MCG-NKU/E2FGVI

CUDA OOM for ~1-2 min long videos

smandava98 opened this issue · 3 comments

smandava98 commented

Hi,

My videos are ~ 1 min long (30 FPS) which leads to between 1k-2k frames... I am on an A100 GPU but I keep running out of memory. I don't want to risk losing precision because small details are important. Do you recommend creating any changes that can help solve this?

smandava98 commented

I tried this on an 80 GB A100 as well as a 40 GB one and it runs out of memory for both (although it is allocating different sizes for both too).

Paper99 commented

Thank you for your interest in our work. Processing 1000 frames at once is far beyond the GPU memory requirements. The best approach is to divide the 1000-frame video into smaller segments, test them separately, and then combine them together.

  • 👍1
smandava98 commented

Thanks. I've come up with a workaround. However, compared to FGT and other methods, the result seems to be worse. I'm using the HQ model.

It seems like it misses inpainting of the edges of my mask. Are there any requirements that the model must take in terms of masks?

I'm using in the wild videos.