Quality checks

Every rule and check is detailed in docs/en/REFERENTIAL.md.

29 of the 64 checks are measured in code, without an agent (which ones apply depends on the profile):

• output length and format;
• cuts made on motion, never between two parked shots;
• frozen frames, dead air and total near-still time (under 1 s per 30 s, end card excluded);
• sounds landing on their picture, when the storyboard lists sound cues (umd sync);
• black frames;
• flashes, as a full-frame approximation of WCAG 2.3.1;
• loudness and true peak;
• silences and abrupt audio cuts;
• render reproducibility, and how faithful the delivered encode is to the master (VMAF);
• caption coverage, sync and reading speed;
• for pages: smoothness, CLS and prefers-reduced-motion;
• in the code: hard-coded durations, AI motion tells, remote fonts and assets, hollow wording;
• asset inventory and licences.

The other checks are judged by the auditors, from the contact sheet and a few targeted frames.

#Scoring

• Each check is a blocker (5 points), major (3) or minor (1). Passed = all points, partial = half, failed or not checked = none.
• A failed blocker caps its domain at 3/10 and the overall score at 5.9/10. Examples:
  • a dangerous flash;
  • text too short to read;
  • a render that is not reproducible;
  • stills not validated before animation;
  • a non-commercial asset in an ad.
• The overall score weights the domains by profile: promo, presentation, explainer, social, film, logo or web. Domains that do not apply are ignored, such as captions for a logo or audio for a web page.
• Thresholds can be tuned per project in the thresholds of .motion/project.json: −14 LUFS, 3 flashes per second, 3 s of dead air, VMAF 93… The dead-air, frozen-picture, near-still, cuts-on-motion, flash and loudness thresholds were checked with zero false alarm on 15 reference films scored 7 to 8 by an AI jury (docs/CALIBRATION.md). A rule that would have wrongly flagged 7 of these films (music dynamics) was dropped. The VMAF, caption and web ones remain conventions.

#The full referential

#Profiles and weights

#Default thresholds

#NAR · Brief and narrative

One goal, one audience, a structure that holds and a script without AI phrasing.

Build rules

• Write the goal, the audience and the expected action in BRIEF.md before any shot.
• Structure: hook, development, payoff (or call to action). In film: the character wants something small, tries, something unexpected happens, it looks at the camera.
• Every storyboard beat has a duration and a single intent, written at the end of the line.
• Key information must be understood with the sound off: it is also on screen.

#TIM · Rhythm and timing

No dead air, regular events, a duration that fits the format.

Build rules

• Something new happens about every 2 seconds; never more than 3 seconds without meaningful motion.
• Target duration set in the brief (social 15–60 s, promo 30–90 s, logo 2–8 s) and met within ±10%.
• Key shots hold long enough to be read; no shot under 0.5 s unless a deliberate rapid montage.
• With music: cut on the beats.

#LIF · Life and density

Nothing is fully frozen, scenes vary, motions overlap.

Build rules

• Nothing is ever fully still, even on the last shot: breathing, blinking, living light or particles.
• Vary framing, composition and rhythm from scene to scene; never the same layout twice.
• Characters: eyes react 6 to 10 frames before the body.
• Offset elements (stagger, overlapping action): they never all move at once.

#ART · Art direction

A look validated on stills before animating, faithful to the brief, without an AI signature.

Build rules

• Render key stills (umd still) and compare them with the reference before animating; correct for up to three rounds.
• Colours and type come from the brief and motion-tokens.json; at most two font families.
• No default purple gradient, glass everywhere, emoji as icons, generic 3D blobs or centred template layouts.
• Respect the format's safe areas (title-safe 90%, platform UI zones in 9:16).
• Logos and icons as vectors; no blurry upscaled asset.

#MOT · Easing and choreography

Consistent durations and curves, a purpose for every motion, no AI motion tells.

Build rules

• Durations and curves only from .motion/motion-tokens.json; no hard-coded values in components.
• Every motion has a role: guide the eye, explain a change, give feedback, link two ideas.
• Entrances ease-out, exits ease-in, on-screen moves ease-in-out; never linear on a spatial move.
• One focal motion at a time; the others follow, offset.
• Ban AI tells: pulsing dots, blur on every entrance, bouncy springs everywhere, everything fading up 20 px, infinite floating.
• On the web, animate only transform and opacity; never transition: all.

#TXT · On-screen text

Long enough to read, contrasted, never clipped or overlapping.

Build rules

• Text display time: at least max(1.2 s, words ÷ 3 + 0.5 s).
• Contrast ≥ 4.5:1 (≥ 3:1 for large text); over moving backgrounds add a scrim or stroke.
• Minimum body text size: about 4% of frame height in video.
• No overlap or overflow; everything stays inside title-safe.

#SAF · Motion safety

No dangerous flashes, no nauseating motion without an alternative.

Build rules

• Never more than 3 flashes per second (WCAG 2.3.1), including white cuts and strobes.
• Limit fast zooms, spins and large parallax; on the web, turn them off with prefers-reduced-motion.

#WEB · Animated web pages

Smooth, stable, respectful of prefers-reduced-motion and native scrolling.

Build rules

• Every non-essential animation stops or shrinks with prefers-reduced-motion: reduce.
• Budget: 95% of frames on time on a 4× throttled CPU; animate transform and opacity.
• No animation shifts the layout (CLS < 0.1).
• No scroll-jacking: scrolling stays native, with keyboard and trackpad.
• Interactions: visible feedback within 100 ms, UI durations of 100 to 300 ms.

#AUD · Audio

Compliant loudness, a clear voice, music ducking under the voice, sounds in sync.

Build rules

• Mix with umd mix (automatic ducking), then normalise with umd normalize (−14 LUFS, true peak ≤ −1 dBTP by default).
• Sound lands 0 to 1 frame after the image it underlines, never before.
• Fade in and out; no silence longer than 0.5 s at the start.
• Voice: humanised script, brand names pronounced correctly, 140 to 170 words per minute.
• Export 48 kHz stereo, AAC ≥ 192 kb/s (or Opus in WebM).

#CAP · Captions

Every word present, on time, readable and spelled right.

Build rules

• Known script: forced alignment (ctc-forced-aligner). Otherwise: word timestamps (faster-whisper, WhisperX).
• Generate with umd captions, check with umd captions-check.
• Two lines at most, 17 characters per second at most, inside the safe area, never hiding the key visual.

#REN · Render and encoding

A reproducible render, in the right format, with no visible loss or artefact.

Build rules

• Time is driven by the frame number (umd capture, useCurrentFrame, timeline.seek); never by the clock.
• Local fonts and assets (Fontsource): no network loading during render.
• High-quality master, then delivery encode H.264 High yuv420p, constant frame rate, faststart.
• Gradients: add grain or dithering against banding.

#LIC · Licences and rights

Every asset, model and dependency is allowed in the deliverable.

Build rules

• Record every asset in .motion/assets.lock.json with its source and licence as soon as it enters the project.
• Commercial deliverable: no non-commercial weights, no Tencent model in the EU, no embedded AGPL, no unlicensed asset.
• Remotion: free up to 3 employees, company licence above; declare it in project.json.
• Describe references in your own words; never have someone else's video copied.