implement modnet + matting on stage 1 training

[johndpope]

known issue

1x single video
https://wandb.ai/snoozie/megaportraits/runs/khl4e01g?nw=nwusersnoozie

4x videos
https://wandb.ai/snoozie/megaportraits/runs/zhos6yxx?nw=nwusersnoozie

looking at whether or not to follow megaportraits paper - to remove background - or provide model some extra weights for face masks like emo.....

or find code that's causing this - usually it's in a batch operation that corrupts this

https://github.com/johndpope/VASA-1-hack/blob/main/data/driving_video.json

UPDATE

so this paper specifically addresses id leakage - https://arxiv.org/abs/2404.19110

I look at the EMOPortraits code - its so full of args - I attempt to untangle it

https://github.com/johndpope/EMOPortraits/tree/feat/take2

reading this - it seems like they may have upgraded expression embedding with emo
https://byteaigc.github.io/X-Portrait2/

We ensure strong disentanglement of appearance and motion in training our expression encoder such that the encoder only pay attention to the expression-related information in the driving video. Our model can achieve strong cross-style and cross-domain expression transfer, covering both realistic portraits and cartoon images. This makes our model highly adaptive to a wide range of use cases including real-world storytelling, character animation, virtual agents and visual effects.

[johndpope]

i think i missed the modnet + matting which is causing this >1 (4x videos) generation bug.
its not the id leakage.
https://wandb.ai/snoozie/megaportraits/runs/zhos6yxx?nw=nwusersnoozie
johndpope/MegaPortrait-hack#59