black images with Z-Image-Turbo-fp16.safetensors & Z-Image-Turbo.safetensors in ComfyUI

#2
by hydraofm0 - opened

All image generated have black images on outputs

https://pastebin.com/1HArHW2D

the fp8 scaled also generates black images

hydraofm0 changed discussion title from black images with fp16 in ComfyUI to black images with Z-Image-Turbo-fp16.safetensors & Z-Image-Turbo.safetensors in ComfyUI

The same, all black pictures.

The bf16 model works fine from:
https://comfyanonymous.github.io/ComfyUI_examples/z_image/#basic-example-workflow

Comparing that to the fp16 model silveroxides created it appears that the model layers are named differently compared to the comfy models.

These layers:
all_x_embedder.2-1
all_final_layer.2-1

Should be named:
x_embedder.2-1
final_layer.2-1

-- edit --
No it's more than that. Comfy is also combining the qkv weights into a single larger array. There's probably more stuff that I'm not aware of, but for sure there is a format issue happening.

The bf16 model works fine from:
https://comfyanonymous.github.io/ComfyUI_examples/z_image/#basic-example-workflow

Comparing that to the fp16 model silveroxides created it appears that the model layers are named differently compared to the comfy models.

These layers:
all_x_embedder.2-1
all_final_layer.2-1

Should be named:
x_embedder.2-1
final_layer.2-1

-- edit --
No it's more than that. Comfy is also combining the qkv weights into a single larger array. There's probably more stuff that I'm not aware of, but for sure there is a format issue happening.

ohh, but fp16 works a little more fast on my 5090 in comparison with bf16, need solution :/

Sign up or log in to comment