Evoken

1.1 Identity and product system

Evoken is described as a Beijing-based company behind LiblibAI, LibTV, and a design-agent product. The available record supports a product-system description rather than a fully disclosed legal-entity profile: LiblibAI is identified as an AI image-generation and model-sharing platform founded in March 2023, while the newer products extend into video creation and design workflow. The product architecture is strategically coherent—community and model assets can feed production and delivery workflows—but the evidence does not establish consolidated legal entities, headquarters beyond Beijing references, or a complete board and governance structure. A company encyclopedia describes LiblibAI as providing online generation, copyright protection, and community functions, while independent reporting places Evoken in Beijing and Shanghai and links it to Lovart. These are useful identity anchors, not substitutes for corporate filings.[CO001, CO002, CO003, CO004]

1.2 Leadership, investors, and control questions

The assigned evidence identifies Chen Mian as Evoken's founder and describes a prior ByteDance commercialization background, but it does not provide a complete executive roster, board composition, voting-control structure, or key-person succession plan. Financing coverage names a broad investor set across the 2025 Series B and 2026 Series B+ rounds, including Granite Asia, Tencent, Shunwei Capital, HSG, Ant Group, and other existing backers. That breadth may provide financing access and strategic distribution, yet the public material does not quantify ownership, liquidation preferences, secondary transactions, debt, or investor concentration. The leadership and stakeholder tables therefore separate reported roles from diligence asks rather than implying governance facts that are not disclosed.[CO005, CO006, CO007, CO008, CO009]

1.3 Scale, milestones, and underwriting gaps

Evoken's reported scale has accelerated sharply: coverage of the June 2026 round cites more than $500 million of combined financing across two major rounds, a post-money valuation above $2 billion, and ARR above $300 million as of May 2026. Tech in Asia independently reports the financing and ARR figures while also describing an unfinalized, early-stage Hong Kong IPO discussion and a 300,000-person beta user base for Lovart after an 800,000-person waitlist. These metrics are not audited financial statements, and the source record does not disclose gross margin, net retention, cash balance, burn, runway, customer concentration, or current headcount. The milestone timeline captures reported events and adverse diligence signals, including the April 2026 content-safety issue and the later risk warning reported by Baidu Baike; neither absence nor presence of these reports proves the current state of remediation.[CO010, CO011, CO012, CO013, CO014]

1.4 Exhibits

2.1 Market boundary and status quo

Evoken's addressable market is narrower than all generative AI spending. The relevant core is software and services that help consumers and professional teams create, edit, distribute, or manage AI-generated visual content, including image, video, design-agent, model-community, and workflow products. Adjacent spend includes general-purpose model APIs, cloud compute, creator marketplaces, advertising production, and entertainment production; those adjacencies can expand the opportunity but should not be counted as Evoken revenue without evidence. The status quo remains a mix of manual production, agencies, stock assets, conventional editing software, and general-purpose model tools. Public descriptions of LiblibAI show a progression from image-model creation and sharing toward video and full workflow production, while market research separately identifies software, services, media and entertainment, and content generation as distinct categories. This boundary supports a platform thesis but prevents the broadest China-AIGC estimates from being treated as directly captured market.[CM001, CM002, CM003, CM004]

2.2 Multiple sizing lenses and contradictions

Three external lenses illustrate both scale and uncertainty. IMARC estimates China's generative-AI market at USD 5,160.82 million in 2025 and USD 19,558.29 million in 2034, a 15.96% CAGR, while Grand View Research's databook reports USD 1,751.6 million in 2025 and USD 27,861.6 million in 2033, a 42.3% CAGR. The difference is not a rounding issue: the publishers use different definitions, segment coverage, and forecast horizons. A separate Xinhua report gives a more focused demand signal, estimating AI-generated comic-style micro dramas at RMB 16.8 billion in 2025 and more than 10,000 monthly releases from early 2026. Evoken-specific databases report a $300 million-plus round and $2 billion valuation, but those are financing facts, not market size. A disciplined TAM should therefore present these as separate lenses, with SAM and SOM left unquantified until Evoken discloses geography, product revenue mix, retention, and paid conversion.[CM005, CM006, CM007, CM008, CM009]

