Z-Image Fine-Tune only noise in ComfyUI

[Christian-H-0212]

Hello everyone,

I have the following problem:

1. I fine-tuned Z-Image using the preset that was added yesterday.
   OneTrainer-Z-Image-config1.json

2. The test samples from OneTrainer all looked good.

3. Then I copied the finished “.\OneTrainer\models\model.safetensors” to ComfyUI in “.\ComfyUI\models\diffusion_models” and renamed it “z_image_fine-tune_bf16.safetensors.”

4. Then I selected my fine-tune in the “Load Diffusion Model” node and generated images.

Result:
All images I generate with my fine-tune only have noise, and I don't understand what I did wrong.

Can anyone give me a tip?

[dxqb]

Comfy have created their own format, that isn't compatible with anything else: Comfy-Org/ComfyUI#12303
There is a conversion script here that you could try: https://huggingface.co/Comfy-Org/z_image_turbo/blob/main/z_image_convert_original_to_comfy.py#L18

If that works let me know, I haven't tried yet.

[Christian-H-0212]

You're my hero... it worked! ❤️

Here's a quick guide for others who might have the same problem:

python -m venv .venv # Optional
.\.venv\Scripts\python.exe -m pip install --upgrade pip # Optional
.\.venv\Scripts\activate # Optional
pip install torch safetensors
pip install packaging
pip install numpy
python convert.py z_image_fine-tune_bf16_converted.safetensors z_image_fine-tune_bf16.safetensors

And don't forget to save the contents from here in a “convert.py” file.