Technique / MotifMontage

Cut among variants without losing the motif

Compare a family of outputs rather than demonstrate a control sequence.

Run the original

6 s · 30 fps · silent

Original neutral assets · 6 seconds · phase values are chosen defaults, not source timings.

npm run render -- MotifMontage out/MotifMontage.mp4 --codec=h264 --muted

Download the complete pinned project · Inputs, phase meanings & full source code

Official source / inspected interval

ChatGPT Images — Chameleon

Watch the official OpenAI upload ↗ · Published 2026-04-21T12:23:50-07:00

Inspected 46–50 s. Consecutive subset: 49–50 s, 30 frames. Dates come from the official watch-page publication field.

Last western-portrait frame before new variant at source 49.134000 seconds — 49.134000 s · Last western-portrait frame before new variant

New character treatment at source 49.167000 seconds — 49.167000 s · New character treatment

Next hard-cut variant after three frames at source 49.267000 seconds — 49.267000 s · Next hard-cut variant after three frames

Another treatment; same recurring subject at source 49.367000 seconds — 49.367000 s · Another treatment; same recurring subject

Source: OpenAI. Bounded excerpt and annotated stills for the analysis below; not included in the project download.

What changes across the edit

Shot boundary: After the longer western vignette, native frames show rapid portrait replacements. Examples begin at 49.167, 49.267, 49.367 and 49.467 s.
Framing: Portrait images sit within a white field; size, crop and angle vary. Recurring character, not exact registration, provides continuity.
Layer ownership: One output image at a time on a stable field; no native application UI.
Motion: Motion inside a vignette is followed by hard cuts among distinct character treatments.
Typography: No reading text in this selected montage interval; later word-based storytelling is a different study.
Action order: Longer subject action → rapid variants → continued sequence outside this window.
Pacing & readable holds: Several portraits last about three 30 fps source frames, roughly 0.1 s. This is a fast visual rhythm, not adequate time to read dense output.

Recreation versus observation

Matches

One recognizable silhouette survives direct variant replacements.

Differs

Original geometric artwork at a fixed footprint, three variants, 0.8 s default cadence and a longer last hold. No generated character or source image is reused.

Measured

Native changes at frames 1475, 1478, 1481 and 1484 (49.167–49.467 s).

Inferred

The recurring subject unifies the montage; no consistency benchmark is inferred.

Chosen defaults

All Remotion phase frames, cubic Bézier (.16,1,.3,1), geometry, colors, font choice and reading holds are chosen defaults. None is a recovered OpenAI specification.

{
  "title": "Make room for the idea.",
  "prompt": "Arrange a small exhibition.",
  "result": "Three studies, ready to compare.",
  "label": "EXHIBITION / 03",
  "accent": "#40685b",
  "action": 24,
  "change": 48,
  "settle": 120,
  "hold": 60,
  "focusX": 0.66,
  "focusY": 0.64,
  "zoom": 1.55
}

Increase hold to extend only the final quiet state. Change the phase frames to change choreography; preserve their ordering. Content limits and the exact meaning of each phase are in the handoff.

Exact original geometry, layer ownership and asset contracts · Schedule against an actual recording