Why can't the Qwen-Image Edit maintain character consistency?

### Custom Node Testing

- [ ] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

I hope qwen has the same ability to maintain character consistency as the online demo
https://huggingface.co/spaces/Qwen/Qwen-Image-Edit

### Actual Behavior


In this example, `Flux Kontext` and `Qwen Online Demo` maintain the consistency of the characters very well, and only replace the pixels where necessary.
However, the offline version of `Qwen Image Edit` (Q8_GGUF or FP8) running on ComfyUI behaves differently. It almost completely redraws the image (similar to I2I with reference input)

Why does this happen?

### Steps to Reproduce

workflow from: https://docs.comfy.org/tutorials/image/qwen/qwen-image-edit

### Debug Logs

```powershell
no logs
```

### Other

_No response_