Does adding Black and White footage to training data help or hinder?

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.

Locked
User avatar
Replicon
Posts: 50
Joined: Mon Mar 22, 2021 4:24 pm
Been thanked: 2 times

Does adding Black and White footage to training data help or hinder?

Post by Replicon »

I'm putting together some source data, and one of the source vids I want to use is black and white. Everything else is in color.

Would the black and white footage still be valuable to have, due to the face features being included, or would the lack of color actually throw things off? Would it be better overall for future trained models to just leave out the B/W input data?

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: 218 times
Contact:

Re: Does adding Black and White footage to training data help or hinder?

Post by bryanlyon »

That's really to complicated to answer. Depending on your data available, it might be that you NEED the extra data to get reasonable results. It could also equally be that you would just get better results from your color data.

Unfortunately there is no one "correct" answer without having all the information.

User avatar
Replicon
Posts: 50
Joined: Mon Mar 22, 2021 4:24 pm
Been thanked: 2 times

Re: Does adding Black and White footage to training data help or hinder?

Post by Replicon »

Hehe fair enough. But theoretically, suppose I have 1000 good, diverse, high quality pictures. Then, suppose I have a separate video that adds maybe 3-500 high quality pictures that might add some facial expression diversity, but are entirely black and white. Would you keep those in, or would you forgo them entirely?

Also, separate but related question: how does the converse of that work? Suppose I have a well-trained model (using color photos only), and I want to convert onto a video that is black and white. Would that work at all? This one, I can probably determine experimentally with minimal effort. :)

User avatar
torzdf
Posts: 2649
Joined: Fri Jul 12, 2019 12:53 am
Answers: 159
Has thanked: 128 times
Been thanked: 622 times

Re: Does adding Black and White footage to training data help or hinder?

Post by torzdf »

I'd put the B+W ones in, but would probably ensure they didn't take more than 15-20% of my total dataset.

This is pure "finger in the air" stuff though.

My word is final

Locked