search

yisol / IDM-VTON

[ECCV2024] IDM-VTON : Improving Diffusion Models for Authentic Virtual Try-on in the Wild
https://idm-vton.github.io/
3.92k stars 613 forks source link

What is the minimum memory requirment? #43

Open nitinmukesh opened 6 months ago

nitinmukesh commented 6 months ago

I have 4060 8GB VRAM + 8GB shared RAM and it fails all the time with CUDA memory error.

Is there any setting which I can modify so it can run on lower requirements. I already tried 360 x 480 image size and same memory issue.

yisol commented 6 months ago

Hello, >18GB VRAM is required for single image inference. You can use optimization scheme like offloading for further memory reducing.

nitinmukesh commented 6 months ago

Thank you for your response. Appreciate your efforts in creating this.

Hello, >18GB VRAM is required for single image inference. You can use optimization scheme like offloading for further memory reducing.

Unfortunately being an end user with no developer background, no idea what need to be done.

theguaz commented 6 months ago

@yisol hello, I'm getting this when I tried to run it on a G4dn.2xlarge AWS instance (GPU NVIDIA T4 16 GB GPU memory).

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 14.58 GiB total capacity; 13.85 GiB already allocated; 15.56 MiB free; 14.38 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

I've tried to use max_split_size_mb with values ranging from 64 to 512 and I still get the same result. Should I get a higher end GPU or is anything I can do to optimize this repo ?

FurkanGozukara commented 6 months ago

works minimum with 8 GB VRAM on 4 bit precision with cpu offloading https://github.com/yisol/IDM-VTON/issues/47

theguaz commented 6 months ago

@FurkanGozukara but that is using your files and not the original repo

FurkanGozukara commented 6 months ago

@FurkanGozukara but that is using your files and not the original repo

true original repo uses a lot

nitinmukesh commented 6 months ago

d to use max_split_size_mb with values ranging from 64 to 512 and I still get the same result. Should I get a higher end GPU or is anything I can do to optimize this rep

Just don't spam all the github with your posts. I have already seen your post to buy from patreon and I'm not interested.

smsali commented 1 month ago

@FurkanGozukara but that is using your files and not the original repo

true original repo uses a lot

Hey! What do you mean by 'your files' and 'original repo'? I have 3060ti with 8gb memory. Want to try this. Thanks!

FurkanGozukara commented 1 month ago

@FurkanGozukara but that is using your files and not the original repo

true original repo uses a lot

Hey! What do you mean by 'your files' and 'original repo'? I have 3060ti with 8gb memory. Want to try this. Thanks!

here our idm vton app : https://youtu.be/m4pcIeAVQD0

also here how to use on cloud : https://youtu.be/LeHfgq_lAXU