Woman in yoga outfit

Generated by using Z-Image Turbo to generate a mildly consistent front and back view of the woman and feeding this to Meta's SAM 3D Body to generate a mesh and MHR70 skeleton. A helical camera path is generated around the mesh and a combination of depth and OpenPose skeleton is rendered for all 81 views and camera intrinsics and extrensics are saved. This, along with the front and back views as the reference input, is fed to Wan 2.2 VACE. The depth map and skeleton keep the views consistent enough to be able to re-use the camera parameters from the depth map renderings. A point cloud is made by sampling from the original mesh. For higher quality the outputs from the WAN video model were passed through Z-Image Turbo again with a low denoising (0.15) setting. Lastly the images were upscaled using SeedVR2 and splats were trained using Brush.