Skip to content

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

44.0 KB
Raw Blame

motion.md: how things move

See also: transitions.md (joins between beats), type.md (word cadence and arrival; meaning on the word), ui-demo.md (cursor, press and camera inside UI beats), layout.md (sizes and placement), director.md (film-level clock and grammar), world.md (light, depth, finish, grain; when to use shutter blur), slop.md (review after rendering; conditions for fades and holds), tools.md (render.mjs --shutter, look.py motion note).

Durations, curves, smear, clocks, camera and physical behaviour for a launch/explainer film. Two kinds of evidence are used. MEAS = frame-sampled statistics over 50 moments and 4,288 distinct moves (hook 12, turn 8, proof 24, cta 6; staggered siblings counted once; three tile-mosaic moments excluded). Moment IDs (M913...) = single-film observations. Where they disagree the MEAS number is the default and the anecdote is a signature move (use rarely, one per film). Transitions between beats are in transitions.md.

Conventions. Time as s (f@30); 1 f = 0.033 s, 3 f = 0.10, 6 f = 0.20, 10 f = 0.33, 15 f = 0.50, 30 f = 1.0. Sizes: % of frame height H or width W; px are on the 1280x720 reference stage (multiply by 1.5 at 1920x1080). Eases are named M.ease.<name>: plain t -> value functions (GSAP takes them directly, ease: M.ease.softLand, duration: frames/30); the runtime exports exactly crashIn, popOver, bounceHard, snapSettle, softLand, whip, glide, cruise, softInOut, gentleIn, accelExit, crashOut. Rule format: WHEN situation -> DO decision -- because effect (IDs).


1. The measured defaults

1.1 Duration by property (MEAS, distinct moves, frames @30)

Property n p10 / p25 / median / p75 / p90 Travel (p25 / median / p75)
translate x 1121 3 / 6 / 12 / 24 / 38 19 / 61 / 194 px (1.5 / 4.8 / 15% W)
translate y 1023 3 / 6 / 12 / 22 / 33 15 / 61 / 179 px (2 / 8.5 / 25% H)
scale 626 3 / 5 / 10 / 22 / 36 0.11 / 0.25 / 0.55 (delta)
rotation 271 3 / 6 / 11 / 21 / 34 11 / 25 / 124 deg
opacity 360 2 / 4 / 6 / 12 / 20 0.73 / 0.96 / 1.0
blur 391 3 / 4 / 7 / 10 / 14 2.5 / 5 / 10 px
size (w/h) 175 2 / 4 / 8 / 16 / 23 15 / 56 / 228 px

All moves pooled: median 10 f (0.33 s). Histogram of k: 2 f 8%, 3 f 5%, 4-5 f 13%, 6-8 f 18%, 9-12 f 15%, 13-18 f 14%, 19-30 f 15%, 31-60 f 10%, 61+ f 2%. So 13% are 2-3 f snaps and 12% run longer than 1 s.

R1.1 WHEN you set a duration with no other information -> DO use translate/scale entries 12 f (0.40 s), opacity 6 f (0.20 s), blur 7 f (0.23 s), exits 13-14 f for translate and 10 f for scale; treat 6 f and 24 f as the p25 and p75 edges -- because these are the measured medians and any film built on 20-30 f ease-outs for everything is slower than the whole sample (MEAS).

1.2 Entry vs exit, and role (MEAS)

Property Entry median k (p25/75) Exit median k (p25/75)
translate x 12 (6/23) 14 (7/27)
translate y 12 (6/22) 13 (7/24)
scale 11 (6/22) 10 (5/24)
rotation 10 (5/25) 10 (5/20)
opacity 6 (4/12) 7 (4/12)
blur 8 (5/11) 7 (4/9)
Role Translate/scale entry Translate/scale exit Opacity entry/exit Blur entry/exit Share "snap" curves Share ease-in Overshoot share
hook 12-14 f 10-12 f 6 / 6 7 / 9 17% 23% 12.3%
turn 12-15 f 14-16 f 5 / 6 8 / 9 10% 33% 11.3%
proof 12-14 f 12-16 f 7 / 7 9 / 6 11% 27% 5.7%
cta 6 f 9-13 f 4 / 6 6 / 6 39% 30% 4.3%

R1.2 WHEN the beat is a CTA/sign-off -> DO halve the entry times (6 f), make 23% of moves snapSettle and 25% softLand, and keep overshoot out; the 6 f median is for supporting moves (stack, pills, copy), while a hero mark or button runs 18-30 f on snapSettle (close.md C1, C8) -- because the close is the most staccato part of a film (CTA median move 6 f vs 10-11 f elsewhere; MEAS). R1.3 WHEN the beat is a turn -> DO give it the longest exits (14-16 f) and lean on acceleration (33% ease-in), with up to 11% of moves carrying a late overshoot -- because turns are built on things being pulled away and replaced (MEAS).

