LiblibAI 2.0 (LibLibrary) Market Research

1. Introduction

LiblibAI (Chinese brand name "LibLibrary"/哩布哩布) is one of China's largest-scale AI creation communities and online image-generation platforms. Built on an "open-source model ecosystem + modular tool flow" core architecture, it moves the models, LoRAs and ComfyUI workflows that were previously scattered across individual GPUs in the Stable Diffusion ecosystem to the cloud and runs them as a community, letting creators generate images with model assets published by others without deploying their own GPU.

In October 2025, the platform released LiblibAI 2.0, upgrading itself from a "model aggregation tool" into an "AI professional creation studio" that integrates image generation, video generation, video effects, an inspiration community, AI applications and model training into the same workbench. Among the 16 platforms in this research, LiblibAI is the only one whose core moat is "community-shared model and workflow assets" rather than "self-developed models" or "end-to-end product experience", which gives it unique sample value within the AI Harness six-layer model — it extends the persistent objects of L4 (Memory & State layer) from "personal assets" to a "community public asset library".

1.1 Basic Information

1.2 Development History

1.3 Position in the AI Harness System

Under this project's six-layer capability model (per the parameter card), LiblibAI's core contribution lies not in model intelligence itself, but in externalizing the engineering knowledge that was originally implicit in individual computers along the chain of "model selection — parameter combination — workflow orchestration — asset accumulation" into retrievable, reusable and tradable community assets.

More specifically within the six-layer model:

• It packages the "prompt + model + LoRA + parameters" of L1 (Context Engineering) into one-click reusable combinations, lowering the barrier to context construction.
• It upgrades L4 (Memory & State) from a "personal asset library" into a three-tier structure of "community model library + workflow template library + personal asset library", making it the platform that does community-level asset persistence most thoroughly in this research.
• On L3 (Orchestration & Control) it provides node-based workflows via built-in ComfyUI, but lacks task-level state machines, interrupt-resume and sub-agent dispatch.
• It is clearly weak on L5 (Evaluation & Observation) and L6 (Governance & Safety): no public Eval Set, and the 2026-04 regulatory naming exposed substantive flaws in its content-safety guardrails.

In one sentence: LiblibAI is an asset-community Harness with "strong L1 + strong L4 (community-level) + moderate L3 + weak L5 + weak L6".

2. Glossary

3. Feature Description

3.1 Three Sets of Operation Interfaces

LiblibAI 2.0 offers three operation interfaces that progress from simple to advanced, covering the full spectrum from complete beginners to professional parameter tuners:

This design is essentially a tiering of "context exposure level": the platform turns the complexity of L1 context engineering into switchable levels, letting users decide how much context-construction responsibility they take on.

3.2 Image Generation & Model Aggregation

LiblibAI does not push its own self-developed image foundation model; instead it aggregates mainstream closed-source and open-source models in a "model supermarket" way (version names change with platform iterations; the following is a snapshot at the time of retrieval):

Supporting capabilities include: text-to-image, image-to-image, multi-LoRA stacking with weight adjustment, style-tag combinations (e.g. "anime + cyberpunk + photographic realism" each with individual weights), automatic prompt optimization and Chinese-English translation.

3.3 Video Generation

Version 2.0 elevated video generation to a position alongside images: it supports text-to-video and image-to-video, and embeds 500+ video-effects templates (covering camera movement, style transfer, action/expression, particle effects, etc.). The operation path is "upload first frame → select effect → generate", requiring no parameter configuration. According to third-party summaries, the platform can invoke models such as Tongyi Wanxiang to generate videos over 10 seconds during peak hours (Low-Medium confidence).

3.4 Online LoRA Training

The platform provides a no-code online training capability, a key step in productizing L4 (asset persistence):

• Training sample volume: upload about 10—50 images of the same theme (different sources cite 20—50 or 10—30 images, ).
• Preset modes: portrait, art style, ACG, product, etc.
• After training, the LoRA can be used directly for image generation with highly consistent style; the trained LoRA can enter the personal asset library or be optionally published to the community.
• Free users get a limited number of training opportunities per month, with more for members (exact counts).

