CineReconJS

CineReconJS

Can a vision model rebuild a film clip as an editable, animated Three.js scene? Here is Claude Opus 5.5 attempting all 50 CineRecon reference videos at L1 (geometric proxies), with every reconstruction open to inspection.

Open the gallery Side-by-side player Repository Original CineRecon

How a run works

CineReconJS adapts CineRecon's Blender benchmark to Three.js. The model never sees the video file; it sees 24 timestamped frames and writes code.

  1. Sample24 frames spread uniformly over the full reference video, endpoints included.
  2. WriteThe model returns scene.mjs: geometry, materials, lights, an animated camera and object motion.
  3. RenderHeadless Chromium renders every frame at the source resolution and frame rate into an MP4.
  4. InspectThe model gets matched reference/render pairs and decides whether to revise or submit.
  5. SubmitThe model chooses when to stop. A 10-revision ceiling existed but was never reached.

L1 · Proxy

Simple geometric proxies preserving subjects, proportions, spatial relations, contacts, motion and camera timing. This run.

L2 · Simplified

Recognizable subjects, silhouettes, materials and articulated actions. Not run.

L3 · Detailed

Instance-specific geometry, appearance and imaging characteristics. Not run.

Results

Run statistics for each model on the same task: reconstruct a reference video at L1 from 24 sampled frames, refine from rendered feedback, then submit.

No quality scores yet. The original CineRecon judge has not been run on these reconstructions, so this page reports what the models did (revisions, renders, cost, stopping) and shows the videos. It does not rank them. Judge scoring would be a separate, paid step.

Revisions before submitOpus 5.5 · number of videos by changed scene revisions
Estimated API cost per videoOpus 5.5 · USD, hypothetical list price · click a bar to inspect that video

Side by side

Reference on the left, reconstruction on the right, playing in sync. Pick any of the 50 videos.

1 / 50

Read this before comparing anything

  • Not protocol-equivalent to the paper. The upstream benchmark's Blender agent sees the whole video and uses tools. These runs use sampled frames and a Three.js runtime, so short events can fall between samples. Scores or rankings from one backend do not transfer to the other.
  • Native agent runs. Opus 5.5 ran as a Claude Code subagent on a subscription, not via an API harness. Cost is a hypothetical API list-price estimate from transcript token usage, not a bill. Rendering used native Windows Chromium with software WebGL, outside the README's disposable-sandbox guidance.
  • Self-chosen stopping. Every Opus run ended by model_submit. Revision counts reflect when the model judged itself done, not a fixed budget.
  • Paused runs. 18 runs were paused by subscription session limits; their elapsed times include the pause and are not comparable. Costs here reprice the cache re-write after a pause as a cache read.
  • Astra covers one video. GPT-6 Astra ran video1 only, with a disclosed 90-minute operational ceiling and no token telemetry. It is shown for comparison on that clip, not as a cohort.
  • Reference media rights. The reference clips are third-party material. Copyright remains with the rights holders, and upstream notes that redistribution permissions are under review. Videos here are downscaled to 640 px for inspection.

Full per-run caveats are in the repository's cohort report. Method, tracking rules and validation: README · run tracking · Astra record. Dataset, annotations and evaluator come from Unc1eW4ng/CineRecon.