AI Agent - Mar 19, 2026

Liblib.art 2026 FAQ: Model Upload, LoRA Training, Copyright Policy, and API Access Explained

Introduction

Liblib.art is China’s largest AI model and workflow community, hosting 120,000+ models and serving hundreds of thousands of creators. As the platform has grown, so have the questions from new and existing users.

This FAQ compiles the most commonly asked questions about Liblib.art as of March 2026, organized by topic. Whether you’re a first-time visitor wondering how to get started or a power user with questions about API access and copyright policy, this guide has you covered.

Getting Started

What is Liblib.art?

Liblib.art is a Chinese-language platform for hosting, discovering, training, and using AI image generation models. Think of it as China’s Civitai — a community-driven hub where creators share Stable Diffusion checkpoints, LoRA models, and ComfyUI workflows.

How do I create an account?

You can register using:

Phone number (mainland China mobile numbers) — the most common method
WeChat login — one-tap authentication via WeChat
Email — available but less common among Chinese users

Real-name verification is required for uploading models or using certain features, in compliance with China’s generative-AI regulations.

Is Liblib.art free to use?

Yes, there is a free tier that includes:

10 daily generation credits (use-it-or-lose-it)
Access to the full model library for browsing and testing
1 LoRA training job per week
Gallery browsing and community participation

Paid plans start at ¥29/month for additional credits, features, and priority access. See our Liblib.art Pricing 2026 article for a detailed breakdown.

Can I use Liblib.art outside of China?

Yes, the platform is accessible internationally, though:

The interface is in Simplified Chinese with no official English translation
Inference speeds may be slower outside China (CDN is optimized for mainland)
Some payment methods (Alipay, WeChat Pay) require Chinese accounts
Real-name verification requires a Chinese ID for some features

International users who read Chinese can use the platform effectively, but the experience is optimized for mainland China.

How do I upload a model to Liblib.art?

Navigate to the Upload page from your profile menu
Select the model type (Checkpoint, LoRA, Embedding, VAE, etc.)
Upload the model file (supported formats: .safetensors, .ckpt, .pt)
Fill in metadata:
- Model name and description (Chinese or English)
- Category tags (anime, photorealistic, etc.)
- Base model compatibility (SD 1.5, SDXL, Flux, etc.)
- Upload 5–10 sample images with generation parameters
Choose visibility (Public or Private)
Submit for review

How long does model review take?

Most models are reviewed within 2–4 hours during business hours (9 AM – 9 PM CST). Weekend and holiday reviews may take up to 24 hours.

The review checks for:

Prohibited content (political, violent, pornographic)
Malware scanning of model files
Basic metadata quality (description, tags, sample images)

Can I upload models I downloaded from Civitai or Hugging Face?

It depends on the license:

Models released under open licenses (CreativeML Open RAIL-M, Apache 2.0, MIT) can generally be re-uploaded with proper attribution
Models with restrictive licenses (no redistribution, commercial restrictions) should not be re-uploaded without the original author’s permission
Liblib.art requires you to credit the original source when re-uploading

In practice, many early Liblib models were mirrors of Civitai uploads. The platform now encourages original uploads and has introduced a “Clean Data” badge for models with verified provenance.

Can I monetize my uploaded models?

Yes, Liblib offers several monetization options:

Method	How It Works	Commission
Premium model sale	Set a price (¥5–50 typical); users pay to download	15% to Liblib
Tip jar	Free model with optional donations	10% to Liblib
Challenge prizes	Win official creative challenges for cash prizes	Varies
Enterprise referrals	Verified creators may receive enterprise client referrals	Negotiated

Payouts are processed via Alipay or bank transfer, with a minimum withdrawal of ¥50.

LoRA Training

What is LoRA training?

LoRA (Low-Rank Adaptation) is a technique for fine-tuning AI image models using a small dataset. Instead of retraining the entire model (which requires enormous compute), LoRA modifies a small subset of model weights, producing a lightweight file (typically 10–100 MB) that can be layered on top of a base model to change its output style.

How does LoRA training work on Liblib.art?

Prepare your dataset — 15–50 images of the style, character, or object you want the LoRA to learn
Open the LoRA Training wizard — Liblib provides a step-by-step Chinese-language interface
Configure parameters:
- Training type (Style, Character, Object)
- Base model (SD 1.5, SDXL, etc.)
- Training steps (typically 1,500–3,000)
- Learning rate (default: 1e-4)
- Image resolution
Upload your dataset — Liblib auto-crops and optionally auto-captions your images
Start training — Jobs run on NVIDIA A100 GPUs; typical training takes 15–45 minutes
Review and publish — Test the trained LoRA, then publish it publicly or keep it private

How much does LoRA training cost?

Training Type	Typical Steps	Credit Cost	Approximate ¥ Cost
Basic style LoRA	1,500	30 credits	¥3
Standard character LoRA	2,500	50 credits	¥5
Advanced multi-concept LoRA	3,000+	60–100 credits	¥6–10

Free users get 1 training job per week. Standard and Pro subscribers have unlimited training.

What makes a good LoRA training dataset?

Best practices for dataset preparation:

Quantity: 15–50 images (more is not always better — quality matters more)
Consistency: All images should share the visual attribute you’re training (style, character features, etc.)
Variety: Within the target attribute, include variety in poses, lighting, and backgrounds
Resolution: At least 512×512; 1024×1024 is ideal for SDXL
Format: PNG or JPEG
Captions: Add text captions describing each image (Liblib can auto-generate these, but manual captions are more accurate)

What if my LoRA training results are bad?

