Why is generated image quality so poor?

#1
by FireCat1 - opened

The paper at https://github.com/IDKiro/sdxs?tab=readme-ov-file shows examples that seem remarkably detailed. The output from this space is very poor quality compared to, say, Stable Cascade.

Also, there's not much variation among runs of the same prompt.

The prospect of single pass image generation is a big deal - if it works.

(Note: this is not intended as a criticism of @ameerazam08 who created the space. It's more of an expression of disappointment with the authors of the study for hyping their results without releasing a model that reproduces the claims.)

this demo has same pipeline as they mention in ReadMe @FireCat1 I already ping them regarding this issue

Thanks for your attention , if you think anything missing in this pipeline then make pr in same repo will merge it @FireCat1

also they will release the sdxs 1024 version

Check this one for more details
https://huggingface.co/IDKiro/sdxs-512-0.9

@ameerazam08 just want to re-emphasize that my comment was aimed at the study authors and not at you. I appreciate the work you put into creating this and I don't think there's anything wrong with your implementation. I just think the model they released falls far short of the results reported in their paper.

Same I'm also thinking

ameerazam08 changed discussion status to closed

Sign up or log in to comment