project for work

Display post links to interesting swaps


Locked
User avatar
sethstrong
Posts: 5
Joined: Thu Dec 10, 2020 8:54 pm

project for work

Post by sethstrong »

Im new to this technology and still have a lot to learn. I had a project from work that I where i was attempting to swap out guilianis head onto that of an actor we hired and shot in our studio. im posting the latest attempt here (we ended up going out of house due to time constraints) but Im continuing to practice with these elements until I get a decent deepfake.

Im extracting 512 and using Dfl-SAE trainer (192px). the result is from after over 200K iterations. as you can see its quite blurry.

Can anyone suggest what I can do to make this look better? Train longer? I think I saw a post that said something about if the graph doesnt look like its going down its not getting better? if thats the case perhaps its a problem with my B source images?

I was thinking it might be a resolution issue in the trainer, but when I go up to 256 it seems to think my GPU wont handle that (RTX2080) I just bought a 3090 soley for this software but just found out there's no support yet (this idiot needs to learn to read announcements).

Any thoughts?

https://youtu.be/g1yyd4A8WvE

Thanks in advance

User avatar
bryanlyon
Site Admin
Posts: 793
Joined: Fri Jul 12, 2019 12:49 am
Answers: 44
Location: San Francisco
Has thanked: 4 times
Been thanked: 215 times
Contact:

Re: project for work

Post by bryanlyon »

Looks to me more like your data is varied enough. I'm assuming you're probably only using a couple of videos of the actor? You really have to get a large variety. Even Giuliani is going to be hard to get the proper variety that I recommend. Just being in news content is usually pretty static lighting and not a lot of pose or expression variety.

User avatar
sethstrong
Posts: 5
Joined: Thu Dec 10, 2020 8:54 pm

Re: project for work

Post by sethstrong »

yes, that is true about the actor, I only have the one shot of him from that one particular environemt. I guess I had mistakenly thought it was only the B footage that had to be as widely varied, so in the future, when we get an actor in studio we need to get a larger variety of shot/environments to get best results then? I could maybe get a PA to shoot him off set with a cellphone camera in different lighting conditionss with multiple angles and expressions.

So you dont think training longer will help given my current somewhat limited data set?

And thanks for the quick reply. very helpfull!

User avatar
sethstrong
Posts: 5
Joined: Thu Dec 10, 2020 8:54 pm

Re: project for work

Post by sethstrong »

By the way, here is the result we got from the guy out o f house (Dr Fakenstein), obvisouly great work... do you have any idea what he uses?

https://www.dropbox.com/s/le37rvlq46en6 ... L.mov?dl=0

User avatar
sethstrong
Posts: 5
Joined: Thu Dec 10, 2020 8:54 pm

Re: project for work

Post by sethstrong »

One other question, I just saw the Keanu convenience store post, which is briliant, and saw in a message down at the end you mentioned pre-training not necessary unless you have limited source material (or something to that nature) do you think pretraining would have been helpful in my specific case?

and is there a forum thread dedicated to pretraining?

Locked