I finally got to play around with the "Apply random color transfer to src faceset" feature while training.
Original -> No color transfer (initially trained with style power 10) -> applied RCT for about 30k iterations before convert.
Animated version
Okay so without the RCT while training, the conversion face looks kind of darker, almost like there's a slightly grey area over the mask. Note that this was trained for about 80k with style power 10 already, and this is what I usually do to fix color issues. The data_src faceset contained all sorts of lighting from normal, dark, colored, bright. With RCT training, it fixes the skin tone differences and makes it more evened out. But the biggest difference is the lip color.
So I can assume if you want to keep specific colors in your data_dst (like eye color, makeup, etc) after conversion, DO NOT use RCT while training.
This is also useful when re-using models to train another face. Usually when switching facesets and re-using a model, there will be some skin tone differences. Before this feature was available I would have to train with high style power to fix skin tones, before training with low style power to keep the conversion looking like the celebrity. But now I can just train with low style power + RCT while training.
**Note RCT on conversion was used in all the examples above.
Original -> No color transfer (initially trained with style power 10) -> applied RCT for about 30k iterations before convert.
Animated version
Okay so without the RCT while training, the conversion face looks kind of darker, almost like there's a slightly grey area over the mask. Note that this was trained for about 80k with style power 10 already, and this is what I usually do to fix color issues. The data_src faceset contained all sorts of lighting from normal, dark, colored, bright. With RCT training, it fixes the skin tone differences and makes it more evened out. But the biggest difference is the lip color.
So I can assume if you want to keep specific colors in your data_dst (like eye color, makeup, etc) after conversion, DO NOT use RCT while training.
This is also useful when re-using models to train another face. Usually when switching facesets and re-using a model, there will be some skin tone differences. Before this feature was available I would have to train with high style power to fix skin tones, before training with low style power to keep the conversion looking like the celebrity. But now I can just train with low style power + RCT while training.
**Note RCT on conversion was used in all the examples above.