2.3 Buyers, users, and adoption path

The buyer map is segmented by who creates content and who controls the budget. Consumer creators and model communities are users and may pay subscriptions or usage fees; professional producers, studios, advertisers, and design teams are both users and budget owners; enterprise customers can buy workflow, team, or API capacity; and cloud or model partners can influence distribution and variable cost without being end customers. The adoption path generally starts with experimentation or a single asset, expands to repeatable production, and only then becomes a team workflow or enterprise integration. The Xinhua evidence suggests AI-generated micro dramas can lower production cost and enable multilingual distribution, while the Baidu encyclopedia record describes B-side image solutions and a LibTV team edition. These signals support a workflow expansion thesis, but the pool does not provide conversion, contract value, churn, or customer concentration data.[CM010, CM011, CM012, CM013]

2.4 Drivers, constraints, and valuation relevance

Demand is being pulled by better multimodal generation, falling production costs, expanding enterprise experimentation, and the ability to produce content at greater speed and scale. IMARC describes adoption across finance, retail, manufacturing, and media, while Xinhua documents a concrete production-cost comparison in which an AI-generated micro drama is said to cost just over RMB 100,000 versus several hundred thousand yuan for live action. Constraints are equally material: regulatory compliance requires data governance, content filtering, and continuing model updates; computing costs and restricted access to advanced hardware can compress margins; and reliability, intellectual-property, and inappropriate-content concerns can slow mission-critical adoption. These constraints connect directly to valuation: growth is more durable if Evoken can convert creator usage into recurring team or enterprise budgets while controlling model and compliance costs. Public evidence does not yet establish that economics.[CM014, CM015, CM016]

2.5 Exhibits

The exhibits preserve the market boundary, competing external sizing lenses, and the buyer workflow without forcing unlike estimates into one total. Values are shown in the publisher's stated currency, unit, geography, and period. Where a source is company-profile or registry material, it is treated as a financing or operating signal rather than independent market validation. The central diligence question is not whether generative AI is a large category, but which repeatable paid workflow Evoken owns, how much of the relevant spend it can capture, and whether compliance and model costs permit attractive contribution margins.[CM005, CM006, CM007, CM010, CM014, CM016]

2.6 Exhibits

3.1 Landscape and buyer alternatives

Evoken's competitive set is broader than a single image-generator peer. RoundOS describes the portfolio as LiblibAI for image generation and model sharing, Xingliu as a design agent, and LibTV as an AI video-production platform serving creators and businesses. The practical alternatives therefore include creator communities such as SeaArt and Tensor.Art, specialist video tools, incumbent creative suites, agency or internal production, and status-quo manual workflows. A 36Kr project profile describes LiblibAI as an AI-native application for content creation and sharing, with a community and model ecosystem; its figures are company-profile claims rather than independent benchmarks. Cybernews' hands-on SeaArt review shows the adjacent category already combines image, video, face-swap, and lightweight model training, so feature breadth alone is unlikely to be durable differentiation. The buyer decision is consequently about workflow completion, distribution, trust, and repeatable production rather than whether a platform can generate one attractive image.[CP001, CP002, CP003, CP004]

3.2 Capability, packaging, and switching tradeoffs

The available evidence indicates a capability race with materially different packaging. The Tensor.Art description emphasizes free online image generation, hosted model sharing, complex node workflows, and LoRA training. Cybernews describes SeaArt's hybrid Stamina/Credits model: daily Stamina is consumed first, while Credits do not expire, and the review reports roughly six Credits per image and 30–90 Credits per video, subject to model and compute complexity. Those are competitor-specific observations, not a directly comparable market price. The same review says SeaArt is browser-first, has inconsistent outputs under repeated prompts, and exposes a broad shared library with ethical and licensing concerns. These tradeoffs create switching friction through accumulated models, workflows, prompts, libraries, and creator reputation, but the pool has no comparable Evoken pricing, retention, or migration data. Cells without evidence are deliberately marked unknown.[CP005, CP006, CP007, CP008]

