*The FP8 handled by this node is NOT properly quantized and intelligent FP8. It is a practical FP8

Until now, what was known in the image generation world as "FP8" could not have been very accurate no matter how you looked at it. Recently, proper FP8 Scaled and INT8 have become popular, and it seemed that they would take over as the mainstream 8-bit quantization format. However, as far as I tested, the quality of FP8 Scaled and INT8 ConvRot was not as expected. The results show that the conventional FP8 is better overall.

At least with Z-Image, it is faster, uses less memory, and produces higher quality output than INT8, which is a hot topic these days. Depending on the type of model, there may be significant deterioration in FP8, but there is not much deterioration in Z-Image.

Therefore, this custom node was created by dedicating Shiba*2's conventional FP8 conversion node, which was a general-purpose node, to Z-Image, and adding skip processing for layers that want to maintain accuracy. Handling is no different from the conventional FP8. Model size will be slightly larger.

Install the original https://github.com/Shiba-2-shiba/ComfyUI_DiffusionModel_fp8_converter to costom_nodes and place the two files in the compressed file. The original custom node is preserved.

ComfyUI doesn't seem to support Block-wise, so I understand that there is a limit to the accuracy of INT8, but even after asking the AI various questions, I still don't understand. Although ConvRot is a smart method, it is true that it has some weak points. In fact, it sometimes loses to methods that only use type conversion, and as far as I have tried it, the probability of that happening seems to be quite high. In terms of fidelity to the output of the BF16 model, the conventional FP8 was overwhelmingly superior. The current method definitely improves the accuracy a little.