README.md

---
license: apache-2.0
base_model: Qwen/Qwen-Image-Edit
tags:
  - lora
  - diffusers
  - image-to-image
  - qwen-image-edit
  - image-editing
pipeline_tag: image-to-image
---

# Qwen-Image-Edit LoRAs

A collection of LoRA adapters I trained for [Qwen-Image-Edit](https://huggingface.co/Qwen/Qwen-Image-Edit),
focused on pose control and subject consistency in image editing.

The adapters were trained on a mix of collected reference images and synthetic images generated with the
Qwen image model.

## Adapters

| File | Purpose | Rank |
|------|---------|------|
| `qwen-image-edit/multipose_v50_rank96.safetensors` | Multi-pose control (latest) | 96 |
| `qwen-image-edit/multipose_v49h_rank96.safetensors` | Multi-pose control (prior revision) | 96 |
| `qwen-image-edit/qwen_Image_male.safetensors` | Male subject conditioning | - |

## Usage

Load an adapter on top of the Qwen-Image-Edit pipeline with `diffusers`:

```python
from diffusers import QwenImageEditPipeline
import torch

pipe = QwenImageEditPipeline.from_pretrained(
    "Qwen/Qwen-Image-Edit", torch_dtype=torch.bfloat16
).to("cuda")

pipe.load_lora_weights(
    "lognat0704/diffusion-loras",
    weight_name="qwen-image-edit/multipose_v50_rank96.safetensors",
)

# image = pipe(prompt=..., image=...).images[0]
```

## Notes

- Adapters are rank-96 unless noted.
- The `multipose` series iterates on pose-conditioning quality; `v50` is the most recent.