**Black Screen Update 1/2/26 -
Apparently, ComfyUI might need to be told to use FP32. Here's how to do it with launch arguments.
--force-fp32 is used in ComfyUI to force all models (U-Net, VAE, and text encoder) to use 32-bit floating-point precision, which can resolve issues like black images or garbled noise in generations, particularly with certain GPUs or models like the Z-Image.
How to use --force-fp32
** Black Screen issues - Some people have issues where it generates a black image. this has to do with the original script, and I'm not sure what causes this. I have tested it on a 4090 and 5090, and even with ALL the same drivers and Pytorch/CUDA, the 5090 gives me the black screen, but not the 4090 - if it gives you a black screen, take note of the hardware you are using and possibly try a different GPU such as the 4090.
There is a clear bump in quality. If not, hang in there until we get the full weights, or someone else puts out another FP32 version. Z-Image already looks great, so this probably isn't going to be a game changer for you, but it does look a bit better.
https://github.com/PixWizardry/ComfyUI_Z-Image_FP32
Edit: also uploaded to Huggingface for transparency
https://huggingface.co/cutecaption/Z-Image_Turbo_FP32_Merged/tree/main
By utilizing this specialized merge script, I was able to merge all of the original full precision pytorch model files on the official Z-Image Turbo:
https://huggingface.co/Tongyi-MAI/Z-Image-Turbo/tree/main/transformer
By taking these 3 files and merging them together, we have merged it into a full precision (FP32) safe tensors file.
This should in theory allow a slight bump in quality and precision. The difference between BF16 and FP32 may not be very significant and this is yet to be determined by testing, but there may be some significant use for this during LoRA training, but this is also yet to be seen.
Either way, this allows us to test the model in FP32, whereas all of the previous models released appear to have been BF16 or FP8.
Please note the readme information provided by the original script creator.
**
I uploaded it to huggingface as well just for transparency, the hash thingies or whatever should match since it's literally the same file.
Description
FP32 merge of Z-Image Turbo official [3] pytorch models.
FAQ
Comments (24)
interesting, will check it out thanks
Yup! Super cool. Please see the read me from PixWizardry
ComfyUI_Z-Image_FP32
ComfyUI VERSION 3.75
THIS IS A BETA MOD, DO AT YOUR OWN RISK.
Backup your orginal "supported_models.py" at "ComfyUI\comfy" Drag and drop this modded version to support FP32 Z-Image model
THIS IS A BETA MOD, DO AT YOUR OWN RISK.
Common Error: "unet missing: ['norm_final.weight']" Image will still generate.
Goto https://huggingface.co/Tongyi-MAI/Z-Image-Turbo Merge their Transformer files into 1 file. (You need at Least 32GB of CPU Ram to merge)
Load your Safetensor and Enjoy.
@CuteCaption ah got it, will test later today and try to make a comparison. Really liking Z-image
@EricRollei21 Yup, I'm doing a comprehensive comparison now with 1024x1024, 1200x1200 and 896x1152
Any chance you could post some 1:1 comparisons? It'd be cool to see how it and Turbo handle the same prompts and settings.
Sure. Will do this later.
still working on these, I still plan on uploading them
Change comfy code base?!? The comfy HF repo includes https://huggingface.co/Comfy-Org/z_image_turbo/blob/main/z_image_convert_original_to_comfy.py all you need to do is add fp32
OK. Not my post. Whatever you're reading is on PixWizardy's page. All I've done is merge the safetensors.
getting a black screen on use. Any recommendations? thanks!
also have this issue. have tried merging it myself and thought i just did something wrong but i'm getting the same problem here.
@Soup_Fusion So, the only thing I can think of is the 2nd half of the steps, where they talk about support for it to ComfyUI - there is a specific supported_models.py file the developer is saying we need to use. I have done the exact steps, and have no issues.
https://github.com/PixWizardry/ComfyUI_Z-Image_FP32/blob/main/supported_models.py
Someone was saying "All you need to do is add fp32" and apparently that is not the case because it's not working for some.
I would encourage you to try using (after backing up your original) their supported_models.py and test it this way.
@CuteCaption i replaced the default supported_models.py file with the new one and the image is no longer black but it is incredibly noisy and doesn't resolve at all. i'm on the portable install so maybe i'm just missing a some things from the most recent builds. i'll just stick with bf16. thank you though.
hmm.. I just tested it on a 5090 and got a black screen. It looks really good on the 4090s I'm testing it on. I'm sure a more stable version will come out soon!
@Soup_Fusion yeah it's something to do with the script i'm not too sure sadly.
确实不能用，黑图。还是感谢楼主
它在4090上能用，但在5090上不行，我们不知道为什么。这似乎是随机的。
@CuteCaption 难怪我这5090不行
Not working. First it gave black pic, after patching with custom py file it gives static (snow) pic.
sorry to hear that. we don't know what causes this. something to do with the original script more than likely. i don't know for sure (ALL the drivers were the exact same), but it seems like it works on a 4090 just not a 5090, but again have no idea. All I did was use the script provided and upload it.
@CuteCaption Its alright, gonna try via Stability Matrix Comfyui today. Thanks anyways!
@vixenius it has to do with ComfyUI not being launched with fp32, apparently you have to add a flag if it doesn't work. I've also noticed, if I start fresh and then only load the FP32, it sometimes works, but if I loaded BF16 in that same workflow I get the black output.
As requested: Test on an RTX 5080 also producing black images. I hope you find out how to fix this! Best regards, Robert
Even for myself, I can't seem to figure out what causes this. Since all I have done is use a script someone else created for this purpose, I won't be of any assistance sadly. We can definitely try leaving some comments on the original script Github page to see if the original dev has any clues.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.