AI micro-drama boom is creating a new business model for fast, low-cost entertainment production. |
| SR001 | China Law Translate | Measures for Labeling of AI-Generated Synthetic Content | These measures are to take effect from September 1, 2025. |
| SR002 | Covington InsidePrivacy | China Releases New Labeling Requirements for AI-Generated Content | The rules will take effect on September 1, 2025. |
| SR003 | Loeb & Loeb | China's AI Labeling Measures and Mandatory National Standards Take Effect September 1 | Companies have until Sept. 1 to study the Measures and the National Standards and build their AI-labeling tools. |
| SR004 | Wolters Kluwer | Beijing Internet Court grants copyright to AI-generated image for the first time | Plaintiff selected over 150 prompts, arranged their order and set specific parameters. |
| SR005 | Future of Privacy Forum | China's Interim Measures for the Management of Generative AI Services: A Comparison Between the Final and Draft Versions of the Text | On August 15, 2023, the Interim Measures for the Management of Generative AI Services came into force. |
| SR006 | Deep Lex | China AI Regulation Tracker | China operates the most extensive binding sectoral AI regulatory regime globally, with no single comprehensive AI law to date. |
| SR007 | Linklaters | China: dual-track AIGC labelling and latest AI regulatory development | They establish a dual-track labelling system to set obligations for responsible parties across the entire AIGC value chain. |
| SR008 | China Legal Experts | China Deep Synthesis Regulation 2025: Essential Guide | Effective January 2023, it regulates AI-generated content (text, images, voice, video). |
| SR009 | Baker McKenzie | China: A landmark court ruling on copyright protection for AI-generated works | Whether or not an AI-generated work is copyrightable will still need to be determined on a case-by-case basis. |
| SR010 | China IP Law Update | Beijing Internet Court Requires Evidence of Creative Effort to Claim Copyright Protection in AI-Generated Images | Creators must explain their creative thinking, input commands, selection and modification process, and evidence of creative labor investment. |
| SR011 | IAM Media | Using copyrighted content to train generative AI can be deemed fair following Ultraman infringement dispute | The court affirmed that the defendant, as a provider of generative-AI services, was not involved in direct copyright infringement. |
| SR012 | China IP Case Updates | The First Case Involving Copyright Infringement Conducted by a Generative AI Platform: Ultraman v Acgnai | Generative AI service providers are not obligated to conduct prior reviews of user-input data...Liability for contributory infringement only arises when the provider is aware of or actively contributes to specific infringing acts. |
| SR013 | National Law Review | Hangzhou Internet Court: Generative AI Output Infringes Copyright | The court ordered the defendant to immediately stop the infringement and compensate for economic losses and reasonable expenses of 30,000 RMB. |
| SR014 | HLC (Hylands) | When AI and Copyright Clash: Chinese Courts Find AI Platform Liable for Contributory Copyright | The AI platform could not benefit from the safe harbor exemption. |
| SR015 | MMLC Group | China AI Copyright | Generative AI service providers should take certain technical measures to avoid generating images that are substantially similar to other's works. |
| SR016 | EU IP Helpdesk | Ultraman AI Case in China: Defining Copyright Liability for Generative AI Providers | The more profitable a platform is, the higher its duty of care. |
| SR017 | Yicai Global | LiblibAI Parent Evoken Valued at Over USD2 Billion After New Funding Round | Evoken's annual recurring revenue had reached USD300 million, more than 30 times the level of a year earlier. |
