Generated with a much more efficient flow that uses an experimental fork of Brush with support for normal map supervision. Normal maps are generated for every view using sapiens2 and used as a loss function in the splatting process. This vastly improves novel view synthesis. Combined with some improvements to the way the VACE model is guided this new flow produces splats with 2 invocations of the video diffusion model instead of 6.
The face is vastly improved by providing better anchoring to the source image as well as the normal loss term, but fighting the video model's priors of humans tending to look at the camera from every angle is tricky (hence the strange eyes on the splat).