Flux.1-Dev Scaled fp8版本v2 (ID: 1158144)

I haven't seen this uploaded on here.

This is the scaled fp8 FLUX.1 [dev] model uploaded to HuggingFace by comfyanonymous. It should give better results than the regular fp8 model, much closer to fp16, but runs much faster than Q quants. Works with the TorchCompileModel node.

The fp8 scaled checkpoint is a slightly experimental one that is specifically tuned to try to get the highest quality while using the fp8 matrix multiplication on the 40 series/ada/h100/etc... so it will very likely be lower quality than the Q8_0 but it will inference faster if your hardware supports fp8 ops.

From HuggingFace :

Test scaled fp8 flux dev model, use with the newest version of ComfyUI with weight_dtype set to default. Put it in your ComfyUI/models/diffusion_models/folder and load it with the "Load Diffusion Model" node.

描述:

Original v1 model was broken. Test v2 has been updated and fixed.

训练词语: