The pointmap finetune of Sapiens2 is surprisingly good for face geometry. Using that to generate a mask, fitting the sam3dbody MHR mesh to it and using that as facial conditioning is too strong of a signal for even the priors of the video diffusion model to override.
I think I'm getting pretty close to the fundamental limits of WAN 2.2. Upscaling helps but 720x1280 just isn't that many pixels, and it's hard to get good coverage with only 81 frames to work with. Working on making the flow more accessible now. I've ditched ComfyUI as the complexity of the workflows became too much for it, so it'll be a docker container with a web interface moving forward.