[Tip] Why you need more sources (with examples)

Want to understand the training process better? Got tips for which model to use and when? 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 with Training a Faceswap model.

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

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

Post Reply
User avatar
abigflea
Posts: 176
Joined: Sat Feb 22, 2020 10:59 pm
Answers: 2
Has thanked: 20 times
Been thanked: 53 times

[Tip] Why you need more sources (with examples)

Post by abigflea »

To create a good model you need many sources. Photos and videos.
Typically you will hear you need a minimum of 1000, although more is better and this is to show you why.

Lets say you have a video you want to transfer your friend to.
The original actor would be A ( "A" for Actor).
You have plenty of Actor A. 30 min of video, and 500 good quality pics.

Your friend is "B" , you have very few. As can be expected, this isn't good.

In these example videos I have used increasingly less faces from A for training.
You'll see the quality decline. This will give you an Idea of why you need more data (pics & videos)for a better quality swap.

Bear in mind, I didn't run these models to completion, only enough to show you the decline in quality.
All trained 200K iterations each. Actor A had 6K faces extracted

Actor B with 5400 faces

Actor B with 600 faces

Actor B with 180 faces

Actor B with 78 faces

Actor B with 32 faces

Actor B with 8 faces

As you can see, it gets bad with less data. The inability of the model to convert expressions, head movements. Just gets ugly.

What you should take away from this is MORE DATA for reasonably good results.

:o I dunno what I'm doing :shock:
RTX 3090 : RTX 3070 : RTX 3060 : RTX: 2070 : Ghetto 1060


Post Reply