| SR018 | AIbase | LiblibAI Secures Nearly $300 Million in Series B+ Financing, Valuation Exceeds $2 Billion | LiblibAI has secured nearly $300 million in B+ round financing, valuing it at over $2 billion. |
| SR019 | KuCoin News | AI Application Layer Unicorn LiblibAI Secures $300M in B+ Round, Valued at Over $2B | LiblibAI secured $300 million in a Series B+ round and is valued at over $2 billion. |
| SR020 | Baidu Baike | LiblibAI | CCTV exposed that LiblibAI could bypass review mechanisms under specific prompts to generate inappropriate content. |
| SR021 | Tencent News | 问AI · 创始人陈冕的字节经历如何助力公司破圈? | 陈冕曾经是字节跳动剪映与 CapCut 全球商业化负责人。 |
| SR022 | 36Kr PitchHub | LiblibAI project profile | 2024 年 2 月,LiblibAI 通过了国家互联网信息办公室第四批深度合成服务算法备案;同年3月,成为国内首家通过国家《生成式人工智能服务管理暂行办法》备案的 AI 社区。 |
| SR023 | Cyberspace Administration of China | 国家互联网信息办公室关于发布生成式人工智能服务已备案信息的公告 | 截至11月1日,累计有611款生成式人工智能服务完成备案,306款生成式人工智能应用或功能完成登记。 |
| SR024 | AIbase | LiblibAI Completes $130 Million Series B Financing | Although the company has not yet turned a profit...capital has become a key barrier for entrepreneurs. |
| SR025 | AIbase | LiblibAI Receives $130 Million Series B Financing | The financing is the largest in the Chinese AI application sector this year. |
| SR026 | Tencent News | LiblibAI announces new financing and 2024 revenue metrics | 2024 full-year revenue was reported as RMB 206 million. |
| SR027 | China Biz Insider | Evoken hits $2B valuation as AI aggregator model faces its next test | Evoken develops no proprietary foundation models...built on third-party model APIs. |
| SR028 | Jiemian News | LiblibAI customer and B-side use cases | Named B-side customers include Kingsoft Office, G-bits Games, Tmall Campus, Wondershare, and Tsinghua University. |
| SR029 | Tech Times | China's AI Short Drama Boom Hit Industrial Scale: 470 Titles A Day, Faces 'Stolen Jobs' Gone | Generative video has been deployed as a mass commercial production system...eliminating entire categories of creative jobs. |
| SR030 | China Biz Insider | MiHoYo Commits RMB 100 Billion to AI, Repatriates Silicon Valley LLM Team | Chinese gaming companies are increasing AI investment for content production. |
| SR031 | AIbase | LiblibAI 2.0 Upgrades to Professional AI Creation Studio | LiblibAI 2.0 upgraded from a model community into an all-in-one AI creation studio. |
| SV001 | SaaSRise | The AI Software Valuation Report 2026 | AI-native companies command a median 21.2x EV/Revenue in VC rounds and 11.5x in M&A buyouts. |
| SV002 | SaaSRise | SaaS Benchmark Report 2026 | AI-native companies grow at roughly 4x the aggregate rate of legacy peers and face 55-70% gross-margin challenges. |
| SV003 | ScaleXP | SaaS ARR Revenue Valuation Multiples | Current public benchmarks show the BVP Nasdaq Emerging Cloud Index around a 6.2x average revenue multiple. |
| SV004 | Acquiry | SaaS Valuation Multiples 2026 | AI-native SaaS above 50% ARR growth trades at 10x to 20x ARR, while non-AI SaaS trades at 4x to 7x. |
| SV005 | We Are Founders | 2025 US SaaS Valuation Multiples: The Founders Benchmark | AI wrappers are seeing multiples collapse; vertical AI with proprietary workflows trades around 9x-12x ARR. |
| SV006 | PitchBook | Granite Asia Overview | PitchBook lists Granite Asia as an active Singapore asset manager with hundreds of investments and LiblibAI as a June 2026 investment. |
| SV007 | Tracxn | Granite Asia private equity profile | Target URL returned 429 Too Many Requests, so it is retained only as a rate-limited corroboration point. |
| SV008 | China Biz Insider | 15 Embodied AI Unicorns in 6 Months: China’s Robot Race Hits a Reality Check | Most startups in the cohort carry cash runways of only 18 to 24 months, making a 2027-2028 reckoning likely. |
| SV009 | Tech Funding News | 16 unicorns in one month: what June 2026 says about where AI money is going | Liblib raised $300 million in a Series B+ round at a valuation of over $2 billion led by Granite Asia, Tencent and Shunwei. |