3.5 Post-Processing & AI Toolbox

The platform embeds a set of post-processing capabilities that need no jump to external tools: high-res fix, intelligent outpainting, one-click background removal, local inpainting, face restoration, prompt reverse, watermark removal, line-art extraction, as well as e-commerce-oriented "quick apps" such as product-scene fusion, AI model generation, virtual try-on and poster layout. Third-party summaries say the tool count reaches 6,000+ AI applications (Low confidence).

3.6 Asset Library & Community Ecosystem

• The inspiration page aggregates a large number of community works, categorized by style, model and popularity, supporting "make the same style".
• A hundred-thousand-scale LoRA style model library, searchable and instantly usable (scale measures conflict; see 1.1).
• The platform claims to use blockchain for creation provenance and copyright notarization, building a full ecosystem chain from creation, sharing and copyright to sales.
• Offers tutorials and a creation center to lower the learning cost for new users.

4. Platform Architecture

图 4-1｜LiblibAI 平台总体架构：五层分层与社区资产核心

数据来源：基于本文分析绘制的示意图。

4.1 Overall Architecture Layers

4.2 Model Aggregation & Routing Strategy

The platform explicitly adopts a "compatible with both open-source and closed-source models" strategy: on the image side it assembles Qwen-Image, FLUX series, Kontext, Seedream, Midjourney, etc., and on the video side it embeds Kling, Hailuo, Vidu, Tongyi Wanxiang and others. This model brings two engineering consequences:

1. Positive: a single account and single top-up system allows cross-comparing and switching models, lowering the cost of evaluating multiple platforms.
2. Negative: the model combination adjusts dynamically with copyright and partnership conditions, so the same instruction may yield different output styles at different times. For production scenarios requiring reproducibility this is a material risk — important projects should lock in the model version at the time and keep intermediate drafts.

4.3 Compute Platform & Metering System

• Metering unit: compute point (Compute Point); a single image at standard resolution is about 1 point.
• Free quota: registered users claim it via daily login, about 300 points/day (the measure says this can generate about 300 standard-resolution images).
• Membership tiers determine the monthly compute quota, cloud-storage capacity, acceleration count and parallel task count (see 1.1).
• Compute packages: purchased per batch, suitable for temporary heavy image output.

5. Harness Design

5.1 L1 Context Engineering Layer

LiblibAI's context construction is composed of four kinds of elements: base Checkpoint + LoRA combination (with weights) + prompt (positive/negative) + sampling parameters (sampler/steps/CFG/seed).

Its engineering value lies in the fact that this combination is fully recorded and one-click reproducible in the community — when a user sees a work they like, they can directly "make the same style", inheriting the other person's model, LoRA, prompt and parameters wholesale. This effectively turns L1's context construction from "personal tacit knowledge" into "community-transferable explicit artifacts".

But this also brings risk: the quality of community works' context varies widely, and the platform provides no context-legitimacy checks (such as the portrait-rights/copyright provenance of LoRA training material). Context can be inherited, but responsibility does not transfer with it.

Evaluation: strong (community-level context reuse), but lacking compliance checks.

5.2 L2 Tools & Execution Layer

The tool forms come in three tiers of "node / quick app / button":

• Node-level (ComfyUI): each capability is a node, freely combinable, supporting controlled-generation plugins such as ControlNet.
• Quick-app level (AI toolbox): encapsulated single-step capabilities such as HD upscaling, deblurring, outpainting, face swap, outfit change, virtual try-on, background removal, watermark removal and e-commerce product fusion.
• Button-level (Minimal generator / WebUI): fixed parameter panels.

No open tool-registration mechanism for external developers or MCP-like protocol support was found (this judgment is based on public materials).

