LibTV Adds Depth-Video Extraction: Copy Motion and Camera Work, Not Faces
卡尔的AI沃茨 · wechat · 2026-09-01
When using reference videos for AI video generation, the original faces, clothing, and background often leak into the output. This post describes a fix: convert the source video to a depth video first, then feed that to the model — motion, camera moves, positioning, and spatial relations are preserved while identity and style interference drops sharply. LibTV now has one-click depth extraction: drop the source into the canvas, convert to grayscale depth, and pipe it into downstream generation nodes.
The author tested chase/fight scenes, classic orbiting camera work, and Monument Valley–style spatial transformations, with strong motion replication and face stability. Caveats: depth reference works best on live-action with clear spatial layers and large motion; flat anime or static shots benefit little. The post also covers LibTV's new talking-head smart editing: auto-removing filler, rough-cut confirmation, auto review, and natural-language tweaks, all in one canvas.
More from coding & agent
- Pi Durable Agent harness on Cloudflare powers fully autonomous apps with zero human input — MicahBerkley · 2026-10-18
- Dev: the right prompts can make your codebase 2x faster every 10 minutes for an hour straight — cephaloform · 2026-10-18
- LLM one-shots a Minecraft 3x3 piston door by building its own Linux test infra — Vjeux · 2026-10-18
- Dev builds Mumbai-inspired 3D manga background generator from OpenStreetMap data — CurieuxExplorer · 2026-10-18
- Solo founder builds a money-printing data service on Claude's Pi Durable harness — MicahBerkley · 2026-10-18
- Dev builds Arch-based OS with custom desktop and browser using Claude Code — Snoo-10493 · 2026-10-18