Z-Image Models Have Landed
loading
Comments
Very nice! However it seems like ZIT inference is slightly broken. Image 1, made on gen-image ibb.co/5hz0pGYz Image 2, made using official hf space (huggingface.co/spaces/Tongyi-MAI/Z-Image-Turbo) ibb.co/xKdR6Hpw Image 3, made on gen-image, using llm-enchanced prompt ibb.co/dwmkCP3b Both of those have the same settings (same resolution, 8 steps, on hf space I also set Time Shift to 3 which was default value) and the same* prompt, but as you can clearly see there's something off about image made here on this website, as if it was undercooked or something lol. You can see it very clearly when comparing faces. *Hf space uses some llm prompt enchancing, so I tried to do the same, used the same model (I think) and pasted the result to gen-image.com that's how I got image 3, which is I guess is kinda better? But generation still looks off imo. Wishing you all the best, and I hope you can figure this out (unless its just issue with quantized model in which case GG lol)
Hello Yapfra, thanks a lot. I'm working on a fix to pass the time shift in parameter depending of the model use. Keep you inform.
Now that I had the time to test these models, or better, I am trying to test these, i have to say "Generation is slower than SDXL" is the understatement of the year. I am sitting here with a queue that has jobs for image generation well over 2700 seconds now.
Hey, thanks for the feedback and sorry about the wait! We're aware of performance issues on the Z-Image instances right now. We're actively working on a permanent fix this week. We'll keep you posted on progress. Appreciate your patience! 🙏