Evaluation: Medium-Strong (broad tool coverage, but no open tool contract).

5.3 L3 Orchestration & Control Layer

• Workflow-as-orchestration: the built-in ComfyUI supports visual node orchestration and can import workflow templates shared by the community for batch generation and AIGC pipelines. A workflow is essentially a serializable JSON graph with the engineering basis for versioning and reuse.
• Missing items: no production-grade orchestration capabilities such as task-level state machines, sub-agent dispatch, interrupt & resume or failure retry were found; batch processing behaves more like "parallel submission of many tasks" than "dependency-aware DAG scheduling".
• Notably, the platform's agent capabilities (Lovart / Xingliu) are not sunk down into the LiblibAI main site, but exist as independent products (see Article 14).

Evaluation: Medium (workflow orchestration is usable, but lacks production-grade control semantics).

5.4 L4 Memory & State Layer

This is the layer where LiblibAI is most differentiated relative to its peer platforms. Its persistent objects are divided into three tiers:

Compared with other platforms in this group: Kling's "Element" and Runway's "Brand Kits" are personal/team-level assets; LiblibAI instead takes the asset library to the community level, making "reusing others' already-consolidated style assets" the default workflow.

Weakness: community assets depend on third-party uploaders for maintenance, and models may be taken down, deleted or replaced, causing workflows that depend on those assets to fail — a unique asset-availability risk of community-type L4.

Evaluation: strong (community-level), but asset availability is not guaranteed by the platform.

5.5 L5 Evaluation & Observation Layer

• Observation measure: compute-point consumption is the unified metric; users can see task duration, parallel queue and generation history.
• No official Eval Set, Golden Dataset or regression-set mechanism was found; model-quality evaluation relies on subjective community feedback (likes, favorites, "make the same style" counts).
• Lacks production-oriented reproducibility guarantees: the model combination adjusts dynamically, and the same workflow may produce different results at different times.
• Reference comparison: FLUX.2 offers a "fixed snapshot endpoint" to explicitly guarantee reproducibility (see Article 9); no such mechanism was found for LiblibAI.

Evaluation: weak.

5.6 L6 Governance & Safety Layer

This is the layer with the most prominent risk for this platform within this set of documents.

• Compliance foundation: the platform claims to be among the first in China to pass generative-AI filing, and uses blockchain for creation provenance and copyright notarization (specific implementation).
• Regulatory events that have occurred: in 2026-04, CCTV exposed serious regulatory loopholes in several AI applications, among which "LiblibAI AI" was named — using vague prompts bought online, it could generate semi-nude female dancing videos within just a few minutes, and the report noted the app could still be downloaded normally from app stores at the time. Its operating entity, Beijing Qidian Xingyu Technology Co., Ltd., subsequently entered the public eye (Medium confidence; recommend re-verification against the regulatory notice).
• Structural risk: the model of a UGC model community as the core asset disperses content-safety responsibility across a large number of third-party uploaders. LoRAs can be used to consolidate specific real-person likenesses, workflows can be used to mass-produce violating content, and the platform's guardrails must cover all three positions of "model layer + prompt layer + output layer" — missing any one of them may be circumvented.
• Labeling compliance: no official implementation description was found for the platform's explicit labeling and implicit metadata labeling under the Measures for the Labeling of AI-Generated Synthetic Content. [To be filled]
• Portrait-rights risk: community LoRAs extensively involve consolidating real-person likenesses, and Articles 1018 and 1019 of the Civil Code of the People's Republic of China regarding portrait rights should apply; no public description of the platform-level authorization chain and review mechanism was found.

Evaluation: weak, and with an actual recorded regulatory event.

5.7 Six-Layer Maturity Summary

6. Real-World Cases

6.1 Verifiable Public Data

The following are quantitative data traceable to media reports or the platform's public statements, which should be marked as point-in-time snapshots:

6.2 Typical Use Cases

The following are typical workflows recorded by third-party summaries (use-case descriptions, not business cases with effect data):

• Design & creation: illustrations, posters, Logo drafts, e-commerce main images, architectural design renderings; reusing high-quality community models reduces ideation time.
• Self-media & short video: covers and supporting images for Douyin and Bilibili short videos, mass-producing alternative options.
• E-commerce operations: product scene images and marketing creative graphics, placing products into target scenes via "product fusion" quick apps.
• Brand-specific models: designers upload batches of images to train brand-specific LoRAs, consolidating a brand's visual style.
• Novel promotion / cultural-creative merch: using style LoRAs to mass-produce series of promotion-image graphics and merchandise design drafts.

6.3 Items Not Found

• Officially published brand or merchant customer cases with quantitative effect data: none found.
• Implementation details of the platform's explicit/implicit labeling under the Measures for the Labeling of AI-Generated Synthetic Content: no official description found.
• Accurate sample-volume requirements for online LoRA training, free users' monthly training count, and the API's open scope and call pricing: no reliable official description found for any of these.

7. Summary

7.1 Strengths

1. Asset scale & reuse efficiency: the model-library scale is in the domestic first tier, and community sharing lets an individual creator directly inherit others' already-consolidated styles and workflows, significantly lowering context-construction cost.
2. Three interface tiers cover the full spectrum: Minimal generator / WebUI / ComfyUI let beginners and professionals each take what they need; one account bridges image and video, avoiding duplicate top-ups across platforms.
3. No-code LoRA training: it reduces model fine-tuning from "requiring a GPU and engineering skills" to "uploading images and choosing a mode", a key productization step for L4 asset persistence.
4. Community-level L4 is a unique value: among the 16 platforms, only LiblibAI takes persistent assets to the community public-good level.
5. Compliance foundation in place: claims to be among the first in China to pass generative-AI filing, and uses blockchain for creation provenance and copyright notarization.

7.2 Weaknesses & Risks

1. Weak L5 evaluation layer: no official Eval Set, and dynamically adjusted model combinations make outputs non-reproducible across time points, unsuitable for production pipelines with strict consistency requirements.
2. L6 governance layer already has an actual incident: the 2026-04 CCTV naming exposed flaws in the content-safety guardrails; the UGC model community model highly disperses responsibility.
3. Asset availability is not guaranteed: models and workflows that depend on third-party uploaders may be taken down, deleted or replaced, causing downstream workflows to fail.
4. Business-model pressure: aggregating third-party models brings ongoing cost pressure; against giants like ByteDance's Jimeng pursuing "self-developed models + one-click consumer experience", traffic measures have already shown MoM decline (2025-09, SimilarWeb).
5. Confused pricing measures: at least three conflicting membership-pricing sets were found in public searches; enterprises must defer to the official real-time prices before purchasing.
6. Portrait-rights concerns: community LoRAs extensively involve real-person likenesses, and no public description of the authorization chain and review mechanism was found.

7.3 Applicability Boundaries

7.4 选型建议

• **个人创作者与中小电商团队**：LiblibAI 是性价比极高的起点，建议以「基础版 VIP + 按需算力包」组合起步，优先复用社区成熟 LoRA 而非从零训练。
• **有品牌一致性要求的企业**：可将 LiblibAI 用于前期风格探索与素材初稿，但**不应**作为最终交付的唯一生产环境；应把选定的模型版本、LoRA 权重与参数固化为企业自有的工作流资产（可考虑导出至 ComfyUI 自托管，详见第 12 篇），以摆脱对社区资产可用性的依赖。
• **对合规要求严格的行业（金融、医疗、政务）**：在平台就《标识办法》的显式/隐式标识实现、以及社区 LoRA 的授权审核机制作出公开说明之前，不建议用于对外交付内容。
• **需要 Agent 化设计编排的团队**：应评估同门的 Lovart / 星流（详见第 14 篇），而非 LiblibAI 主站。

