PrunaAI released three LoRAs with rank 32: the 6-step and 8-step versions require stacking with their own P-Q-Image-2.1 distilled LoRA (object removal LoRA strength 0.7), while the base version runs directly on Qwen-Image-2.1 for 40 steps, without the offset typical of few-step models. Usage involves roughly outlining the object in the photo with a red contour and adding a text instruction; the repository includes a script for drawing contours, generated based on the original image area, with the author's example being approximately 200 ten-thousand pixels.
Why it matters It removes object shadows and reflections along with the object itself, producing outputs aligned with the original image that are ready for use; combined with the distilled LoRA, it generates images in few steps without needing CFG.
Hugging Face likes 20 · Hugging Face trending score 20 · Hugging Face downloads 162
License: Research only · ComfyUI: Likely compatible · VRAM: Not specified
Reasons
- The model card provides a complete parameter table for all three versions and runnable diffusers examples, along with an online demo space
Usage
- Install the specified version of diffusers as per the model card, load Qwen-Image-2.1 and the corresponding LoRA, use draw_outline.py to draw a red contour on the photo before generating; the 6-step version is recommended by default
Risks
- Inherits the research license (qwen-research) from Qwen-Image-2.1, so commercial use is not permitted
- Requires a CUDA GPU; ComfyUI compatibility is not individually confirmed, though the base model natively supports it