search

TemryL / ComfyUI-IDM-VTON

ComfyUI adaptation of IDM-VTON for virtual try-on.
GNU General Public License v3.0
370 stars 40 forks source link

Veeeery slow #22

Open vkleinmp opened 4 months ago

vkleinmp commented 4 months ago

Rtx 4080 16 GB Vram on Win 11 here. Your demo workflow is eating all my 16 Gb Vram and is at about 30 s/it very slow. Saying you need min 16 Gb Vram is very optimistic. Yes it's doing something but at that speed it's unusable. Please for the love of god, what can be done?

whitepapercg commented 4 months ago

Probably you can load unet in fp8, it degrade the quality, but not that much. Check "SUPIR Model Loader" node for code example.

cardosofelipe commented 3 months ago

Unfortunately, this model is heavy by nature, you are lucky to even be able to launch it with only 16GB of VRAM. The solution here is to get a GPU with 24GB of VRAM. And don't fool yourself thinking that it will execute in 2 seconds. With an RTX 3090 24GB it takes about 13s. Maybe you can get it in 9-10s with an RTX 4090. At least with the current version at bf16 precision mode.