1.3 Landing decelerates, leaving accelerates (MEAS)

Direction (x/y/scale/rot/blur toward or away from neutral) Ease-out Linear In-out Ease-in
Landing (ends closer to 0 / 1) 58% 14% 10% 16%
Leaving (ends farther) 35% 17% 11% 37%
Scale leaving 33% 15% 9% 42% (49% of labelled scale exits)
Blur leaving 3% 19% 15% 63%
Translate exits (labelled) 26-29% 14% - 47-48%

Typical shapes, % of travel done at 10 / 25 / 50% of the time: snap 50/86/95, firm 26/60/85, soft 9/35/73, linear 9/24/50, S-curve 2/11/41, gentle accel 2/7/23, hard accel 0/2/8. Where ease-in sits: exits 45%, depart 29%, mid-move 24%, entries 16%, settles 14%, moves still going on the last frame 68%.

R1.4 WHEN an element lands -> DO ease-out (softLand 12 f, glide 10 f, snapSettle 11 f for hero/CTA) -- because it reads as heavy and certain; 58% of landings do this (MEAS; M913, M914, M916, M919 land 60-72% of travel in the first 4 f). R1.5 WHEN an element leaves across the frame or into a cut -> DO accelerate (accelExit 13 f, crashOut 17 f, gentleIn 9 f) and end at peak speed; 23-29% of exits still ease-out or run linear (shrink-to-centre, blur-outs, things swept by a camera move) so do not make every exit ease-in -- because an accelerating exit is "thrown away" and clears the frame with no dead frames; peak speed sits at t~0.9 (MEAS; M920 card 17 -> 111 px/f = 6.5x, M919 button edge 0.37 -> -0.16 W in 5 f).

1.4 Opacity, blur, cuts versus fades (MEAS)

  • Opacity: 6 f entry, 7 f exit (p25-75: 4-12 f). 23% of opacity tweens are 2-3 f (near-cuts), 44% 4-9 f, 33% >= 10 f. Shapes: linear 32%, softInOut 27%, glide 19%, gentleIn 10%.
  • Blur: median travel 5 px (p25-75 2.5-10), 7-8 f. Blur-in lands ease-out 58% / blur-out accelerates 63%.
  • Elements mostly arrive and leave by moving, not fading: 3,392 transform entries vs 584 opacity fades; exits 60% moves, 15% fades, 25% hard steps. Where an element simply appears or disappears: entry 57% hard 1-frame cut / 43% fade, exit 63% cut / 37% fade.
  • 90% of moments use an opacity tween; 42% also use an opacity jump >= 0.5 in one frame; 96% use hard visibility switches (median 20 distinct cut-frames per moment: hook 27, turn 16, proof 22, cta 6).
  • Slow colour/opacity ramps exist as a minority (>= 10 f is 33% of opacity moves): ground tint 0.5-0.7 s, grey -> black word 0.33-0.67 s, title alpha 0.8 s, last-phrase fade 10 f linear (M926, M941, M944, M920).

R1.6 WHEN something enters -> DO make it move (offset, scale or blur decaying) and use opacity only as a 4-6 f helper (0.12-0.3 -> 1) -- because motion carries the arrival; a pure opacity fade is the AI default (MEAS; M919, M920). R1.7 WHEN a UI piece, row or word should read as a system event -> DO hard-cut it on (1 frame, no fade) -- because 16% of entries and 25% of exits in the sample are hard steps and a step reads as state, not animation (MEAS; M939, M940, M921).

1.5 Stagger and overlap (MEAS, 97 stagger groups in 25 moments)

Interval between sibling onsets 1 f 2 f 3 f 4 f 5 f 6 f 7-9 f 10-14 f 15-24 f
Share of all gaps 56% 13% 8% 6% 7% 2% 3% 1% 4%
  • Per-group median interval 3 f (p25-75: 1-5 f; p10-90: 1-8.6 f). By class: translate 2.2-2.5 f, scale 3, opacity 3, blur 4.8. By role: hook 3.0, turn 1.5, proof 2.2, cta 8.0 (widest). Median group size 4 (p90 15).
  • Metronomic (CV < 0.25) 45%, uneven 55%. Overlap: interval / move duration median 0.5 (p25-75 0.27-0.83): siblings overlap, each tail runs under the next entrance.
  • What the 1-f gaps are: letters in a row, tiles in a cascade (13 tiles 1 f apart, M918), rows. Whole words run slower (2-5 f; 8-12 f when voice-led; see type.md).

R1.8 WHEN repeated siblings (letters, tiles, rows, avatars) arrive -> DO space onsets 1-3 f apart with uneven gaps (e.g. 3,2,2,2,2,2) and run each move ~2x longer than the gap so tails overlap -- because a whole wave under ~0.6 s reads as one gesture; a uniform timer-delay reads as a loading list (MEAS; M933, M935, M918). R1.9 WHEN a CTA stack of elements arrives -> DO widen the gaps to ~8 f (0.27 s); CTA words in a blur-resolve still run back-to-back at 4 f each (type.md 3.4) -- because the sample's CTA groups are the only ones that slow down (MEAS, 5 groups).

