32K iterations result

Display post links to interesting swaps


Locked
User avatar
DJ_Patron
Posts: 2
Joined: Wed Dec 18, 2019 6:21 am
Has thanked: 4 times

32K iterations result

Post by DJ_Patron »

Hello, guys. Original model has been used, masker was set for Extended.
Batchsize was set for 256: I read that this parameter affects the result and than NN can clearly highlight the main signs of the faces when it's bigger value of Batchsize.
This time I choose speed and it's finished by the slightly blurry picture. This is not the problem, I keep experimenting now.
I'm using MSI RTX 2070S, training speed is reaching about 220 EGs/sec. Is it normal? Hasn't seen such a big values here.
There is 32662 iterations total results:

Face A:Image
face B:Image
Please, tell me, when face loss value is about 0,02 should be the result blurry or not? What quality you get at this value of face loss?
Maybe I should delete very similar faces from the faceset?
Somewhere there was information that with a large number of similar input frames, the neural network "averages" them, and this leads to blurring result. Is it true?

Big thanks to torzdf for his great guides, it is very useful information!

User avatar
Howden
Posts: 2
Joined: Sat Mar 21, 2020 4:04 pm
Been thanked: 1 time

Re: 32K iterations result

Post by Howden »

You kinda want to get images every second from a video. You want as much information in as little images as possible, this means, get more different videos with different lighting. Get all angles etc. Exporting every frame from a video doesn't make sense at all and slows down your progress. In that case, you need more iterations for a similar result. I would assume it gets blurry that way.

I am not an expert, just a critical thinker. Always think how you would do it if you were the computer. Thats also how you write programs and automations.

User avatar
Howden
Posts: 2
Joined: Sat Mar 21, 2020 4:04 pm
Been thanked: 1 time

Re: 32K iterations result

Post by Howden »

sort by blurriness and remove the bottom faces, then remove faces from allingment file.

Locked