2024 Production Tool Prototype Complete

SAM2 Roto Assistant

AI-powered segmentation for VFX rotoscoping, reducing manual work by 78%. Interactive prompting with temporal tracking across frame sequences.

Project Screenshot

Challenge

Rotoscoping is one of the most time-intensive prep tasks in VFX—often requiring 2-4 days per shot for complex subjects. While AI segmentation models exist, they lack the temporal coherence and artist control needed for production-quality mattes.

After managing 5,000+ shots at Netflix, I saw this pattern repeatedly: talented artists spending days on tedious, frame-by-frame masking when they could be solving creative challenges instead.

Approach

Outcome

Achieved 78% reduction in manual roto time for simple-to-medium complexity shots. Artist feedback indicated high satisfaction with "rough pass" quality, requiring only edge refinement rather than full frame-by-frame work.

The system successfully identified its own failure modes: extreme motion blur, reflective surfaces, and rapid lighting changes. This self-awareness allowed artists to selectively apply the tool where it added value.

78% Time Reduction
92% Frame Accuracy
4.2/5 Artist Rating

Key Learnings

Integration matters more than accuracy. Even with 92% frame accuracy, the tool only succeeded because it fit into existing workflows. Artists could export results directly to Nuke, refine edges using familiar tools, and version control through ShotGrid.

Confidence scoring builds trust. By flagging low-confidence frames, the system let artists focus QC efforts where needed rather than reviewing every frame. This transparency was critical for adoption.

Failure modes inform product strategy. Rather than trying to solve every edge case, we documented where the tool worked well and steered artists toward those scenarios. Honesty about limitations built credibility.

SAM2 PyTorch FastAPI React OpenCV NVIDIA Triton