1.6 Overshoot: rare, late, small (MEAS)

  • 8.8% of moves (k >= 4) overshoot by > 3%: translate x 9%, y 12%, scale 9%, rotation 15%, opacity/blur ~1%. 37 of 50 moments contain at least one. Median overshoot = 19% of that move's own travel (p75 38%, p90 66%).
  • Shape: 88% of overshoots > 5% peak after 35% of the duration (median t = 0.65-0.70): a slow lean past the mark and back over 15-24 f, not a fast spring. 5% have an anticipation dip first. Early-peak crashes (peak before 35%) are ~1% of moves.
  • Where: entries 147, exits 91, depart 31, settle 26. Roles: hook 12.3%, turn 11.3%, proof 5.7%, cta 4.3%.
  • Absolute sizes seen in single films are small: button 1.055 (M914), bullet +4.7% (M963), badge +4% (M973), pill 0.51 -> 0.58 W (+14% width) held 3 f then back (M957), card width 0.63 -> 1.06 -> 0.82 W once (M939), phone/rim none (M922).

R1.10 WHEN you want a bounce -> DO use popOver (peak 1.08 at t=0.72, 15 f) as a small late lean, at most once per beat (hook/turn ~1 in 8 moves, proof/cta ~1 in 18); the one VISIBLE spring of the film is R1.11 -- because an overshoot on everything is the AI default; the sample puts it on under 9% and mostly late (MEAS; M914, M948, M973). R1.11 WHEN the payoff of a beat needs a physical reward (a card docks, a pill springs out) -> DO make it the film's ONE visible spring: snap to 0.9x of the target, spring to 1.0 over ~0.17 s with a small wobble, children on 2-f stagger with ~0.06 H overshoot (signature move; M948, M957, M939).

1.7 Drift and no dead frame (MEAS)

  • Drift moves (k >= 20 f, flat velocity, linear within RMSE 0.01): 221 distinct in 35 of 50 moments. Median 26 f (0.87 s), translate ~2.5 px/f (median travel 69 px), scale ~1.16x. Slower background creeps seen in single films: 1.1-1.5 px/f (M920, M921 -88 px = 6.9% W over 67 f@60), 0.02 W per 2.9 s (dot grid, M914), 0.027 W/s (M923), 0.07 W/s (tile field, M964).
  • Motion on ~99% of frames (in short moments). This was measured on 3-7 s moments from a gallery of dense, prompt-card clips; it is context for how much a held frame should stay alive, not a target for a whole film. Human-made whole films are calmer: their mean frame-to-frame luma motion has a median of 6.8 (IQR 4.4-9.8), a bright-ground film such as a major hardware launch is still for about 55% of its length, and the user has picked the calm, legible film over the frantic one. look.py prints your film's number as a question, not a gate: above ~10, does the film breathe anywhere? Below ~3, does it exist? Median share of perceptually still frames 1.1%; median longest still run 1 f; p90 3 f. Overall 8.5% of frames in the sample are still (604 of 7,091), mostly short stepped-motion stills. Still runs >= 6 f in 20% of moments, >= 12 f in 10%, >= 24 f in 2% (one end-card hold, 110 f, M946). At the "noticeable" threshold (0.5 px/f) 28% of moments park for >= 6 f; CTAs park longest (median run 12 f, 50% of CTAs >= 12 f; hook 0.6% parked frames, turn 6.8%, proof 3.0%, cta 9.0%).
  • Legit long holds are all payoffs: readable list 0.8 s (M939), shape pile 0.8-1.0 s (M938), static lockup 0.67 s (M968), headline 0.67 s with solid caret (M969), living mark 3.66 s with tip sway +-1% W (M946, a measured outlier: your end hold stays <= 1.4 s, gate G3), single number 0.47 s (M945).

R1.12 WHEN a thing must hold for more than 3 f -> DO keep something alive on it (no dead frame): linear drift 1-3 px/f, 0.5-1% scale creep, a sweep, a caret, a counter tick, a sway, video, light moving on the ground -- because a locked frame reads as a slide. A designed breath (0.6-1.5 s, a reason you can state, immediately before a payoff or after a matched cut) is wanted and may be quiet, with one living thing on it; only the payoff the viewer must read may be perfectly still, and then for under ~1 s (hero figure up to 1.2 s; a resolved CTA lockup 0.6-1.4 s with something alive, close.md R2.6) (MEAS; M913, M916, M920, M946, M955). R1.13 WHEN a clip ends -> DO end mid-motion (accelerating exit, still-shrinking lockup, half-faded field) -- because the cut is the edit and the next frame inherits the speed (M913, M916, M920, M929, M931, M944; ends-at-rest only in M946).

1.8 Statistic vs signature move

Topic Default (MEAS) Signature move Use the signature when
Entry curve softLand 15%, glide 14%; snapSettle only ~8% 54-72% of travel in the first 4 f + 0.4-0.6 s tail (M913, M914, M916, M929) the entry is the hero object or the first beat; one per scene
Exit curve 37% of leaving moves accelerate every exit accelerates off-frame (M919, M920) the exit crosses the frame or ends on a cut
Overshoot 8.8% of moves, late lean none at all except one spring (M948) or only +4-5% pops pick "none" for technical films, "one spring" for consumer films
Word stagger 1-3 f median (letters, rows) 6-7 f @60 = 3-4 f@30 inside a phrase, 10 f@60 before the stinger (M921) whole-word headline builds (see type.md)
Camera 18 f push/pull, median 1.46x one-frame snap 0.56-0.77x + 0.3-0.6 s tail (M913, M914) the camera reveals or reframes the product
Stills motion on ~99% of frames in 3-7 s moments (not a whole-film target; human films: mean motion median 6.8, IQR 4.4-9.8) 0.67-1.0 s payoff holds (M938, M968; M946's 3.66 s is an outlier, not a licence past the 1.4 s end-hold gate); a designed breath 0.6-1.5 s before a payoff after the claim has landed; never in hooks; always one living thing on the frame

2. The 12 measured eases

Fitted to the median curves of clusters of real moves (RMSE over 21 points: 0.02 = identical, 0.06 = close). 89% of measured moves lie within 0.12 of one of these, 53% within 0.06; the eases are anchors on a continuum, interpolate in between. "Share" = distinct moves nearest this ease (n = 3,573).

