Render model comparison grid

render-model-comparison-grid

Render the 'model comparison grid' format from a config. The signature of this format is a "Same prompt. N models." gauntlet: a dark stage where, per beat, a PROMPT eyebrow + the (condensed) prompt fades in centered in monospace and holds readable ~0.8s, then docks to a small top strip while a grid of 2-4 labeled panels staggers in (0.15s apart) and holds for side-by-side comparison. A persistent model/variant label sits under each panel; column order is identical on every beat. Ends on a minimal end card (headline + column names only — no meta-stats line).

The grid is media-agnostic per cell: any cell is a static image or a muted video clip (i2v outputs, screen recordings), mixable within one beat. Video cells loop during the hold and are frame-seeked deterministically (the renderer awaits each seek), so the render never depends on wall-clock playback timing.

The renderer itself is FREE/deterministic (Playwright frame-step + FFmpeg). The paid inputs are separate capabilities: the cell images come from create-image-fal, the cell clips from create-video-fal, and the music bed from create-music-elevenlabs. Prompt text and labels are real DOM — never AI-rendered.

Default shape: 5 beats × 4.5s + 2.5s end card = 25.0s @ 1280×720/30fps, all configurable from one config.json.

Run

build_composition.py --config config.json --output hyperframe.html ; render_seekable_hyperframe.py hyperframe.html master-silent.mp4 <duration> --fps 30 --width 1280 --height 720 — dark stage, staggered grid, deterministic, $0. The config schema is documented at the top of scripts/build_composition.py; scripts/config.example.json IS the shipped worked example (re-point the cell paths at your own media).

build_composition.py validates every cell path and the column count (2-4), infers each cell's media type from its extension (.png/.jpg/.jpeg/.webp → image; .mp4/.mov/.webm/.m4v → muted video), and emits a self-contained HTML that exposes window.mediaReady() + window.renderAt(t). render_seekable_hyperframe.py awaits both, so <video> cells seek to the right frame before each screenshot — never a frozen first frame.

Contract

• Deterministic + FREE (Playwright frame-step + FFmpeg); no paid calls in this capability.
• Columns = panels-per-beat (2-4); every beat supplies exactly that many cells, same order.
• The template recipe (DB) supplies the config; cell images/clips + music are separate capabilities.
• State is computed entirely in renderAt(t) — never CSS animation-delay/transitions (Playwright scrubbing traps delayed animations in pre-state).
• Video cells must decode in the render Chromium (H.264 yes, ProRes no — transcode .mov ProRes to H.264 first). An images-only grid has no decode dependency.