Hunting for view-consistent image-to-video models for game character animation
Flying_Scorpion · reddit · 2026-10-31
A game developer's 3D character animation pipeline needs one split-screen video showing the exact same motion from front and side cameras, frame-by-frame consistent.
The pipeline: UniRig auto-rigging → train a DINOv2-based joint finder on 3,000 random front/side renders → animate a split-screen reference via image-to-video → reject clips where views disagree → triangulate joints into 3D → drive the original rig, with a ground-truth "exam" to validate the joint finder.
The blocker is video generation: Google Flow and Wan 2.2 give inconsistent results; Seedance 2.5, MiniMax H3 and Stability AI's SV4D 2.0 are being tried. Asking which models/workflows best preserve cross-view motion consistency, whether split-screen generation is workable vs. dedicated multiview models, and what reference layouts or camera controls help — open-weight options preferred.
More from Embodied
- Open-source lidar_slam_ros2 adds multi-session map joining without GPS or initial guess — rsasaki0109 · 2026-10-31
- Robot demo asked to do a kickflip: "sim2real just got a lot harder" — RemiCadene · 2026-10-31
- Agile robot demo: vaulting, climbing, jumping, running on PC and phone — ChongZzZhang · 2026-10-31
- ETH RSL puts its Unitree G1 parkour policy in the browser with 50Hz MuJoCo sim — ChongZzZhang · 2026-10-31
- a16z investor: scaling robotics requires re-engineering the middle supply tier — jaimalik · 2026-10-31
- Tau humanoid's in-home cleaning demo shows smoother teleoperation — CyberRobooo · 2026-10-31