Runtime name Shape: % done at 10 / 25 / 50 / 75% Frames @30 (p25 / median / p75) Share What the humans used it for
M.ease.cruise 10 / 25 / 50 / 75 (linear) 6 / 9 / 16 18% fades (32% of opacity moves), drift, camera creep, counters, ambient motion; every drift is linear
M.ease.glide 16 / 39 / 69 / 91 (power 1.7, ~sine.out) 7 / 10 / 18 14% all-purpose medium x/y/scale/blur lands and some fades; entries and exits alike
M.ease.softInOut 4 / 17 / 50 / 83 (S-curve) 6 / 9 / 16 14% opacity (27%), blur (21%), counters off a handle, ground/colour ramps; no entry/exit bias
M.ease.softLand 30 / 59 / 85 / 95 (power ~3, expo k 3.4) 6 / 12 / 21 9% text rises, panels slide, scale-ins; 85% are entries; hook/proof workhorse
M.ease.gentleIn 1 / 6 / 24 / 55 (power 2.07) 6 / 9 / 17 9% exits (44% of it), blur-out and fade-out starts, slow push-ins running into a cut
M.ease.snapSettle 54 / 84 / 96 / 98 (fast snap + soft tail) 6 / 11 / 29 6% hero entries; CTA 23% of moves; hook 7%, proof 4%, turn 2%; 79% are entries
M.ease.whip 23 / 52 / 83 / 97 (power 2.5) 8 / 13 / 21 6% big fast moves; true whips median 7-8 f; 78% of entry whips land with it
M.ease.accelExit 0 / 3 / 16 / 47 (power 2.6) 9 / 13 / 19 5% the leaving curve: scale exits (44%), pans off-frame, peak speed t~0.9
M.ease.crashOut 0 / 0 / 6 / 32 (power 4) 11 / 17 / 31 4% camera pull-offs and exits that end in a hard cut at full speed; ends in an immediate stop
M.ease.popOver 10 / 46 / 96 / 108 (peak 1.08 at t 0.72) 11 / 15 / 27 2% the late lean past the mark: 168 source moves, 88% of overshoots; scale and translate entries
M.ease.crashIn 91 / 118 / 110 / 103 (peak 1.18 at t 0.27) 9 / 10 / 27 1% slams: ty/tx drops that touch the mark almost at once, pass it, relax back by t 0.9
M.ease.bounceHard 13 / 62 / 138 / 139 (peak 1.48 at t 0.62) 11 / 22 / 28 <1% the big overshoot (ty drops, scale pops, rotations); 1.5% of moves; use sparingly

Share by role (nearest ease): CTA snapSettle 23% / softLand 25% / cruise 8%; hook glide 16% / cruise 17% / whip 8%; turn cruise 22% / softInOut 16% / gentleIn 11%; proof cruise 20% / softInOut 16% / glide 15%. By where: entries mostly glide 17%, cruise 15%, softLand 14%, softInOut 13%, snapSettle 9%, whip 9%; exits mostly cruise 18%, softInOut 17%, gentleIn 15%, accelExit 10%, crashOut 9%.

2.1 Pick the ease

What is moving Ease Frames @30 Notes (IDs)
UI card / panel / modal landing softLand 12 (6-21) hero card in a close-up: snapSettle 11 f then a 0.4-0.6 s tail (M913 card 70% in 4 f; M916 card 62% in 4 f)
Card/panel inside a CTA snapSettle 6 (CTA entries median 6 f)
Whole-word arrival softLand 12-15 offset 0.03-0.07 H; hero first word snapSettle 10-18 f (M938 words ~15 f; M964 decays 0.6/f; M929 59% in 5 f, 91% by 16 f, tail to 36 f = snapSettle stretched)
Logo crash-in from 3.5-3.75x snapSettle 18-30 39% of the drop in frame 1, 86% by 6 f, 1.0 at 0.67 s (M929, M932); then cruise creep 1.0 -> 0.93 over 1+ s; use crashIn only if it should punch past the mark
Camera push (zoom in, pan toward) glide / cruise 10-22 pushes are 48% ease-out, 31% in; slow creeps are linear
Camera punch / reveal snap one-frame step, then softLand tail 9-18 ratio 0.56-0.77 in 1 f (M913, M914)
Camera pull / exit / pull-off accelExit or crashOut 13-17 camera pulls are 62% ease-in (median 1.8x, 17 f)
Whip (entry half) whip 7-8 peak speed t~0.25; blur-matched across a cut (see 6.4)
Whip (departing half) accelExit 6-10 ends at peak speed on the cut (M916, M919, M931)
Drift / creep / ambient cruise 26+ 1-3 px/f; 1.16x scale
Press down glide 2-6 cursor 2 f; button 0.63-0.80 min (M914, M919, M956)
Release popOver (or softLand for none) 6-10 +5% rebound (M914); none in M922, M931
Fade in / out cruise 4-7 softInOut for ground/colour ramps; gentleIn for fade-out starts
Blur-in (focus reveal) glide / softInOut 4-8 12 -> 0 px in 4 f (M919, M920)
Blur-out / defocus-out gentleIn 7 63% of leaving blurs accelerate
Counter / number roll values stepped per frame, steps decaying; or softInOut for a number tied to a handle 7-23 f M936 (+8-9/step, 7 f), M954 (16 values, steps 0.75 -> 0.07), M923 (50 f ease-in-out)
Exit off-frame / shrink to cut accelExit 13 scale exits 10 f
One-off physical reward popOver, rarely bounceHard 15 / 22 once per film (R1.11)

