e4s2024 / E4S2024

Official Implementation of 'E4S: Fine-grained Face Swapping via Editing With Regional GAN Inversion'
https://e4s2023.github.io/
MIT License
129 stars 18 forks source link

Time-consuming Problem #7

Closed YYY7C closed 11 months ago

YYY7C commented 12 months ago

The results of this method are amazing. But it takes almost two hours to process a 3 second 1080p video on the 4090, is it normal for the video to take a long time to process?

` tuning e4s_g 79/80: 100%█████████████████████████████| 74/74 [01:08<00:00, 1.08it/s] 100%|████████████████████████████████████| 80/80 [1:29:25<00:00, 67.07s/it]

Finished fine-tuning e4s generator! `

aquawaves commented 12 months ago

I guess it just don't utilize your GPU power at all. Can you provide screenshots of your GPU and VRAM load while processing a swap?

ygtxr1997 commented 11 months ago

It could be time-consuming because we use STIT to finetune the model for all the video frames with default 80 epochs. Despite this, we suggest you to make sure your code is running on GPU by checking the GPU rate and VRAM.

YYY7C commented 11 months ago

Thanks for the answer, I'll check out the resource usage.