Skip to content

Looksmaxxing for motion graphics. An agent skill that turns AI slop videos into films that look made by a motion designer.

7.0 KB
Raw Blame

Known limits

motionmaxxing is an early release, so the glow-up is not finished. Even a maxxed film has flaws, and this page is the honest list of what it cannot do, what it cannot see, and where it is still rough. If a gate passes, that is a floor, not a verdict on taste: the scripts report numbers, a person (or the agent, looking at the render) still decides whether the film is good.

What the gates can and cannot see

Gate Limit
G1 proof readable No script. Judged by eye on look/sheet.jpg and the stills (text >= 0.04 H, fragment >= 0.55 W or cropped by the frame). The agent is asked to describe every sheet frame in one sentence first, and to say in NOTE.md that G1 was judged by eye.
G4 one hero per frame No script. Same method as G1.
G5 page chrome (lint.mjs) Reads the DOM. It cannot see text inside <canvas>, WebGL, images or video, so chrome baked into a generated plate or drawn in a canvas passes. A lint is a lead, not a verdict, and some real films will trip a rule that a designer would accept (text inside a faithful UI rebuild is exempt only when marked data-ui="captured").
G2 / G3 (look.py) Computed from frame differences on the render. A shutter-blurred or grained film changes the numbers; a held frame with a tiny living element (a caret) can read as moving.
Motion band The "human band" (median 6.8, IQR 4.4-9.8) is a question printed by look.py, never a gate. Banners, calm brand films and GIF loops sit below it on purpose.

look.py

  • It treats show and hide events from render.mjs (window.__events) as cuts. A film built from many show/hide swaps inside one continuous scene will report more "cuts" and shorter "shots" than a viewer would see (the hero banner in this repo reports 33 "cuts" from its overlay show/hide steps), which also moves the final-shot share (G3). Without an events file it falls back to detecting cuts from the picture.
  • With no cuts at all the final-shot share is not computed.

Rendering

  • WebGL is slow. A hero3d scene renders at about 7-11 fps at 1080p on an Apple GPU (bloom and depth of field cost more); software GL is slower still. --scale 0.5 does not shrink the WebGL draw. Render only the hardest beat with --from/--to, and use --still for look checks.
  • GPU versus software GL differ by a small amount (about 38 dB PSNR on a chrome scene). Do not mix modes inside one film, and do not re-render one beat on a different machine than the rest if the move is continuous across the seam.
  • --shutter multiplies render time by --subframes and is for smooth-clock films only. Stepped content stays crisp by design.
  • Smear and blur filters need room: the SVG filter region is -50%..+150% of the element's box.
  • Fonts are measured once the page and fonts have loaded. Motion.loaded() waits for that; code that measures text earlier gets fallback-font widths.
  • Chrome is required (any recent Google Chrome or Chromium). Headless rendering through the DevTools protocol; no other browser is supported.

Sound

  • ElevenLabs music tempo and fit. The music endpoint does not reliably hit a requested tempo or length: it can start late, end short, carry a silent tail, or restart a loop mid-film. voice.py music-check and music-fit repair what they can, and look.py flags a silent tail and late start, but a score still has to be listened to against the picture.
  • The sfx normaliser can corrupt very short clips. voice.py sfx-norm trims leading silence and peak-normalises to -3 dBFS; on a clip of a few milliseconds the trim and the mp3 re-encode can leave a click or an empty file. Check short hits by ear, and prefer a longer clip with a hand-placed gain over a very short one.
  • voice.py vo-local (macOS say) gives estimated word timings, not aligned ones.
  • With no key there is no voice, sfx or music. The skill says so in a line and carries on; the picture then runs on a fixed cadence.
  • Nothing here masters audio for broadcast. mix.py targets -14 LUFS / -1 dBTP stereo 48 kHz, which suits web and social.

Brand capture and logos

  • fetch_logo.mjs may return a header ribbon instead of the mark. The order is SVGL, Simple Icons, then the site's own icon or header SVG; when it falls through to the site, the "logo" it finds can be a wide header strip or a lockup rather than the symbol. Render the SVG once and look at it before building on it. Exit 1 means no SVG mark was found; the skill then sets the name in type and never redraws a logo.
  • brand.mjs cannot capture pages that block headless Chrome (an empty page is reported as blocked), pages that need a login, or fonts that are not served as web fonts. It flags colour roles heuristically: verify ink and accent on board.png by eye.

Image generation

  • imagegen.py needs the Codex CLI with the image_generation feature on. Plates take 1-2 minutes each. The guard refuses prompts that ask for text, logos, UI, charts, people or faces, but it checks the prompt, not the picture: a model can still draw lettering, so look at every plate.
  • Generated plates are for a missing surface only, are labelled illustrative in NOTE.md, and never replace product proof.

The runtime

  • Every frame must be a pure function of time. Code that uses setTimeout, Math.random, Date, CSS animations or gsap.set outside the build will not render the same when seeked; Motion.selfTest catches most of this but only on the frames it samples.
  • Two tweens must not overlap on the same property of the same element (put one on a parent).
  • M.words does not split words into letters, and fixedSlots:false supports a single line only.
  • hero3d loads an ES module, so a double-clicked file:// page will not draw 3D; use render.mjs or a local server. extrudedText reads TTF/OTF/WOFF but not WOFF2, and ignores variable axes (use runtime/hero3d/tools/fontconv.py).
  • Clock modes: on clock:'twos' quantisation happens inside __seek; seeking __timelines.main directly gives smooth poses.

The research behind it

  • Genre skew. The measured moments come from short SaaS-style launch clips, so the statistics describe 3-7 s UI-forward moments, not a whole film. "Motion on ~99% of frames" is a property of those moments and is not a target for a full film (human whole films are calmer: median luma motion 6.8).
  • Model-decoded rules. The precedent library (library/moments.jsonl) was decoded by model analysts from frame-accurate recreations. Rows are unverified; open the cited moment before trusting a number. They never outrank references/ or taste/verdicts.md.
  • One rater. The blind A/B history that shaped the taste layer has a single human rater (and a model judge in some rounds). "9 of 10 picks with the skill" is an earlier generation of this skill on five briefs and two models; it is encouraging, not a benchmark. No A/B exists yet for this generation of the skill.
  • Taste is not a template. The skill is written to find an idea per brand. It will still sometimes land on a default; the "Never ship" list and the lint exist because it has.