Research Notes

Evaluated SAM2's handling of fast-moving subjects with heavy motion blur. Found that temporal consistency breaks down above 180-degree shutter angles. Investigating frame interpolation preprocessing as a workaround.

Built automated MaterialX export from SDXL-generated texture sets. Successfully validated PBR compliance in Arnold and Karma. Next: testing displacement map fidelity at production resolutions (8K+).

Designed Neo4j schema linking shots, assets, annotations, and model checkpoints. Can now query "which shots were affected by model version X" and "show me all annotations that influenced this training run." Promising for pipeline observability.

Testing ControlNet with depth conditioning for rapid environment generation. Can produce coherent scene layouts from director sketches in under 30 seconds. Quality sufficient for previz but requires manual cleanup for finals.

Extracted 47 features from ShotGrid data (FX elements, character count, camera complexity, etc.). XGBoost model achieving 82% accuracy in predicting artist hours within ±20% margin. Ready for pilot deployment.

Trained NeRF models on set reference photography to provide interactive relighting previews for DOP review. Reduced feedback cycle from 24 hours (traditional render) to real-time. Artists love it for client approvals.

Compared 3D Gaussian Splatting and NeRF for asset capture workflows. Gaussian Splatting trains 10x faster and exports cleaner point clouds. NeRF still wins on view synthesis quality. Hybrid approach looking promising.

Interviewed 12 artists about acceptable latency for AI-assisted tools. Consensus: under 2 seconds for "feels instant", under 10 seconds for "feels responsive", above 30 seconds loses context. Design accordingly.

Designed annotation capture system for artist corrections on ML-generated masks. Data goes into retraining pipeline automatically. Early results: 15% quality improvement after one production cycle of feedback.

Standardized USD schema for AI-generated geometry and materials. Enables round-tripping between generative tools (Stable Diffusion, NeRF) and production DCCs (Maya, Houdini, Katana). Schema now in internal production use.