Project page — The Missing Temporal Link

Live page: https://dagroup-pku.github.io/Temporal-Context-Routing.github.io/

Paper: arXiv:2609.02367 · Hugging Face Papers · Official Hugging Face demo

Project page for The Missing Temporal Link: Temporal Context Routing for Script-Driven Audio-Video Generation.

Running it

Open index.html in a browser. Clip metadata is loaded from assets/data.js, so the page works from the filesystem (file://) as well as over HTTP.

After replacing assets/data.json, refresh the JS payload:

python3 tools/sync_data_js.py

tools/build_data.py writes both data.json and data.js.

To serve the folder:

python3 -m http.server 8901
# then open http://localhost:8901

What is on the page

The page opens with an 81-second video overview assembled from concise excerpts of the held-out paper demos and application-oriented candidates. Four English-language section cards group the evidence into dense dialogue scheduling, off-screen and cross-cut dialogue, coordinated audio-visual moments, and temporal editing. The video then presents the method figure and key quantitative comparisons against the base model.

Seven held-out scripts, all generated by the same model at the same training step used in the paper (7k), follow the video. Each card shows the clip together with the script's own clock:

• shots lane — the requested interval of every shot, alternating blue and amber
• dialogue lane — the requested interval of every spoken line
• dashed red marks — where the cut or the line actually landed, as measured by the evaluator
• playhead — synchronised with the video, so a viewer can see a cut arrive on the written second

Hovering a block reads out what the script asked for and the measured error. Clicking a card plays or pauses it; starting one card pauses the other.

Files

index.html          page structure and copy
style.css           styling
app.js              card rendering, timeline, playback sync
assets/data.json    per-clip script intervals, measured intervals, metrics
assets/data.js      same payload as a JS global, used by the page at runtime
assets/videos/      generated clips plus the 81-second video overview
assets/posters/     first-frame thumbnails

Regenerating the data

assets/data.json is derived from the evaluation run, joining the input scripts with the per-clip measurements:

• scripts: bench_additive_single_step7000_latest/additive_single/eval_input_all.json
• measurements: bench_additive_single_step7000_latest/additive_single/eval_results.json

The seven paper-demo clips were selected by a fixed rule rather than by eye: at least three shots, shot count matched, boundary MAE ≤ 0.08 s, every line detected with onset error ≤ 0.15 s. Candidates were then checked frame by frame to exclude clips carrying burned-in subtitles, and the final eight were chosen to cover distinct settings.

Publishing status

• The arXiv, Hugging Face Papers, code, weights, and official Hugging Face demo links are live.
• The video overview is 1280×720 H.264/AAC with faststart and is about 14 MB.