QwenEdit InSubject LoRA
Model Description
QwenEdit InSubject is a LoRA fine-tune for QwenEdit that significantly improves its ability to preserve subjects while making edits to images. It works effectively with both single subjects and multiple subjects in the same image. While the base model can perform various image edits, it often loses important subject characteristics or distorts the main subjects during the editing process. This LoRA addresses these limitations to provide more accurate subject-preserving image editing.
How to Use
To get the best results, use this prompt format:
Make an image of [subject description] in the same scene [new pose/action/details]
You can include "in the same scene" to preserve the original scene and background while modifying the subject's pose, clothing, or other details.
For example:
Make an image of the horned woman in the same scene seated on a low pink ottoman, adjusting the buckle on one of her matching blue heels while her other leg is delicately crossed, wearing a blue and gold dress with a ruffled collar, red lips and freckles, the vibrant pink background still filling the frame behind her.
use with diffusers
import torch
from diffusers import QwenImageEditPipeline
pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=torch.bfloat16)
pipe.to("cuda")
pipe.load_lora_weights("peteromallet/Qwen-Image-Edit-InSubject", weight_name="InSubject-0.5.safetensors")
Strengths & Weaknesses
The model excels at:
- Preserving subject identity and key characteristics during edits
- Maintaining subject proportions and anatomical accuracy
- Making targeted edits without affecting the main subject
- Strong subject-aware prompt adherence
The model may struggle with:
- Complex multi-subject scenes where subject boundaries are unclear
- Very dramatic lighting changes that fundamentally alter subject appearance
- Edits that require significant subject pose or orientation changes
Training Data
The QwenEdit InSubject LoRA was trained on a curated dataset of high-quality image editing pairs that focus on subject preservation. You can find this data here.