Introduction
Liblib.art is China’s largest AI model-sharing community, and its feature set — model hosting, cloud generation, LoRA training, and community tools — generates many questions from both new and experienced users. This FAQ addresses the most common questions based on community forums, user reviews, and platform documentation.
All information is current as of March 2026.
Model Uploading
How do I upload a model to Liblib?
- Log into your Liblib account
- Navigate to “Upload Model” (上传模型) in your creator dashboard
- Select the model type (Checkpoint, LoRA, Textual Inversion, ControlNet, Workflow)
- Upload the model file (supported formats: .safetensors, .ckpt, .pt)
- Fill in model metadata: name, description, tags, base model, trigger words
- Upload sample images showing the model’s output
- Select a license type (open, restricted, or commercial)
- Submit for review
What file size limits apply to uploads?
- Single file: Maximum 10 GB
- Total storage per free account: 5 models (no total storage limit per model)
- Total storage per Standard account: 20 models
- Total storage per Pro account: 100 models
- Total storage per Premium account: Unlimited models
Do uploaded models go through review?
Yes. All uploaded models go through a review process that checks for:
- Malware or malicious code embedded in model files
- NSFW content in sample images (prohibited on the platform)
- Copyright infringement (obvious cases are flagged)
- Accurate metadata and descriptions
Review typically takes 1–24 hours. During peak periods, review may take up to 48 hours.
Can I upload a model and keep it private?
Yes. You can set model visibility to:
- Public: Visible to all Liblib users, searchable, downloadable
- Unlisted: Accessible via direct link only, not searchable
- Private: Visible only to you (requires Pro or Premium plan)
Can I monetize my uploaded models?
Yes, through Liblib’s creator program:
- Free models: Earn a small per-download payment from Liblib’s creator fund
- Premium models: Set a credit price for downloads; you receive 70% of the credits
- Tips: Users can tip creators directly
- Revenue sharing: Models used in cloud generation earn micro-payments per generation
What model formats are supported?
- .safetensors (recommended — most secure format)
- .ckpt (supported but .safetensors preferred for security)
- .pt (for textual inversions and embeddings)
- ComfyUI workflow files (.json)
LoRA Training
How do I train a LoRA on Liblib?
- Navigate to “LoRA Training” (LoRA训练) in your dashboard
- Select the base model (SD 1.5, SDXL, or SD 3.x)
- Upload your training dataset (10–50 images recommended)
- Configure training parameters:
- Learning rate (recommended: 1e-4 to 5e-4)
- Number of epochs (recommended: 10–30)
- Batch size (2–8 depending on image resolution)
- Network dimension (typically 32–128)
- Network alpha (typically equal to dimension)
- Regularization images (optional but recommended)
- Start training (costs 50–300 credits depending on settings)
- Training completes in 30–90 minutes
- Download the trained LoRA or use it directly in cloud generation
What images should I use for LoRA training?
Recommended:
- 10–50 high-quality images (more is not always better)
- Consistent subject with varied poses, angles, and lighting
- High resolution (1024×1024 or larger)
- Clean backgrounds or varied backgrounds (depending on your goal)
- Proper captioning (describe each image accurately)
Avoid:
- Blurry, low-resolution, or watermarked images
- Fewer than 10 images (underfitting risk)
- More than 100 images without careful curation (overfitting risk)
- Inconsistent subjects across images
- Screenshots with UI elements or text overlays
How much does LoRA training cost?
| Training Type | Approximate Credit Cost | Approximate CNY Cost |
|---|---|---|
| Basic (SD 1.5, 20 images, 10 epochs) | 50 credits | ~¥5 ($0.70) |
| Standard (SDXL, 30 images, 20 epochs) | 100 credits | ~¥10 ($1.40) |
| Advanced (SDXL, 50 images, 30 epochs, high dim) | 200–300 credits | ~¥20–30 ($2.80–$4.20) |
Can I train on copyrighted images?
Liblib’s terms of service state that users are responsible for ensuring they have the right to use training images. The platform does not actively verify copyright ownership of training data. However:
- Using copyrighted images to train a LoRA for personal, non-commercial use is generally lower risk
- Using copyrighted images to train a LoRA for commercial distribution may constitute copyright infringement
- Models that produce output closely resembling specific copyrighted works may be flagged and removed
Best practice: Use images you own, have licensed, or that are in the public domain.
Can I keep my trained LoRA private?
Yes. Trained LoRAs can be set to private (visible only to you) or shared with the community. Private LoRA storage counts toward your account’s model storage limit.
Copyright and Intellectual Property
Who owns AI-generated images on Liblib?
Under Chinese law as of 2026, AI-generated images are in a legal gray area. Liblib’s terms state:
- Users who generate images retain rights to their output
- Liblib does not claim ownership of user-generated content
- Commercial use of generated images is permitted on paid plans
- Users are responsible for ensuring their output does not infringe on third-party IP
Can I use Liblib-generated images commercially?
Yes, on paid plans (Standard, Pro, Premium, Enterprise). The commercial license covers:
- Marketing and advertising
- Product design and merchandise
- Game assets and concept art
- Social media content
- Print and digital publication
Important: The commercial license covers the generated output. It does not guarantee that the output does not inadvertently resemble copyrighted material. Users should exercise due diligence.
What happens if my model or generation infringes copyright?
