MrDeepFakes Forums
  • New and improved dark forum theme!
  • Guests can now comment on videos on the tube.
   
Total Likes Received: 9 (0.05 per day | 0.49 percent of total 1834)
(Find All Threads Liked ForFind All Posts Liked For) Total Likes Given: 9 (0.05 per day | 0.48 percent of total 1871)
(Find All Liked ThreadsFind All Liked Posts)

godinshackles
(LightFaker)
LightFaker

Registration Date: 03-01-2019
Date of Birth: Not Specified
Local Time: 09-16-2019 at 04:52 PM
Status:

godinshackles's Most Liked Post
Post Subject Numbers of Likes
Getting a SAE model started well and other clarifications 3
Thread Subject Forum Name
Getting a SAE model started well and other clarifications Questions
Post Message
Hi, I've currently built my first SAE model and ran it up to 257k epoch ish. It has yielded some very positive results and I am quite happy with it. I'm trying SAE mostly because I read that it was the most englobing model so far. Quite satisfied with it so far. Might be able to start uploading quality material soon who knows? Still chickening out on the legal basis of deepfakes but definitely thinking about it. Won't do until I produce quality stuff though.

I have some questions before I start my next project since I intend on getting better at this. You may not have all the answers but just try/talking from experience can help. Again, I took the time to read the descriptions provided in the SAE model. Just trying to understand them in practice. If any section or thread with lots of those questions answered have already started, please refer me to them instead of answering unnecessary questions of course. I'm actively looking for tutorials and if I get better at this, I'll look into writing some myself.

My workflow has greatly been enhanced with the help of guides/comments from dpfks and iperov (like most of us I'm sure) so if you're reading this, big thanks. Getting somewhere with this because of y'all and credit's gotta be given.

Here are my questions. It's a lot but I'm fairly sure this covers all my issues. Hope this won't give the tl;dr syndrome to too many. But thanks in advance to anyone who answers any of these.  

I have a GTX 1070. I can give full specs if needed.


Question 1:
I struggle with side angles, is it because of me using the SAE model or is it because of my src facecet not having the "almost perfectly exact" angles




Question 2: Model reusing. I'm trying to understand the logic behind building a model
Can I swap one of the two actresses in the model?
Can a model be used for other actresses? If so, is it mandatory that one of the actresses is the one used in the model?
Can I reuse the same model for the same two actresses BUT a different scene and obtain satisfying results?




Questions 3: What does erode do. I'm quite confused by the picture in the guide (not sure if it enhances the color or changes the mask width)



Section 4 : Model First Run (things I can't override and need to choose day 1 so I'm trying to understand them before getting my next model started)


- Learn Mask: I think dpfks mentioned in another one of my threads that if I put this setting to y/true (which I can't override), I won't be able to fix faces with the erode option (which I'm still looking to actually understand by the way). Is it worth putting to true or not so much?

- Resolution: 128, 256. I'm "assuming" 256 would yield better results for larger faces? I'm sure it would bleed my GPU if it can even take it (which I doubt). How much vram does it take actually to run 256 and is it all it's cracked up to be?

- Half-face vs Full-Face: I'm taking this is better for scene with...um...um... mouth obstructions? But since this is a pre-setting, is it worth building a model for such specific scenes?

- Batch size: Since it can cause error, I've trained my model at 16 and got to 257k without issues. I tried 24 but it didn't seem to work. Is there a way to know the proper limit of a GPU or is it trial and error?


- Feed faces sorted by yaw: Does it actually help the model get a better grasp and not just select random pictures when training? Is it a must?

- Auto encoder dIMs: How can I find out what is best for my GPU? Trial and error?
- Encore/decoder dims per channel: Same question as above


- Pixel Loss: This is said to be worth putting after 15-25k epochs so I should keep a look out and override this setting after a while?
- Face Style Power: This is said to be worth putting after 10k epochs so I should keep a look out and override this setting after a while?

godinshackles's Forum Info
Joined:
03-01-2019
Last Visit:
Yesterday, 05:29 AM
Total Posts:
30 (0.15 posts per day | 0.29 percent of total posts) Find All Posts
Total Threads:
11 (0.06 threads per day | 0.49 percent of total threads) Find All Threads
Time Spent Online:
18 Hours, 9 Minutes, 43 Seconds
Thanks/Likes:
Given: 9 | Recieved: 9
Members Referred:
0
godinshackles's Contact Details