Summary
This tutorial from the Veteran AI channel walks through a three-stage pipeline for extracting 3D character motion from existing footage and re-shooting it with a completely redesigned camera path — without re-animating the character from scratch.
The workflow begins with SAM 3D Body, a ComfyUI-native tool that recovers frame-by-frame skeletal motion from ordinary video using RT-DETR for person detection and MoGe for camera field-of-view estimation. That motion data is exported as an FBX file and imported into Blender, where the creator designs a custom camera path using a spline and Follow Path constraint — moves like passing through the space between three seated characters that would be impossible to capture on set. The untextured Blender render then serves as a motion-and-camera reference fed into MiniMax H3’s reference-generation workflow alongside a character image and detailed prompt, producing a fully rendered video that preserves the new camera animation.
The tutorial covers both a simplified single-person workflow and a more complex multi-character version using SAM 3.1 for mask tracking, and explains a key prompt-engineering insight: explicitly stating in the prompt that motion comes from the camera rather than the character is essential for the AI to correctly interpret the reference. Workflows are also available on RunningHub for those without a local ComfyUI setup.
📺 Source: Veteran AI · Published September 02, 2026
🏷️ Format: Tutorial Demo