Common issues and fixes:

Problem	Likely Cause	Fix
LoRA has no effect	Too few training steps	Increase steps to 2,500+
Output looks distorted	Overfitting — too many steps or too small dataset	Reduce steps or add more diverse images
Wrong features captured	Inconsistent dataset	Remove images that don’t match the target attribute
Color shift	Learning rate too high	Reduce learning rate to 5e-5
Only works with specific prompts	Overfit to caption text	Use simpler, more generic captions

Copyright and Licensing

Who owns the models I upload?

Liblib.art’s terms of service state that you retain ownership of models you upload. By uploading, you grant Liblib a license to host, display, and distribute the model according to your chosen visibility settings.

Who owns images I generate on Liblib.art?

Under Liblib’s terms and Chinese law as of 2026:

Images generated using open-source models — ownership is generally attributed to the person who generated them, subject to the model’s license terms
Images generated using premium/paid models — the model author may specify usage restrictions
Commercial use — generally permitted unless the specific model’s license prohibits it

Important caveat: AI-generated image copyright is an evolving legal area in China. The Beijing Internet Court’s 2023 ruling suggested that AI-generated images can be copyrighted if sufficient human creative input is demonstrated. However, case law is still developing.

Can I use Liblib-generated images commercially?

In most cases, yes. However, you should check:

The base model’s license — most open-source SD models allow commercial use
The LoRA’s license — some model authors restrict commercial use
Your Liblib plan — free-tier generated images may have watermark requirements
Chinese content regulations — AI-generated content used commercially must comply with labeling requirements

What about NSFW content?

Liblib.art prohibits NSFW content in compliance with Chinese regulations. This includes:

Pornographic or sexually explicit images
Graphic violence
Politically sensitive content
Content depicting real individuals without consent

Automated content filtering scans all uploads and generated images. Violations result in content removal and, for repeat offenders, account suspension.

What is the “Clean Data” badge?

The Clean Data badge is a voluntary certification for models trained exclusively on:

Creator-owned original artwork
Licensed stock imagery
Public domain source material
Datasets with verified consent from all depicted individuals

Models with the Clean Data badge provide stronger IP protection for commercial users and receive priority placement in search results.

API Access

Does Liblib.art have an API?

Yes, Liblib.art offers a REST API for Pro and Enterprise plan subscribers.

What can the API do?

Endpoint	Functionality
`/generate`	Submit generation jobs with model, LoRA, prompt, and parameter specification
`/models`	Search and list available models
`/status`	Check generation job status
`/result`	Retrieve generated images
`/train`	Submit LoRA training jobs
`/workflows`	Execute saved ComfyUI workflows

What are the API rate limits?

Plan	Concurrent Requests	Requests per Minute	Monthly Quota
Pro	5	30	Based on credits
Enterprise	Custom (typically 20–50)	Custom (typically 120+)	Custom

How do I get API access?

Subscribe to the Pro plan (¥139/month) or above
Navigate to Settings → API in your Liblib dashboard
Generate an API key
Review the API documentation (Chinese language, with code examples in Python and JavaScript)

Is there a Python SDK?

There is no official SDK, but the community has developed several unofficial Python wrappers. The API uses standard REST conventions with JSON payloads, making integration straightforward with libraries like requests.

Account and Billing

How do I cancel my subscription?

Navigate to Settings → Subscription → Cancel Plan. Cancellation takes effect at the end of your current billing cycle. You retain access to all paid features until then.

Do unused credits expire?

Daily free credits expire at midnight CST each day
Monthly bonus credits from paid plans roll over for up to 3 months (capped at 3× your monthly allotment)
Purchased credit packs never expire

Can I get a refund?

Liblib offers refunds within 7 days of purchase for unused subscriptions and credit packs. Partially used credits are not refundable. Contact support via the in-app WeChat channel.

How do I delete my account?

Navigate to Settings → Account → Delete Account. Account deletion is permanent and includes:

Removal of all uploaded models
Deletion of all generated images
Loss of all remaining credits
Removal from all community features

A 30-day cooling-off period allows you to reactivate before permanent deletion.

Technical Questions

What model architectures does Liblib support?

As of March 2026:

Stable Diffusion 1.5 — full support
Stable Diffusion XL (SDXL) — full support
Flux (Dev and Schnell) — full support
SD 3.0/3.5 — partial support (inference only)
AnimateDiff — supported via ComfyUI workflows
ControlNet — supported for all compatible base models

What browsers are supported?

Liblib.art’s web interface works best on:

Chrome 100+ (recommended)
Edge 100+
Safari 16+
Firefox 100+

The mobile app is available for iOS 15+ and Android 10+.

What file formats are supported for model upload?

.safetensors (recommended — safer and faster to load)
.ckpt (supported but less preferred)
.pt (for embeddings and some LoRAs)
Maximum file size: 10 GB per model file

How do I report a bug or request a feature?

Bug reports: Submit via the in-app feedback form or the official WeChat support group
Feature requests: Post in the community forum under the “Feature Requests” category
Urgent issues: Contact support directly via the WeChat channel linked in your account settings

Conclusion

Liblib.art has evolved into a comprehensive platform with features that serve everyone from casual hobbyists to professional studios. The key to getting the most out of it is understanding the credit system, choosing the right plan for your usage pattern, and taking advantage of the community ecosystem — especially LoRA training and workflow sharing.

If your question wasn’t covered here, the Liblib.art community forums and WeChat groups are active and responsive. The platform’s Chinese-language documentation is thorough, and community members are generally helpful to newcomers.