search

kijai / ComfyUI-VEnhancer

53 stars 1 forks source link

Is there a plan to reduce VRAM usage? #7

Open Willian7004 opened 1 month ago

Willian7004 commented 1 month ago

I'm apperciate that you use sequence cpu offload and vae tiling to readuce VRAM usage on the CogVideoX nodes. Can these methods be used in this program to readce VRAM usage? I tried to input a video with 720x408 resolution and 8fps,the output uses the same parameters and it takes more than 11g VRAM.

motoyagugu commented 1 month ago

11g VRAM? I wonder whether 24g is usable or not. Does that work? The original Repo of VEnhancer said it needs A100 80g VRAM at least! Then consumer level devices require multiple GPUs! If the comfy node works, the author is a genius.

Willian7004 commented 1 month ago

11g VRAM? I wonder whether 24g is usable or not. Does that work? The original Repo of VEnhancer said it needs A100 80g VRAM at least! Then consumer level devices require multiple GPUs! If the comfy node works, the author is a genius.

I used the same parameters as the video I input to test if the model run properly. I tried to upscale the video to 1944x1100 16fps on a server with the fp32 model and it took about 42g VRAM on sampling and about 52g VRAM on VAE decodeing. Maybe 24g VRAM is not enough for the fp16 quantialized model without optimisation methods to reduce VRAM usage.