3.3 Moat durability and displacement risk

Evoken's strongest defensible possibility is a connected production chain: an image/model community can supply creative assets, while design-agent and video products can capture more of the downstream workflow. The Tencent News fetch assigned to this pool is only an app download page and does not substantiate the article headline, so it is excluded from claims. Independent evidence is also mixed: Cybernews reports both broad utility and ethical, licensing, browser-dependence, and consistency drawbacks for a close adjacent platform. Hurun's 2026 index establishes that AI is a large and fast-moving unicorn category, not that Evoken has a unique moat. Competitor convergence, model commoditization, and incumbent distribution remain credible displacement risks. The key diligence test is whether Evoken's cross-product users, proprietary community assets, and professional workflows produce lower acquisition cost or higher repeat usage than single-product alternatives; public evidence in this pool does not answer that question.[CP009, CP010, CP011, CP012]

3.4 Exhibits

The exhibits separate company-profile evidence from independent competitor observations and avoid presenting unlike token, user, funding, or valuation measures as a common performance scale. The positioning map is an evidence-backed ordinal synthesis, not a benchmark: its axes reflect breadth of the documented workflow and degree of community or ecosystem integration, while exact quality, price, and retention remain unknown. It should be read as a diligence prioritization aid, not as a claim that any plotted provider wins on customer outcomes.[CP003, CP006, CP009, CP012]

3.5 Exhibits

4.1 Monetization and public traction

The available evidence describes three monetization surfaces rather than a single homogeneous SaaS product. LiblibAI is presented as a subscription community with compute, LoRA training, original models, and advanced generation as paid benefits, while B2B deployments span commerce, design, games, education, and training. LibTV adds video creation and team plans, and Lovart/Xingliu adds an agent-style design product. TMTPost reports a subscription model and enterprise use cases, while Baidu Baike records more than 30 million cumulative users, but this is not an audited revenue measure. A later industry article attributes more than $300 million of group ARR to Evoken and $80 million ARR to Lovart; that attribution should be treated as company-linked or media-reported until contracts, billing cohorts, and revenue recognition are disclosed. Peer figures are useful only as scale references: GetLatka reports $200 million of Midjourney revenue, while Sacra and TechCrunch report or estimate $4 billion of Canva ARR. Revenue and ARR are not interchangeable, and neither benchmark establishes Evoken's margins.[CI001, CI002, CI003, CI004, CI008, CI009, CI015, CI016]

4.2 Pricing, cost structure, and unit economics

Public pricing evidence is incomplete, but the June 2026 account describes LibTV annual tiers from RMB 569 to RMB 8,499 and off-peak generation as low as RMB 20 per minute. Those are list or observed consumer prices, not realized blended ARPU, gross profit, or cash contribution. The same account says Evoken relies on third-party model APIs and therefore earns a spread between provider costs and retail pricing. This creates a clear sensitivity: usage intensity, provider volume discounts, queue capacity, refunds, and support costs can move gross margin even when subscription revenue grows. No assigned source discloses CAC, sales cycle, payback, churn, net retention, gross margin, working capital, or capex. The appropriate underwriting stance is to model these fields as unavailable and request cohort-level evidence rather than infer SaaS economics from ARR or user totals.[CI005, CI006, CI007, CI013]

4.3 Capital adequacy and underwriting verdict

The strongest financing evidence is a reported Series B+ of nearly $300 million at a valuation above $2 billion, alongside an article describing more than $500 million of cumulative funding. The financing improves near-term liquidity, but it is not equivalent to cash on hand: the pool does not provide a balance sheet, cash reconciliation, debt schedule, burn rate, runway, or use-of-funds disclosure. The same reporting highlights a structural risk in subsidized aggregation: providers can change API prices or launch competing first-party products, while moderation and content-compliance failures can impose remediation or regulatory costs. The financial verdict is therefore track rather than buy: growth and financing are credible enough to warrant diligence, but margin durability and capital sufficiency remain material blockers.[CI010, CI011, CI012, CI013, CI014]

