Stable diffusion thread

[doogyhatts]

4-5 mins per image for those using 8gb vram GPU?

 

[eno_lc]

4-5 mins per image for those using 8gb vram GPU?

That’s quite good considering it doesn’t oom

 

[doogyhatts]

That’s quite good considering it doesn’t oom

Once MagCache is ready for Flux2, it will help reduce the inference time.

SVDQuant might take a while, as the model needs to be quantized to FP4.
A smaller model also reduces loading time and inference time.

 

[jadedRomeo]

I personally gave up trying to run any models locally. It has gotten like bitcoin mining last time. Too slow, even though it's good to bypass censors.

 

[doogyhatts]

 

[doogyhatts]

Z-Image's architecture for their diffusion model is a slightly modified and larger version of the Lumina 2 architecture trained from scratch.
https://comfyanonymous.github.io/ComfyUI_examples/z_image/
https://huggingface.co/unsloth/Qwen3-4B-GGUF/tree/main

17.7 seconds for the second generation ans 22 seconds for the 1st generation. 3080TI here.

 

[eno_lc]

Z-Image's architecture for their diffusion model is a slightly modified and larger version of the Lumina 2 architecture trained from scratch.
https://comfyanonymous.github.io/ComfyUI_examples/z_image/

Can’t wait for z-image edit. Hope the character consistency is good

 

[doogyhatts]

@psyger-zero Z-Image Turbo has copyrighted anime data.

 

[doogyhatts]

 

[doogyhatts]

Yay! I got Z-image Q8_0 gguf to work Qwen3-4B-UD-Q8_K_XL gguf.

 

[doogyhatts]

So now people are using Z-Image as a refiner to Flux1.

 

[doogyhatts]

 

[doogyhatts]

LoRAs are easy to train for Z-Image Turbo.

 

[doogyhatts]

 

[doogyhatts]

@BlackWing1977
Z-image works on AMD.

 

[focus1974]

So now people are using Z-Image as a refiner to Flux1.

actually.. both looks alright to me.

 

[focus1974]

wa.. this one is good! got feel!

Ok.. going to try this next week!

 

[BlackWing1977]

@BlackWing1977
Z-image works on AMD.

What is that?? Online or can download to install?

 

[doogyhatts]

What is that?? Online or can download to install?

A new image model by Alibaba.
https://tongyi-mai.github.io/Z-Image-blog/
https://github.com/Tongyi-MAI/Z-Image

Need to download model and use it inside either ComfyUI or Wan2GP.
Wan2GP will auto-download the model.

 

[BlackWing1977]

A new image model by Alibaba.
https://tongyi-mai.github.io/Z-Image-blog/
https://github.com/Tongyi-MAI/Z-Image

Need to download model and use it inside either ComfyUI or Wan2GP.
Wan2GP will auto-download the model.

Hmm... I think still like quite complex to setup.... that is why I until now still yet to go try ComfyUI... and really dun have the interest to keep trying all these also..

But I do think eventually will be so common until a more refined solution will be available also.