概要
The repository provides distinct models for generating new images and editing existing ones. Its examples cover Qwen-Image-2512 for text-to-image generation and Qwen-Image-Edit-2511 for editing with multiple reference images. Choose between a prompt-only workflow and an image-plus-instruction workflow before loading a pipeline. Product announcements in the repository are separate from the checkpoints available in its deployment examples.
主な機能
- Targets image generation scenarios involving Chinese and other text.
- Provides a dedicated text-to-image pipeline.
- Edit-2511 examples accept multiple reference images.
- Exposes prompts, seeds and inference-step settings.
- Includes generation examples for multiple aspect ratios.
- Supports Diffusers integration and local demo workflows.
要件、インストール、クイックスタート
2. Install compatible Transformers as documented and run pip install git+https://github.com/huggingface/diffusers.
3. For generation, load Qwen/Qwen-Image-2512 with QwenImagePipeline.
4. For editing, load Qwen/Qwen-Image-Edit-2511 with QwenImageEditPlusPipeline.
5. Configure device, precision, image dimensions and seed for the selected pipeline.
6. Run the official minimal example before substituting your own prompt or reference images.
使用方法
Implementation notes
Generated lettering can still be wrong, and multi-image edits need checks for identity and placement. Change a small number of requirements per iteration and retain intermediate outputs. Measure peak memory, processing time and retry behavior before deploying a service.
モデルの互換性とユースケース
The generation example includes CPU and CUDA paths, while the editing example uses CUDA with bfloat16. Large checkpoints require substantial memory and storage; offloading and quantization change the deployment configuration and performance.
ライセンスとリスクに関する注意事項
The official README identifies Qwen-Image as Apache-2.0. Apply the terms associated with each downloaded checkpoint, dependency and input asset.