Summary
Veteran AI presents a detailed ComfyUI workflow tutorial that extends MiniMax H3’s Ref2V (reference-to-video) face replacement capability with a full masking and subject-tracking system. The core problem being solved is precision: without a mask, replacing a person in a multi-person scene inevitably affects unintended subjects. The solution is the MaskVidExperiments ComfyUI extension, which uses SAM 3.1 (Segment Anything Model) to track a specific subject through every frame, crops the editing region to that subject alone, runs MiniMax H3 inference on just the cropped area, then composites the result back into the original video.
The tutorial walks through the full node graph step by step — including MVEx Mask Cleanup (which removes flickering segmentation artifacts), MVEx Subject Crop (which records the bounding box for later recomposition), latent-space mask conversion, and MVEx Subject Uncrop at the end of the pipeline. Three scenarios are demonstrated: replacing a face in a continuous outfit-change video while preserving all clothing transitions, isolating and replacing one dancer in a two-person scene while leaving the second person untouched (using SAM object IDs to select the target), and tracking a single subject through a crowded plaza full of moving pedestrians.
The tutorial also covers failure modes — cases where tracking confidence drops — and repair techniques using local cropping and upscaling. All workflows are available on RunningHub for cloud-based ComfyUI access, making them accessible to creators without local GPU hardware. The 10Eros Max model with built-in Turbo acceleration is used as the base generation model throughout.
📺 Source: Veteran AI · Published August 31, 2026
🏷️ Format: Tutorial Demo