Liblib operates a takedown process:
- Copyright holder submits a complaint with evidence
- Liblib reviews the complaint (typically within 72 hours)
- If valid, the infringing content is removed
- Repeat offenders face account suspension
- Users can appeal takedowns through the platform
Are there restrictions on model training data?
Liblib’s terms prohibit training on:
- Images depicting real minors
- Pornographic or explicit content
- Political content prohibited by Chinese regulations
- Images with clear “do not use for AI training” notices from rights holders
API Access
Does Liblib offer an API?
Yes, API access is available on Pro, Premium, and Enterprise plans.
API capabilities:
- Image generation (all supported models and settings)
- LoRA application at generation time
- Model search and discovery
- Batch generation
- Generation status tracking and result retrieval
API limitations:
- Rate limits based on plan tier (Pro: 10 req/min; Premium: 50 req/min; Enterprise: custom)
- No direct LoRA training via API (training is UI-only)
- No model upload via API
- Authentication via API key (no OAuth)
How do I get API access?
- Upgrade to Pro, Premium, or Enterprise plan
- Navigate to “API Settings” in your account dashboard
- Generate an API key
- Review the API documentation (available in Chinese)
- Integrate using standard REST API calls
What is the API pricing?
API calls consume credits at the same rate as UI generation. There is no separate API pricing — you use your daily credit allocation.
Is there an SDK?
Liblib provides a Python SDK for API integration. It is documented in Chinese with code examples.
Content Moderation
What content is prohibited on Liblib?
Per Chinese internet regulations and platform policy:
- NSFW/explicit content: All pornographic or sexually explicit content
- Political content: Images depicting political figures, sensitive historical events, or political commentary
- Violence: Graphic violence or gore
- Deepfakes: Realistic depictions of real people without authorization
- Illegal content: Anything illegal under Chinese law
- Spam: Mass-uploaded low-quality or duplicate models
How is content moderated?
Liblib uses a multi-layer moderation system:
- AI filter: Automatic screening of all generated and uploaded images
- Upload review: Human review of all model uploads (sample images and descriptions)
- Community reporting: Users can flag inappropriate content
- Periodic audits: Platform staff review trending and featured content
What happens if my content is flagged?
- First offense: Content removed, warning issued
- Second offense: Content removed, temporary generation restriction (24–72 hours)
- Third offense: Account suspension (7–30 days)
- Severe violations: Permanent account ban
Are AI-generated images watermarked?
Yes. Per Chinese AI content labeling regulations (effective 2024), all images generated through Liblib’s cloud service include:
- Invisible watermark: Machine-readable metadata identifying the image as AI-generated
- Optional visible watermark: Users can enable/disable a visible “AI Generated” mark
Account Management
How do I create an account?
Registration requires:
- Phone number (Chinese mobile number for SMS verification)
- OR WeChat account login
- OR email address (with verification)
International users can register with email but may have limited access to certain payment methods.
Can I delete my account?
Yes. Account deletion removes:
- All personal data
- All generated images stored on the platform
- All uploaded models (public models will be removed)
- All trained LoRAs stored on the platform
Account deletion is processed within 7 days. It is irreversible.
Is there multi-user/team support?
Team management is available on Enterprise plans only. Pro and Premium plans are single-user.
Can I use Liblib from outside China?
Yes, but with limitations:
- The interface is in Chinese only
- Payment is primarily through Chinese methods (Alipay, WeChat Pay)
- Some international users report slower access from overseas locations
- Customer support is in Chinese
Troubleshooting
My generation is stuck in queue — what should I do?
During peak hours (evenings in China, ~7–11 PM Beijing time), free-tier users may experience 1–10 minute queue times. Paid plans have priority queues that reduce wait times.
Options:
- Wait for the queue to process
- Try generating during off-peak hours
- Upgrade to a paid plan for priority queue access
- Reduce generation complexity (lower resolution, fewer steps)
My LoRA training failed — what happened?
Common causes:
- Insufficient credits (training requires credits upfront)
- Corrupted or incompatible training images
- Images too small (minimum 512×512 recommended)
- Training parameters too aggressive (learning rate too high)
- Server-side error (retry after 30 minutes)
My model upload was rejected — why?
Common rejection reasons:
- Sample images contain NSFW content
- Model description is incomplete or misleading
- Model file appears corrupted or is an unsupported format
- Model duplicates an existing upload without meaningful improvement
Conclusion
Liblib.art is a comprehensive platform with clear policies and capabilities. The most important things to understand are the credit system (daily allocation + packs), the content moderation rules (Chinese regulations are strict), and the API access requirements (Pro plan or above). For most users, the platform is straightforward to use once the credit economics are understood.
References
- Liblib.art Help Center — https://www.liblib.art/help
- Liblib.art Terms of Service
- Liblib.art Privacy Policy
- Liblib.art Creator Program Documentation
- Liblib.art API Documentation
- “Chinese AI Content Labeling Regulations,” Cyberspace Administration of China, 2024
- “Copyright and AI-Generated Content in China,” China IP Law Blog, 2025
- “LoRA Training Best Practices,” Stable Diffusion Community Wiki, 2025
- “Chinese Internet Content Moderation Requirements,” China Law Blog, 2025
- “Credit-Based Pricing in Chinese SaaS Platforms,” TechNode, 2025