How is VRAM distributed during Extraction?

Want to know about the Faceswap's Face Extraction process? Got tips, ideas or just want to learn about how it all works? Then this is the place for you

Forum rules

Read the FAQs and search the forum before posting a new topic.

This forum is for discussing tips and understanding the process involved for Extracting and preparing face sets for training a model in Faceswap.

If you have found a bug are having issues with the Extraction process not working, then you should post in the Extract Support forum.

Please mark any answers that fixed your problems so others can find the solutions.

Post Reply
User avatar
Posts: 4
Joined: Fri May 12, 2023 3:22 am
Has thanked: 3 times

How is VRAM distributed during Extraction?

Post by mat426 »

I'm planning on swapping multiple videos that are 20-30 minutes long and 25 fps. It's always one face in the video and its mostly frontal. Whats the best approach to settings for the Extraction of the Alignmentfiles for Conversion? (RTX 4090 24gb)

Does it make sense cranking the batchsizes up and running in Singleprocess or cut the videos in smaller parts and running them one by one?
I´ve tested some different approaches already but the speed changes randomly from 12 it/s to 2 it/s and everything slows down even more when reaching the 90% completion range.

How exactly is the Vram used during this process so i can get it done the fastest way?

Thank you.

User avatar
Posts: 2373
Joined: Fri Jul 12, 2019 12:53 am
Answers: 146
Has thanked: 116 times
Been thanked: 538 times

Re: How is VRAM distributed during Extraction?

Post by torzdf »

Honestly, at extract, the biggest bottleneck is likely to be the CPU processing (aligning the face, saving to disk, etc. etc.) rather than GPU. I doubt you would get much speed up, if any, increasing the batch size.

My word is final

Post Reply