3. Entry and exit grammar

3.1 Entries

An entry spends two or more decaying channels at once, never opacity alone. Order of use in MEAS: transform moves 71% of entries, blur-in 8%, opacity-only 12%, hard cut 16%.

Channel Start state Decays over IDs
Offset, words/lines 0.03-0.07 H below the slot (20-50 px) or 0.03 W beside it, +14-46 px to the entry side; hero first word 0.10-0.24 H, shrinking per word (24/18/13/13% H) 12-15 f; tail 0.4-0.6 s M941, M938, M958, M929, M920, M917, M955
Offset, cards/panels rise from below the frame 0.6 H; slide 0.14 W; already mid-open at frame 0 62-72% in the first 4 f M916, M913, M933
Scale oversize start (see 3.3) or 0.46-0.75x of rest for identity-bearing objects 8-30 f M929, M947, M968, M967
Blur 3-6 px type, 8-12 px status words, 12-17 px hero counter, 18 px logo 4-8 f M919, M920, M954, M968
Smear directional, frame 0-1 only (see section 4) 3-4 f M916, M929
Colour accent tint -> ink 4-17 f (6 f at 30 fps, 0.28 s at 60) M967, M974, M966
Opacity 0.12-0.3 -> 1 4-6 f M920, M919

R3.1 WHEN anything enters -> DO start the beat already in progress on frame 0 (a card mid-open, a word half-risen, a hand at speed) and cut onto it -- because the first frame of the clip carries the momentum of the previous one (M933, M934, M938, M914, M918). R3.2 WHEN an item pops in whole -> DO give it a 1-f reveal of its final form with a small settle (3-10% over 7-14 f) rather than a morph -- because a hard cut into ~90% of the final state then a settle reads as engineered (M962, M964). R3.3 WHEN several entries share a beat -> DO give each a different offset size and ease tail (fan that closes), not one shared curve -- because different decays read as separate bodies (M937 words, M938).

3.2 Exits

Exit type Mechanics Use when IDs
Accelerate off-frame ease-in 6-10 f, 17 -> 111 px/f; object stays sharp, only the label may smear the element is "thrown away" M920, M919, M914
Shrink into the centre scale to ~0.3 with ease-in 4 steps/0.24 s, strips jittered +-1 px for the last 0.4 s the old act collapses before a hard cut M947, M967
Clip / mask wipe through glyphs clip edge sweeps through the line in 5 f, accelerating, while the line slides the same way text leaves M962, M947
Words out one frame apart 1 f apart, 3-4 f each, up 0.07 H and grey; overlap the next entrance by ~2 f phrase replaced M938
Dissolve to cells or strips real cells, stepped coverage the thing becomes the next scene transitions.md
Hard cut off 1 f; 25% of exits state change MEAS
Single-object fade linear 10 f or stepped 4-5 f one object only (a line, a last phrase), never between scenes M920, M934, M938

R3.4 WHEN an element is removed -> DO let it exit by moving (60% of exits) or by one hard frame, and give the exit ~3-4 f when it is a word (exits are faster than entrances) -- because a fade leaves a ghost that the eye has to wait out (MEAS; M938). R3.5 WHEN a drawn highlight or stroke leaves -> DO exit with NEW motion (the tail pushes out, curls to a smaller arc), not the entry reversed -- because reversed motion reads as an undo (M933).

