FP8 is deprecated

Use the official int8 version.

Note:

• The "txtfusion.projector" layer (aka. where people are trying to remove the filter) in this ckpt is still in original bf16 precision. You don't have to worry the precision. (In fact all layers and modules that handle text input are in bf16 precision. So, even if the target layer changed, it doesn't matter.)

Krea 2 in ComfyUI mixed precision format.

Converted from official bf16 ckpt.

Compared to official fully quantized ckpt, this ckpt keeps some important layers in original bf16 format. Thus it's a little bit larger, has a little bit better precision.

For nvfp4, ckpt size is ~9.5GiB, should be good for GPU that has 12GiB VRAM, without swapping.

For scaled fp8, ckpt size is ~14.5GiB, should be good for GPU that has 16GiB VRAM, without swapping.