4.4 Exhibits

5.1 Product surface and customer workflow

The clearest product description positions Evoken above individual generation models. LiblibAI is described as an image-creation community, LibTV as an AI-video platform, and Xingliu as an AI design agent, while the three are presented as a chain from source material through production to delivery. A separate LiblibAI 2.0 description says creators can move from an idea to a finished work in one studio and generate both images and videos. This evidence supports a workflow-layer thesis: creators discover assets, select a model or effect, generate media, and carry the output toward production. It does not establish which users pay, whether reported teams are in production, or whether cross-product movement is technically integrated rather than a portfolio-level marketing claim.[CE001, CE002, CE007, CE008]

5.2 Architecture and operating dependencies

The available material describes compatibility with open-source and closed-source models but does not disclose Evoken's routing, inference, storage, or observability stack. General practitioner references clarify the dependencies behind one visible community workflow. LoRA adapters modify a base checkpoint rather than operating alone, and Stable Diffusion generation iterates from noise through denoising. Local Stable Diffusion deployment also requires GPU capacity. These mechanics make model compatibility, checkpoint governance, compute availability, and inference cost plausible operating dependencies, but they are not proof of Evoken's implementation. Investors should therefore separate the observed product catalog from an unverified technical architecture and request direct evidence on model vendors, proprietary components, latency, unit cost, failover, storage, and release controls.[CE003, CE009, CE010, CE011, CE012]

5.3 Maturity and differentiation signals

Breadth and ecosystem density are the strongest public maturity signals. Company-issued coverage reports more than 20 million creators, multiple video models, more than 500 visual effects, and a style-model library spanning illustration, photography, and e-commerce. Separate company-issued coverage describes LiblibAI, LibTV, and Xingliu as a connected product group. These claims suggest differentiation through creator supply, reusable model assets, and workflow breadth rather than a publicly documented proprietary foundation model. The evidence does not reconcile registrations with active creators, catalog entries with safe or useful assets, or portfolio adjacency with actual interoperability. There are no retained customer benchmarks, uptime records, API documents, patents, or model-quality evaluations that independently demonstrate defensibility or production-grade performance.[CE004, CE005, CE006, CE007, CE008]

5.4 Trust, reliability, and technical risk

Independent market analysis identifies data governance, content filtering, watermarking, compute cost, and access to advanced hardware as material constraints for Chinese generative-AI operators. Those are sector risks, not evidence that Evoken has suffered a specific incident. The reviewed company-specific sources do not document security certifications, penetration testing, privacy controls, content provenance, copyright screening, model-upload review, incident history, service levels, disaster recovery, or regulatory approvals. No public developer repository, package, API forum, or engineering benchmark was available in the usable evidence. The product may therefore be commercially broad while still carrying unquantified third-party-model, compute, rights, moderation, and reliability exposure. Technical diligence should treat controls and service quality as open evidence requests rather than infer them from user or catalog scale.[CE013, CE014, CE015]

5.5 Exhibits

6.1 Reported reach and customer segments

A May 2026 report describes Evoken as operating three connected products: LiblibAI, LibTV, and Xingliu. The article reports more than 30 million LiblibAI users, 500,000 original models, more than 10 million Xingliu users, and nearly 1,000 teams using LibTV. It also says the core revenue mainly comes from professional content producers in short dramas, films, and advertising. These are third-party-reported company metrics rather than audited customer records, and the article does not define active users, paying users, customer contracts, geography, cohort dates, or revenue contribution by product. The defensible segmentation is therefore consumer/community reach, professional production teams, and design-agent users; payer and user identities remain only partly observable.[CU001, CU002, CU003, CU004, CU005, CU006]

6.2 Adoption trajectory and named proof

The available source set does not provide a customer-by-customer reference list, procurement record, contract term, renewal record, or independently measured utilization series for Evoken. The reported LibTV figure of nearly 1,000 teams is a useful indicator of commercial surface area, while the reported 30 million LiblibAI users and more than 10 million Xingliu users indicate broad reach; neither figure proves paid conversion, repeat use, production deployment, or outcome quality. Two prefetched pages that contain the name Evoken are clearly about a New Jersey funeral-doom metal band, not the Beijing AI company, and must not be treated as customer evidence. This name collision materially increases the need for direct customer references and a product-level activity denominator.[CU003, CU007, CU008, CU009, CU010, CU011, CU012]