3.3 Start oversized and come down

Object Start Rest Time and shape IDs
Logo/wordmark sign-off 3.5-3.75x (ink wider than the frame) 1.0, then creep to 0.57-0.93 39% of drop in frame 1, 86% by 6 f, 1.0 at ~20 f (0.67 s) M929, M932
Brand mark (calm) 1.86x 1.0 48% in 0.13 s, 67% by 0.33 s, last 0.5% over 2 s M959
Logo plate scaleX 1.46 1.0 excess x0.72 per frame@24, 90% in 0.25 s M965
Logo emblem 0.28 W, 18 px blur 0.10 W excess halves every 2 f, blur 18 -> 0 M968
Opening word 0.39 H / 0.58 W ink 0.09 H 0.30 s strong ease-out; ~86% of the shrink in 0.13 s M947
Thesis word 0.91 W ink, 0.33 H slow shrink 18% in 0.6 s, then accelerate then blur 0 -> 18 px M937
First kinetic word 1.94x, accent tint 1.0 1.94 -> 1.31 at 0.30 s -> 1.0 at 0.67 s M917
Cursor 1.5-2.2x (name-pill cursors), 0.36 W (glossy), 5.1x mostly off-frame and blurred 1.0 scaled about the TIP; 60% of travel in step 1, creep 0.5-0.9 s M953, M956, M959
Camera on a UI fragment 3.3x (card), 1.4-1.6x (modal, card), 4.39x (button zoom), 2.5x (grid) 1.0 one-frame snaps + tail; or two surges M913, M914, M918, M919, M937
Logo grid / tiles 2.5x, tiles are giant cropped squares 0.875 two surges, near-stall between M937

R3.6 WHEN an object with identity (logo, word, cursor, card, mark) enters -> DO start it at 1.4-5x and scale down to rest, never grow it from zero, never fade it up -- because "arrives from the camera" reads as authority and makes the hit without a bounce (M929, M932, M947, M959, M968). R3.6b WHEN you apply R3.6 -> DO read it as a rule about arrival (the thing decelerates in from oversize, never grows from zero), not as a recipe for a giant cropped first word or a logo that crashes in at every end. Spend scale once per film and let the ending get smaller (director.md section 6); a calm brand's logo may arrive at 1.1x with a long ease (close.md C4) -- because "start big" repeated in every beat is a house style, and an arrival that is always the same size says nothing. R3.7 WHEN the rule seems to fail -> DO allow growth from small only in these cases: items in a 1-f-apart cascade (13 tiles 0 -> 1.12 -> 1.0 in place, M918), a growing object that IS the next scene (bead 0.04 W -> disk 17x, M922; dot -> disc 0.33 H, M962), a loader dot -> QR (M954), and identity objects that start no smaller than ~0.30-0.60x (cards 0.30, button 0.60, disc 0.46: M967, M959, M945) -- because in each the growth is the content.


4. Smear and blur

Three grammars, chosen once per film (R4.1): per-element smear, crisp steps, or camera shutter blur (R4.8). All use blur as a focus tool (type resolves, backgrounds rack-focus); only the smear grammar adds per-element directional smear on fast moves, and only the shutter grammar blurs everything by its own speed.

R4.1 WHEN you choose a film's blur language -> DO pick ONE and keep it: smear (directional blur along the travel axis on every fast move; consumer/AI-agent/premium, M916-M920, M926, M929, M931, M937), crisp steps (hard steps and cuts, no smear even at 165 px/f; technical, developer, indie, flat vector, M913, M914, M921, M934, M939, M962-M969), or shutter blur (R4.8: a real camera shutter applied at render time to a smooth film). In a crisp film change content during the move instead (panels exit with the pan, M934) -- because mixed grammars read as different authors. Never combine shutter blur with per-element smear (double blur) or with an on-twos stepped film (the blur would smear the held poses).

R4.2 Directional smear (smear films). WHEN an object makes a fast move -> DO apply a blur along the travel axis only, on the FIRST frame of the move (peak), decaying ~halving each frame over 2-4 f (kernel 20 -> 4 -> 2 -> 0 in 3 f; a thin 2 -> 0.7 -> 0.25 gaussian over 6 f), plus 1-3 low-alpha trail copies (0.22/0.12/0.06) peaking on frame +1 -- because the smear reads as speed from frame 1 and is gone before the object is read (M929, M919, M920).

Situation Smear length / blur IDs
Word arrival 8-30 px along the entry axis (10-20 px vertical for rising words, 24-30 px horizontal for slide-ins) M917, M920, M929
Card/panel move 15-35 px M926, M944, M942
Cursor, mascot sweep 8-25 px along travel M919, M943
Whip grows 3 -> 65 px (blur 2 -> 5) during the last 3-5 f; 100-140 px on a whip-out M916, M931, M919
Pan jump / one-frame camera smear up to ~500 px around a 140 px core for ONE frame M928
Drop/burst on a vertical move 100 px vertical smear on one frame M919

Batch data: 3-65 px in ordinary use, 100-140 px on whips. PRACTICE (extrapolation): set the length to ~0.5-1x of that frame's displacement and keep it under ~0.15 W.

Implementation. SVG <filter><feGaussianBlur stdDeviation="sx sy"/></filter>: "20 0" horizontal, "0 20" vertical, applied via filter: url(#id); tween stdDeviation per frame. For a diagonal travel (M916 45 deg, M918 -30 deg) put the element in a wrapper rotated to the travel angle, blur along x, rotate the element back. For radial streaks (crash-in logos) stack 64-96 scaled low-alpha copies whose length halves each frame (M929, M932). One-sided smears (trail only behind the object) read better than a symmetric blur for cursors (M943).

