Corridor Cast EP#239 | Tech Talk - Coding, Building a Greenscreen Tool, Guassian Splats & More
12 snips
Jan 23, 2026 A lively tech deep dive into building an AI tool for automatic green screen keying and the data, backbones, and refiner models behind it. A tour of Gaussian splats and neural rendering alternatives, their rapid rise, and creative uses in film and VR. Chat about LLM coding assistants, installation headaches for research tools, and plans to open-source a production-ready workflow.
AI Snips
Chapters
Transcript
Episode notes
Combine High-Res Backbone With Edge Refiner
- Use a higher-resolution backbone (e.g., HIRA) for concept prediction and a secondary refiner model to recover full-resolution edges.
- Upsample concept predictions and then restore sub-pixel detail with a dedicated mini-model for hair and fine edges.
First Public Demo Was ‘Mid’ But Informative
- Niko demoed an early result that left dark fringe and motion-blurred hair issues but removed tracking markers and largely separated subject from background.
- He used that imperfect output as a milestone before rebuilding with a newer backbone and additional data.
Alpha Alone Isn’t Enough
- Generative Video Matting (GVM) yields extremely clean alpha channels but only provides mattes, not corrected foreground color.
- Combining a strong alpha predictor with color unmixing is required to avoid fringing when compositing.