AI Rotoscoping & Depth Pipeline — combining state-of-the-art models into a production VFX toolset.
Meta's SAM 3 for video object segmentation. Given a text prompt (e.g. "person"), it generates per-frame binary masks with temporal consistency. Acts as the foundation stage — its masks guide all downstream methods. Runs on-device via GPU inference.
Refines SAM3's hard binary masks into soft alpha mattes with fine hair/edge detail. Takes SAM3 output as guidance to produce broadcast-quality alpha channels — the highest-quality matte path (also the slowest; runs on the high-VRAM box).
Best for green-screen plates. Takes the refined alpha matte and extracts clean foreground, composite previews, gray cards, and pre-multiplied EXR sequences. Outputs industry-standard formats ready for Nuke/Fusion/After Effects compositing.
Alternative matting path using MatAnyone2 for subjects where VideoMaMa may struggle. Particularly strong on complex semi-transparent materials (glass, smoke, fabric).
BRIA's lightweight model for fast background removal. Runs in FP16 mode for speed. Ideal for quick previews or when a rough matte is sufficient. Uses SAM3 mask as guidance.
Generates per-frame monocular depth maps and surface normals from single camera footage. Useful for parallax effects, relighting, defocus simulation, and 3D integration.
Large model variant that produces temporally consistent depth maps across video frames. Eliminates the flickering common in per-frame depth estimation methods.
Keyer combines multiple cutting-edge models into a unified VFX pipeline that didn't exist before: