Summary
Veteran AI presents a masterclass on combining NVIDIA’s open-source Kimodo motion generation model with SCAIL-2 to create custom AI character animation without any real-world video footage. Kimodo is a text-to-3D-motion diffusion model trained on motion capture data; SCAIL-2 then applies that motion to a character reference image to produce a final video output. The entire pipeline runs inside ComfyUI using the ComfyUI-Kimodo extension, with RunningHub offered as a cloud-hosted alternative.
The video is structured around honest testing rather than polished showcases — the presenter leads with failure cases to calibrate expectations before explaining what actually works. Key technical findings include: SCAIL-2 successfully recognizes motion from Kimodo’s 3D skeletal preview recordings even without real people or natural backgrounds; SAM3 generates adequate masks from these synthetic inputs; and camera angle changes in the Kimodo 3D preview are misread as character movement rather than camera shifts, limiting multi-angle workflows.
Parameter tuning is a practical highlight: the video systematically tests pose_strength values of 1.0, 0.75, and 0.5, finding 0.75 offers the best balance between motion fidelity and natural character movement for realistic reference images — both too high (stiff, overstretched limbs) and too low (motion loss) are demonstrated with visible results. The tutorial also covers Kimodo’s BVH and NPZ export options for downstream rigging workflows, and documents the screen-recording workaround required to convert Kimodo’s preview output into a video format SCAIL-2 can ingest.
📺 Source: Veteran AI · Published June 17, 2026
🏷️ Format: Tutorial Demo
