Open shivshankar11 opened 4 months ago
Yea saw that, I just don't see any gain here. They said there is a warm-up stage at the beginning so the first iteration takes longer and with LCM Lora you are generating quality pictures in 1 -2 steps. Does anyone have any convincing arguments?
If adding support for stable fast is not impacting anything in negative way then should include it, 1.8x speed is lot to be left on table if someone is video faceswaping. I think stablefast also has memory optimization, 1-1.5gb vram. Btw can pipeline support unet and vae as input instead of checkpoint?
stable-fast is able to optimize InstantID pipeline and achieves a 1.8x speedup! https://github.com/chengzeyi/stable-fast/blob/main/examples/optimize_instant_id_pipeline.py