R4.3 Colour split. WHEN a first word or a hero object arrives with smear -> DO add an RGB split (cyan -/+10 px, yellow +10/-5 px, blue) on the FIRST word only, peak 1-4 f, gone by <= 9 f; for a crash-in use a cyan/amber fringe of ~15 px decaying in ~5 f -- *because* the effect reads as speed and is spent where it matters (M929, M932, M927, M930). **R4.4 Blur-focus reveal.** WHEN a status line, CTA word or label should appear without moving -> DO resolve it in place: blur 12 -> 0 px, opacity 0.1 -> 1, 4 f per word, words back-to-back (or 8 -> 6 -> 4 -> 1.5 -> 0 px with opacity 0.12/0.3/0.65/0.9/1) -- because it reads as inference/focus, nothing types (M919, M920). R4.5 Rack-focus backgrounds. WHEN a UI sits behind a headline -> DO blur the UI, not the type: 0.7 -> 5.5 -> 10 -> 11 px over ~0.7 s (cap ~11 px) with 0.1 darkening while the headline stays crisp; for a defocused-but-readable backdrop hold ~4 px and let the cursor blur with it; pull focus back by animating blur 4 -> 0 over ~0.46 s, never by opacity -- because the proof stays visible and the type stays legible (M918, M942, M943). R4.6 Depth defocus. WHEN a UI exits or recedes -> DO defocus by depth (top row 0.9 px, bars 3.2, labels 4 px) and leave the hero stat sharp; secondary panels sit at 7-15 px (M953, M956, M958). R4.7 Blur peaks to expect. Entering card 6 px, disc dock 12, orb 15, button burst 25, roll 22, hero counter 12-17, whip 9, logo 18 -> 0 (M939-M946, M954, M958, M968). R4.8 Shutter blur (render-time). WHEN the film is smooth (30 or 60 fps interpolation, camera moves, type that glides) and you want the look of a real lens -> DO render with node scripts/render.mjs --shutter 180 --subframes 8 (add --grain 0.03 for a little film grain): each output frame is the average of 8 sub-frames spread across a 180-degree shutter (half the frame interval), so every moving thing, including type and the camera, is blurred by its own speed and a still thing stays sharp -- because continuous, physically consistent blur is what separates a rendered camera from stacked tweens, and it costs no authored smear. Rules: this is the film's blur grammar (R4.1), so author no per-element smear and no radial streaks on top; fast cuts and whips need no extra blur; stepped, on-twos, glitch or crisp films must not use it; render time grows with the sub-frame count (use 4 for drafts, 8 for finals); keep grain at one level for the whole film (world.md finish). Check the result at real speed: type that must be read should be sharp within 3-4 f of landing.


5. Frame clocks

R5.1 WHEN you start a film -> DO choose ONE clock per moment: (a) smooth 30 or 60 fps for brand morph/UI dive (M921-M924, M973, M974); (b) smooth 24-30 for camera/type with stepped reveals (M964-M966); (c) on twos: every pose holds 2 frames, every onset lands on an even frame (M947-M949, M967-M969, M945 pairs); (d) 12 fps drawn things on a 24 fps world (hand-drawn loops, ring/fill/check poses) while the camera, type and cursor stay smooth (M933, M934); (e) 30 fps hard cuts on a constant interval (M962: every 10 f = 0.333 s) -- because mixed step rates read as a bug. R5.2 On-twos implementation. Quantise time: const f = Math.floor(frame/2)*2 and seek every timeline to f/30. Land all onsets and durations on even frames; the camera timeline gets the raw frame, the drawn layer's timeline gets the quantised one (two timelines, one seek). In a 24 fps comp 12 fps drawn poses are plain 2-f holds; in a 30 fps comp use 2-f holds (15 fps) or quantise t as Math.floor(t*12)/12. Stepped typing, glitch rectangles, radius steps and dial states are discrete states by design and are never tweened (M929 rectangles step every 2 f, M926 radius at five instants). R5.3 Stepped keyframes. WHEN a thing must feel hand-posed -> DO reach each pose in 1 f and hold 1-2 f (a phone: 6 poses at 0.03/0.07/0.10/0.20/0.30/0.40 s; a camera: held steps at 1.70/1.80/1.87/1.90-1.97 s), and run the easing curve across the poses (M955, M957) -- because pose holds are what make the motion look drawn. R5.4 Real 60 fps morphs. WHEN you want a liquid brand transform (flip, dive, rosette spin) -> DO use a true 60 fps clock with no holds on the morph (glyph squash 2-2.4 f per slot, icon spin to ~30 deg/f, dive x1.5-2 per 4 f@60) (M973, M974). R5.5 Held duplicate frames. Real frame-accurate holds do appear inside smooth moves (a repeated frame in a camera jump, M940, M946 world jerk, M928 whip) -- one repeated frame in a fast move adds impact; use sparingly (1-2 per film).


6. Camera

6.1 One group, not many tweens

R6.1 WHEN you simulate screen use -> DO put the entire UI (card, panels, chips, scrubber, video) in ONE camera group and move that single transform; dot grids/backgrounds stay screen-fixed; cursor, mascot and drag stack live on their own top layer outside the camera and scale with camera scale; screen-fixed headlines are not parented to it; text is never zoomed separately -- because one transform reads as a lens, many tweens read as animation (M913, M914, M917, M918, M919, M953, M958). R6.2 WHEN you reframe -> DO prove the scale with a fixed texture: a dot grid pitch doubling, the cursor growing, a hairline thickening -- because a 2x zoom must read as 2x and not 3x (M931, M914).