| SV010 | TechCrunch | Almost 40 new unicorns have been minted so far this year; here they are | TechCrunch tracked VC-backed startups becoming unicorns in 2026 using Crunchbase and PitchBook data. |
| SV011 | Nation Press | China mints 67 unicorns in H1 2026 | China created 67 new unicorn start-ups in the first half of 2026, with AI and robotics accounting for more than 53 percent. |
| SV012 | Yicai Global | LiblibAI Parent Evoken Valued at Over USD2 Billion After New Funding Round | Evoken completed a USD300 million Series B+ funding round, valuing the company at more than USD2 billion. |
| SV013 | AIbase | LiblibAI parent Yanyu Technology completes nearly $300M B+ financing | LiblibAI has over 30 million users, 500,000 original models and ARR projected to exceed $300 million. |
| SV014 | KuCoin News | AI application-layer unicorn LiblibAI secures $300M B+ round | LiblibAI closed nearly $300 million and reached a valuation exceeding $2 billion after a prior $130 million round. |
| SV015 | BestHub | How LiblibAI’s $300M ARR fueled a near $300M funding round | ARR reached $300 million and group revenue grew over 3000% year over year by May 2026. |
| SV016 | Tencent News | 问AI · 创始人陈冕的字节经历如何助力公司破圈? | On June 18 Evoken announced nearly US$300 million B+ financing at a valuation above US$2 billion. |
| SV017 | Baidu Baike | LiblibAI | Baidu Baike records the CCTV content-safety exposure and company rectification context. |
| SV018 | AIbase | LiblibAI announced $130M Series B financing | AIbase reported the Series B and noted the company was not yet profitable. |
| SV019 | INCE Capital | INCE Capital portfolio news on LiblibAI financing | INCE Capital described the October 2025 Series B and investor participation. |
| SV020 | Cyberspace Administration of China | 国家网信办关于发布生成式人工智能服务已备案信息的公告 | As of November 1, 611 generative AI services had completed filing and applications using filed models must disclose filing or registration status. |
| SV021 | Hurun Report | Global Unicorn Index 2026 | Hurun found a record 1,603 unicorns globally, with AI driving 36% of unicorn value and China adding 80 in the year. |
| SV022 | China Biz Insider | Evoken hits $2B valuation as AI aggregator model faces its next test | At a US$2 billion valuation on US$300 million ARR, Evoken is assigned a roughly 6.7x revenue multiple. |
| SV023 | ARR Club | ARR Club | ARR Club tracks verified revenue data and growth strategies of top AI products. |
| SV024 | Sacra | Canva company profile | Canva profile lists revenue of $4.00B, valuation of $42.00B and 44% growth. |
| SV025 | Sacra | Canva at $4B ARR, growing 43% YoY | Sacra estimates Canva hit $4B ARR and Gamma had $102M ARR at a $2.1B valuation for a 20.6x multiple. |
| SV026 | TechCrunch | Canva gets to $4B in revenue as LLM referral traffic rises | Canva had more than 265M monthly active users, more than 31M paid users and $4B ARR by end-2025. |
| SV027 | GetLatka | Midjourney revenue and valuation profile | Midjourney hit $500M revenue in May 2025 and its most recent disclosed valuation is $10B. |
| SV028 | ElectroIQ | Midjourney Statistics | Midjourney revenue reached almost US$500 million in 2025 and had more than 21 million Discord members. |
| SV029 | CB Insights | Granite Asia investor profile | CB Insights lists Granite Asia competitors and investor context. |
| SV030 | Migrant Times | DBS and Granite Asia sign $110M AI fund deal | DBS and Granite Asia launched an AI fund to back Asian startups. |
| SV031 | Jiemian News | AI creative platform customer proof and industry adoption | Jiemian cites benchmark B customers and professional creator demand for Liblib. |
| SV032 | China Law Translate | AI Labeling Measures | China’s AIGC labeling rules require explicit and implicit labeling obligations for generated content. |