[Bug]: Z-Image turbo Black images

[Lordtzeentch38]

ComfyUI and Frontend Versions

I am using ComfyUI Desktop Version, currently the latest version 0.3.76.

Description

When I open the template for Z-Image and start generating, I get a completely black image.
I have checked the model, clip, and vae several times, and I have re-downloaded the model, clip, and vae, but nothing has changed; the problem persists, and the images are completely black.
I noticed the problem with my workflow, which worked perfectly in version 0.3.75, but with the current version of ConfyUI it gives black images, so I restarted ComfyUI and opened the template for Z-Image, as mentioned before, completely black image.

Steps to Reproduce

I open the template and start the generation.

Debug Logs

got prompt
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:1, offload device: cpu, dtype: torch.bfloat16
loaded partially; 9728.80 MB usable, 9501.73 MB loaded, 2237.82 MB offloaded, 225.00 MB buffer reserved, lowvram patches: 0
100%|██████████| 9/9 [00:24<00:00,  2.73s/it]
Requested to load AutoencodingEngine
Unloaded partially: 3557.98 MB freed, 5943.75 MB remains loaded, 225.00 MB buffer reserved, lowvram patches: 0
loaded completely; 248.25 MB usable, 159.87 MB loaded, full load: True
Prompt executed in 26.64 seconds

Browser Logs

It is not the Portable version. I am using the Desktop version.

Setting JSON

comfy.settings.json

What browsers do you use to access the UI ?

No response

[comfyui-wiki]

Can you share a screenshot of your settings -> about page?

[Lordtzeentch38]

[Lordtzeentch38]

However, I noticed that in the terminal logs, the models are automatically loaded into cuda:1, which is the other graphics card, identical to the one in cuda:0.
In fact, I did a test, loading the model onto cuda:0, with MultiGpu nodes, and the generated image is visible.
But if I don't use MultiGpu nodes and go back to using core nodes, the resulting image is black again, precisely because the models are loaded into cuda:1. Although, for me, this doesn't make sense, since the models are loaded into the VRAM of the other 3060 anyway.

[Blue828]

I had the same problem. When I switched from "z_image_turbo_bf16.safetensors" (12 GB) to "z-image-turbo-fp8-e4m3fn.safetensors" (6 GB) it is OK. Maybe it is related to VRAM (I have RTX5060 Ti 16GB).