Continuity Detective

About the project

Continuity Detective is a local video inspection workbench for creators assembling short videos from ordered clips. It focuses on wardrobe continuity for one subject: you import clips, confirm a reference, and review timestamped evidence of visual drift.

Inspiration

When you cut together AI-generated or phone footage of the same person, small wardrobe changes are easy to miss until you’re already deep in the edit. Existing tools either assume every difference is an error or try to infer narrative intent automatically. I wanted a workbench that shows real evidence and lets the creator decide what counts as intentional.

How I built it

I used the Devpost Learn Skill Pack from the start. The skills produced scope.md, prd.md, and spec.md in the devpost/ folder before any implementation. The result is a React + Vite frontend talking to a local FastAPI backend and a separate Python worker. The worker uses MediaPipe for torso proposals and compares colour distributions against the creator-confirmed reference. “New wardrobe from here” creates a new reference boundary so earlier frames keep the old reference and later frames use the new one. Everything runs locally—no cloud inference.

Challenges

The hard part was keeping evidence honest. Computed differences are appearance measurements, not proof of a different garment or inferred story. Getting the reference-change logic right (preserving earlier references while applying the new one only after the declared boundary) took the most care. Quality gates for occluded or unclear frames also had to stay visible so the tool never invents results.

What I learned

Planning documents forced the boundary early: one subject, wardrobe only, local analysis, creator-declared intent. The unique kernel turned out to be the reference boundary itself—not just detecting difference, but letting the creator say “this change is intentional from here.” I finished with a working first draft that shows timestamped A/B evidence and lets me export findings for editing.

What’s next for Continuity Detective

The current version is a focused proof of concept: one subject, wardrobe only, short local clips. Next steps that stay true to the same idea:

Support prop and face continuity as separate, validated checks. Allow more than two or three clips and longer sequences. Improve the torso crop proposal and add clearer “not assessable” states for heavy occlusion. Add a simple way to jump from an exported finding straight into an editing timeline. Optional cloud analysis path for people who don’t want to run the Python worker locally.

The core stays the same: creator-declared references, real evidence, and no automatic guessing of intent.