Krea 2, 2.42x faster than BF16 on an RTX 3090 (up to 2.90x), at a third the size. Native INT8/W4A4 quantization with an SVDQuant low-rank branch, no calibration dataset needed. Turbo and base.
https://github.com/alperktt/Krea-2-SVDQuant-ComfyUI
git clone https://github.com/alperktt/Krea-2-SVDQuant-ComfyUI
comfy node install krea-2-svdquant