search

AIGODLIKE / ComfyUI-ToonCrafter

This project is used to enable ToonCrafter to be used in ComfyUI.
Apache License 2.0
294 stars 9 forks source link

added half precision weights link to the model prepare section #7

Closed FizzleDorf closed 2 months ago

FizzleDorf commented 2 months ago

as the title says. the fp16 weights require much less VRAM and won't run out of memory like the full weights do on a 4090.

FizzleDorf commented 2 months ago

the model isn't being applied yet since it's always looking for the full weights. I'll add a model select instead so it should be a lot better. I might be training a 1024 model as well so the ability to select models would be ideal to implement early

KarryCharon commented 2 months ago

the model isn't being applied yet since it's always looking for the full weights. I'll add a model select instead so it should be a lot better. I might be training a 1024 model as well so the ability to select models would be ideal to implement early

Model selecttion is finished! We are testing fp16 version now.

FizzleDorf commented 2 months ago

It is possible to make it work on a 3080 10GB. With Xformers you will be able to fit a 512x240 generation into VRAM. Should be around two minutes to do a ~16 frame animation. also, you can run in fp8 to save ~2gb of VRAM on top of that.