Nunchaku Releases 4-Bit Lightning Qwen-Image-Edit-2509 Model
The Nunchaku team has recently released a new 4-Bit Lightning version of the Qwen-Image-Edit-2509 model, bringing users a faster image editing experience.
Key Updates
Two days after the original 4-bit Qwen-Image-Edit-2509 model was released, the Nunchaku team launched this Lightning version, specifically designed for users who found the original model’s running speed insufficient. The new version incorporates Lightning LoRA technology and supports 4/8 step fast inference.
System Compatibility
An important feature of the new version is its more hardware-friendly requirements. It can run smoothly even on relatively low-specification devices:
- Minimum Configuration Requirements: 8GB VRAM + 16GB RAM
- Optimization Suggestions: Optimal performance can be achieved by adjusting the
num_blocks_on_gpuanduse_pin_memoryparameters
It’s worth noting that users do not need to update their existing software version (wheel v1.0.0 or ComfyUI-nunchaku v1.0.1) to use the new model.
Download Options
Users can download the new model through the following channels:
- Hugging Face: https://huggingface.co/nunchaku-tech/nunchaku-qwen-image-edit-2509
- ModelScope: https://modelscope.cn/models/nunchaku-tech/nunchaku-qwen-image-edit-2509
Usage Instructions
The team has provided detailed usage examples:
- Diffusers Usage Example: https://github.com/nunchaku-tech/nunchaku/blob/main/examples/v1/qwen-image-edit-2509-lightning.py
- ComfyUI Workflow: https://github.com/nunchaku-tech/ComfyUI-nunchaku/blob/main/example_workflows/nunchaku-qwen-image-edit-2509-lightning.json
Note that the ComfyUI workflow requires ComfyUI version ≥ 0.3.60.
Future Plans
According to the development team, they are currently working on the following features:
- FP16 Support: Providing higher precision inference options
- Custom LoRA Support: Allowing users to utilize personalized LoRA models
Additionally, the team is actively developing the Wan2.2 project, bringing more innovative features to users.