I tried fp8-mixed, and it works well enough.
#4
by AlexTheSandGrinder - opened
Just a report, and also, if you know what speed lora specifically would work well with this? I'd rather do 4 steps than 20!
The FireRedTeam team said they would create a distill model, so just wait for it.
cocorang changed discussion status to closed
I am patiently waiting
by the way I tried Qwen-Image-Lightning-4steps-V2.0-bf16.safetensors and it works well enough. There's obviously visible degradation, but it's acceptable.