6.3 Retention, expansion, and concentration

No assigned source discloses Evoken's NRR, GRR, logo churn, renewal rate, contract length, cohort retention, net dollar expansion, customer concentration, or channel mix. The article's statement that revenue mainly comes from professional content producers identifies a plausible buyer segment, but it does not establish whether those producers are direct subscribers, API users, agencies, or one-time project customers. The wider market evidence shows that AI video is being adopted in Chinese content production, including claims of industrial-scale output and sharply shorter production cycles, but those market conditions are not evidence of Evoken retention or share. Underwriting should treat expansion and concentration as open questions until account-level and cohort-level evidence is supplied.[CU005, CU013, CU014, CU015, CU016]

6.4 Evidence limitations and diligence priorities

The customer case is directionally positive but evidence-constrained. The strongest current signal is a reported combination of consumer-scale reach and professional-team adoption across linked products. The weakest links are independent corroboration, payment status, production-versus-pilot classification, and durability. The assigned search log recorded nine searches, but only two returned the name-collision pages retained as result URLs; the other relevant-looking pages were rediscovered prior URLs or general market coverage rather than current search hits. A diligence process should request an anonymized customer waterfall, active-paying definitions, monthly cohort retention, top-ten account share, contract and renewal samples, and named customer permission for reference calls.[CU001, CU003, CU007, CU012, CU013, CU014, CU015, CU016]

6.5 Exhibits

7.1 Regulatory and legal exposure

China's synthetic-content rules are an operating requirement, not merely a registration exercise. The Measures took effect on September 1, 2025 and require both perceptible labels and metadata labels, including when users download, reproduce, or export generated files. They also connect labeling evidence to algorithm-filing and safety-assessment procedures and authorize relevant regulators to address violations. This makes end-to-end label persistence, export testing, log retention, and clear compliance ownership central diligence items for Evoken's image and video products. Historical reporting says LiblibAI completed Chinese algorithm and generative-AI service filings in 2024, but that company-linked account does not verify current implementation under the later labeling regime. The thesis-break trigger is a substantiated filing defect, enforcement action, or inability to demonstrate labels and traceability across exported content.[CR001, CR002, CR003, CR004, CR005, CR010]

7.2 Copyright and content-control exposure

Copyright risk is unusually concrete for a creator platform that supports user-uploaded models and workflows. A Hangzhou appellate judgment upheld contributory infringement where users trained a LoRA model on protected Ultraman images and the platform failed to take necessary measures. The accompanying legal analysis says duty of care can rise with profitability and recommends stronger copyright-recognition systems for dedicated model training. A separate Beijing ruling recognized copyright in a prompted AI image, but the commentary stresses that Chinese courts lack common-law stare decisis and have not reached consensus. Evoken therefore faces both infringement-control risk and ownership uncertainty. Mitigation must be evidenced through rights provenance, repeat-infringer handling, notice and takedown metrics, model screening, and escalation records rather than a general claim that users bear responsibility.[CR006, CR007, CR008, CR009]

7.3 Operating, partner, and model dependencies

The assigned evidence identifies several dependencies but does not quantify them. Regulatory analysis places responsibility on providers for training data quality, illegal-content response, user agreements, and information security. A republished industry analysis says the company did not train a general-purpose foundation model and instead competes at the application layer, where external models can be purchased, integrated, and replaced. That may provide flexibility, but it also leaves model-provider pricing, availability, quality, licensing, cloud capacity, and migration readiness as material unknowns. The public pool contains no Evoken-specific outage, breach, certification, complaint, or enforcement history. Absence from this bounded evidence set is not proof of absence; management should supply an incident register, dependency map, concentration shares, termination terms, tested fallbacks, and recovery objectives.[CR009, CR012, CR016]

