howto get detailed output
Whatever I seem to try, how many steps, whatever quant (well, except fp32) I seem to get blurry results. Or a sort of dot pattern. But it seems to remove a lot of details and create artifacts in generated parts. A prompt like 'remove people from the background' is still way clearer and better handled by Flux1 Kontext.
What am I doing wrong? Is there a trick like a special scheduler or special resolution?
Flux.1 Kontext is inferior compared to Qwen Edit. What did you do? What's the text encoder, you might use a non-working one
Just the Comfyui example workflow. So qwen2. 5 as text encoders if I recall correctly.
The model 'listens' very well. Prompt adherence is extremely good. But there is a grain like dot pattern in everything , and areas 'made up' by the model seem to be lacking in quality or detail (for example, if the prompt says 'remove all people' it needs to fill in the area where the people were standing. This inpainting quality is bad and lacks texture . Flux still does this better).
It seems very depended on resolution or aspect ratio. Even when sticking around 1 megapixel sometimes the results are quite ok , sometimes just unusable .
This was a known issue with the first qwen-image-edit and the 2509 version. But they explicitly mention this to be improved for the 2511 version, but I'm not seeing it.