信息缺口声明

1. **模型库规模口径冲突**：存在「10 万+」与「50 万+」两套公开口径，未能定位官方权威数据。****
2. **会员定价**：检索到 ¥35/月、¥36/月、¥70/月、¥199/月、¥299/年 等多套冲突口径，均来自第三方梳理，非官方定价页。**[待填写]**
3. **运营主体称谓**：「北京奇点星宇科技有限公司」与「演语科技（Evoken）」在报道中交替出现，两者法律关系未获权威确认。****
4. **API 开放程度**：官网设有 API 入口，但未检索到公开的 API 文档、调用定价与并发限制。**[待填写]**
5. **在线 LoRA 训练样本量要求**：不同来源给出 10—30 张、20—50 张等不同口径。****
6. **2026-04 央视点名事件**：目前仅检索到百科词条转引凤凰网财经与潇湘晨报，未定位到央视原始报道或监管通报。**建议以官方通报复核。**
7. **区块链创作溯源与版权存证**：平台宣称有此能力，未检索到技术实现说明与存证效力说明。**[待填写]**
8. **《标识办法》合规实现**：未检索到平台就显式标识与隐式元数据标识的官方说明。**[待填写]**
9. **官方客户案例与量化效果数据**：未检索到，第 6.3 节已如实标注。
10. **AI 工具箱「6,000+ 应用」「500+ 视频特效模板」等数量口径**：均来自第三方导航站，**低置信，**。

8. 参考资料

1. LiblibAI 官方网站 — 北京奇点星宇科技有限公司，2026。https://www.liblib.art
2. 单笔融资额超越 Manus，这家 AI 公司瞄向全球化 — 证券时报，2025-10-23。https://www.stcn.com/article/detail/3399818.html
3. 陈冕 — 百度百科（含 LiblibAI 成立、Lovart 上线、2026-04 央视点名等条目与引注）。https://baike.baidu.com/item/%E9%99%88%E5%86%95/66304026
4. 陈冕 — 搜狗百科（含演语科技融资历程与 ARR 口径）。https://baike.sogou.com/v10000726100.htm
5. 超 Manus！LiblibAI 获 1.3 亿美元融资成 2025 年 AI 应用单笔最大 — 太平洋电脑网，2025-10。https://g.pconline.com.cn/ai/article/1457520.html
6. LiblibAI 简介 — 映技派（含模型库规模、定价档位、视频模型清单）。https://www.yjpoo.com/site/997.html
7. LiblibAI — 十万模型一站生图创作 — 今日头条（含模型规模、备案与区块链存证表述）。https://www.toutiao.com/a7674479798168715822
8. Liblib AI 画图模型 — 快搜（含 LiblibAI 2.0 三大模块整合说明）。https://www.kuaisou.com/docs/20260613-liblib-ai-hua-tu-mo-xing.html
9. LiblibAI 平台功能梳理 — 6035 工具站（含模型聚合清单与工具箱描述）。https://www.6035.com.cn/sites/830.html
10. LiblibAI·哩布哩布 AI 使用评测 — 百易 AI 导航（含会员档位与训练能力描述）。https://www.baiyiai.com/ai-image-tools/site/493.html
11. 星流 AI 产品解析 — 青衣网络（含 Lovart / 星流 / LiblibAI 三者关系）。https://www.ra0.cn/?p=13081/
12. 《人工智能生成合成内容标识办法》— 中央网信办等四部门，2025-03-14 发布，2025-09-01 施行。https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm
13. 《人工智能生成合成内容标识办法》解读 — 中国政府网，2025-03-16。https://www.gov.cn/zhengce/202503/content_7014281.htm
14. 《中华人民共和国民法典》第一千零一十八条、第一千零一十九条 — 全国人民代表大会，2020。（肖像权定义与禁止以信息技术手段伪造侵害肖像权）