fp8 quantized Z-Image for ComfyUI using its quantization feature "TensorCoreFP8Layout".
Scaled fp8 weights. higher precision than pure fp8.
Also with "mixed precision". Important layers remain in bf16.
There is no "official" fp8 version for z-image from ComfyUI, so I made my own.
All credit belongs to the original model author. License is the same as the original model.
FYI: many people might think that fp8 model has huge quality loss. That's because "fp8 model" saved by ComfyUI is ... just a model with fp8 weights. And many creators made their fp8 model in that way.
Normally when people talk about "fp8 model", they mean "quantized fp8 model", like scaled fp8 and gguf q8. The weights are "compressed".
If you see creators complaining about the poor quality of fp8 models saved by ComfyUI, send them this link, or make your own quantized fp8 model from bf16.
https://github.com/silveroxides/ComfyUI-QuantOps
I just share the tool, I'm not using it. I'm using my own old script.
Base
Quantized Z-Image. Aka. the "base" version of z-image.
https://huggingface.co/Tongyi-MAI/Z-Image
Note: No hardware fp8, all calculations are still using bf16. This is intentional.
Rev 1.1: An updated version with better "mixed precision". More bf16 layers, so the file is bigger. Previous version will be deleted.
Turbo
Quantized Z-Image-Turbo
https://huggingface.co/Tongyi-MAI/Z-Image-Turbo
Rev1.1: An updated version with better "mixed precision". More bf16 layers, so the file is bigger. No hardware fp8. Previous version will be deleted.
v1: It contains calibrated metadata for hardware fp8 linear. If you GPU supports it, ComfyUI will use hardware fp8 automatically, which should be a little bit faster. More about hardware fp8 and hardware requirement, see ComfyUI TensorCoreFP8Layout.
Qwen3 4b
Quantized Qwen3 4b. Scaled fp8 + mixed precision. Early (embed_tokens, layers.[0-1]) and final (layers.[34-35]) layers are still in BF16.
Description
FAQ
Comments (6)
I've been using your first fp8 turbo model for quite some time and when I compared speed with silveroxide's full fp8 quant your model was even faster than his.
The warnings in the model description and comments got me thinking.
Conventional FP8 (which I also using) simply performs type conversion without proper scaling. It's strange to call it "pure FP8." Isn't it important to spread the word that "the FP8 you're using is practically fake"? Creating a TE-compatible FP8 on Windows is difficult, so I gave up, but the problem may be that even a simplified 8-bit version can produce decent images. However, the benefits are enormous, and I think it's a great achievement. The current FP8 is still necessary.
Also, while many people are hoping for faster speeds from FP8, it should be widely known that RTX's AI performance is limited, so FP8 cannot achieve the same high performance as B200.
Really high quality and same speed as some of the lower GGUF models. Works OK on my 1080ti card.
Not for amd ROCm for sure :( . Constant dtype casting (FP8 BF16) every layer transition. GPU scheduler gets confused: one moment it's blasting full BF16 compute units, next it's doing slower FP8 passes → power draw oscillates wildly (e.g., 200–300 W spikes → drops to 100–150 W → repeat).
which model rocm friendly did you find that works good? I was using a bf16 + sdxl and my gpu was not enough, I have a 7900xt, Im running on linux currently.
@Mith_Arath https://civitai.com/models/2237711/z-image-turbo-nsfw-photorealistic-zit bf16 versions and run them in node as weight_dtype fp8e3fn you should be fine