7.4 Financial-model risk, monitoring, and kill criteria

Capital availability is not the same as loss-absorption capacity. A 2024 financing account said proceeds would fund compute infrastructure, model and plugin development, and creator-ecosystem support. A later republished analysis relayed company disclosures of more than $300 million of ARR, more than 30 million users, and a nearly $300 million B+ round at a valuation above $2 billion. Those figures indicate scale and substantial capital access, but they do not disclose audited revenue recognition, gross margin, retention, burn, cash, runway, or customer concentration. Monitoring should therefore focus on verified label-control exceptions, rights complaints and takedown recurrence, model or cloud concentration, service availability, gross margin after model and compute costs, cash runway, and renewal quality. Any injunction, repeated unremediated infringement, prolonged critical outage, funding shortfall, or inability to reconcile recurring revenue to cash collections should pause investment.[CR011, CR013, CR014, CR015]

7.5 Exhibits

8.1 Investment call and price discipline

Evoken has a credible scale signal: an independent venture-news account reports a $300 million Series B+ at a post-money valuation above $2 billion, while the same account reports ARR above $300 million as of May 2026. That arithmetic suggests a headline valuation-to-ARR ratio below 6.7x, but it is not a verified enterprise-value multiple because the valuation, ARR definition, revenue mix, debt, cash, and preference stack are not disclosed. A separate June 2026 funding roundup also reports Liblib, an Evoken product, raised $300 million at a valuation above $2 billion and had more than 30 million cumulative users. The decision implication is track: the price may be reasonable if ARR is recurring, gross margins are strong, and growth is durable, but public evidence does not support a buy recommendation or a target return calculation.[CV001, CV002, CV003, CV004]

8.2 Thesis versus anti-thesis

The positive thesis is a multi-product workflow rather than a single generation feature. The reported Evoken portfolio links LiblibAI's creator community, LibTV's video workflow, and Xingliu's design agent; the account says LiblibAI had more than 30 million users, 500,000 original models, and that LibTV served nearly 1,000 teams. Industry evidence supports a large addressable use case: a May 2026 report describes China's micro-drama market as projected above 120 billion yuan, with more than 50,000 AI-native titles added to Douyin in March and production costs around one-tenth of live action. The anti-thesis is that capital and attention may outrun monetization: the same market report says output quality is uneven and cites an overseas micro-drama company that generated approximately $400 million in 2024 revenue but remained unprofitable. Evoken therefore needs cohort-level retention, contribution margin, and customer concentration evidence before the thesis can support a premium multiple.[CV005, CV006, CV007, CV008]

8.3 Scenarios, comparables, and downside

A scenario framework is more defensible than a single point valuation because the public record does not disclose audited revenue, gross margin, net revenue retention, burn, or cash runway. SaaS valuation guidance distinguishes ARR from total or forward revenue and says investors should test ARR against growth, retention, gross margin, CAC payback, and profitability. A 2026 market report places AI-native software at a median 21.2x EV/revenue in venture rounds and 11.5x in M&A, versus 5.5x and 3.8x for legacy SaaS, but those are broad category benchmarks rather than Evoken comparables. A second benchmark source reports 4.4x median implied ARR for public software in October 2026 and 21.7x for the top ten, underscoring dispersion. Applying a conservative 4x–8x range to the reported $300 million ARR would imply $1.2–$2.4 billion before adjustments for ARR quality, cash, debt, and preferences; this is an analytical range, not a fair-value estimate.[CV009, CV010, CV011, CV012]

8.4 Diligence asks and thesis-break triggers

Exit readiness is not established by a private round alone. The next diligence package should reconcile ARR to invoices and recognized revenue by product, show gross margin after model and inference costs, provide cohort retention and expansion, disclose customer and channel concentration, and provide the capitalization table with liquidation preferences, option pools, debt, and the exact post-money basis. The company also needs to evidence whether reported growth came from repeat professional production budgets or from launch-period usage and promotional activity. Regulatory review matters because the target market is content generation: a cited industry account describes tiered review and platform checks, including provincial approval for productions above one million yuan and removal of more than 25,000 episodes for violations. A thesis break would be material ARR restatement, sustained gross margin compression from third-party model costs, net retention below plan, a down round, or inability to renew major production accounts.[CV013, CV014, CV015, CV016]

8.5 Exhibits