6.2 The one-frame snap and its tail

Move Numbers IDs
Reveal snap scale 2.50 -> 1.41 in 1 f (ratio 0.56), y -832 -> -226; then 0.56 s decelerating settle to 1.0; strongest frame of the moment M913
Reframe snap 1.469 -> 1.137 in 1 f (ratio 0.77), fast ease-out to canonical by 0.32 s, regrow to ~1.13x over 0.36 s M914
Stepped zoom mid-typing 1.0 -> 2.46x in ONE frame; peripherals step to 0.65-1 px blur and 0.4-0.5 opacity in the same frame M923
Push burst 1.064 -> 1.299 over 2 f (-82 -> -389 px), decays to 1.368 by 0.4 s, 0.6 s hold with under 1 px creep M913
Hard scale cut 0.037 W -> 0.136 W (3.7x) in one frame, background keeps its own scale M964
Hard scale cut + content jump scale x0.42 and a multi-word tail appears in the SAME frame, then x3.86 M949
Discontinuous camera jump button 0.04 W -> 0.27 W after an ease-in push, then pushes on to 0.31 W M956
50% double-exposed frame, then 2.33x close-up no zoom tween M943

R6.3 WHEN the camera should reveal or reframe -> DO one-frame snap (ratio 0.56-0.77) followed by a 0.3-0.6 s decelerating tail (softLand, 9-18 f); never one global ease-out zoom; sometimes fire the snap on a trigger (the last checklist tick releases the pull-back 0.40 s later, M913) -- because the jump has the punch of a cut, the tail lets the frame be read (M913, M914, M923).

6.3 Pans and parallax

R6.4 Two pushes. WHEN a pan crosses the scene -> DO build it as two pushes with different accelerations, never one ease: push 1 fast at frame 0 (x 1.09 -> 0.87 in 0.25 s, -> 0.78 at 0.79 s), slow drift 0.78 -> 0.74 to 1.08 s, push 2 accelerating to a peak 0.025 W/f at 1.54-1.63 s, settling by 2.08 s; or fast/slow/fast/slow zoom speeds that never hold (M918) -- because a single ease reads as a keyframe, two pushes read as an operator (M966, M918, M917 exponential decay tau 0.36 s). R6.5 Parallax. WHEN layers pan -> DO run them on separate tracks at 2-2.4x speed ratios (tiles > diagram > orb: diagram peaks 0.05-0.06 W/f vs orb 0.025), stagger the exits (text clipped at 1.54 s, grid at 1.92 s), and let some layers not scale at all (tile field keeps its own scale during a 3.7x cut; the unit label travels 4x less than the timer) -- because depth comes from speed difference (M966, M964, M957 stars ~0.07 H/f, M936). R6.6 One shared track. WHEN you pan a wide timeline or list -> DO run every part (sidebar, ruler, blocks) on one track, peak ~126 px/f, ~2.9 frame widths, long tail, no zoom, clipped by a fixed left gutter of 0.03 W -- because a rigid track reads as camera, growing bars read as charts (M940). R6.7 Accelerating scroll. WHEN a long log scrolls past as proof of volume -> DO accelerate from ~6 px/f to ~250 px/f (0.35 H per frame) over ~1.8 s, no blur, let the viewport run empty at the end, decelerate over 9 f onto one readable block and hold it 0.8 s (M939). R6.8 Slow-in, fast-middle, long-tail. WHEN a card/panel carries the camera settle -> DO 27% of travel in the first 0.46 s, 42% in 3 f, the last 25% over a 0.7 s tail with scale 1 -> 0.6, no overshoot, leaving 0.20 W of the edge on screen (M933); for a lateral card swap 0.5 -> 0.19 -> 0.12 W with 84% by 4 f (M913).

6.4 Whips

MEAS (207 whip-sized moves, <= 14 f and >= 300 px or >= 0.25 scale, in 37 moments): median 8 f, 68% decelerate (entry whips 79% ease-out), 24% accelerate (mostly exits), peak speed at t = 0.25; 159 are entries, 32 exits. Peak per-frame displacement ~13% W (M934: 12.9% W/f at its peak; 27% of travel over 10 f of ease-in, then 31% in 2 f and 49% in 4 f, then a 15-f tail).

R6.9 WHEN the camera whips between regions -> DO give it 3-5 f of travel around the peak (entry: whip, 7-8 f; departing half: accelExit), with the cut landing at peak speed and the new content starting along the same vector -- because motion continues in the viewer's head across the cut (M934, M931, M916, M919). R6.10 Blur-matched. WHEN the film uses smear -> DO ramp blur and smear on the last 3 f of the old view (text 8-20 px, ~100 px smear, up to 140 px), cut with a one-frame discontinuity into a view that is already blurred (button as a smeared pill, 2x scale swap) and resolve over ~8 f; in a crisp film do not blur, change the content during the whip (panel, logo and link lines exit with the pan) (M931, M937, M934). R6.11 Content in the whip. WHEN a whip exits one thing and enters another -> DO have an anchor ride the pan (cursor, bead, card), size constant (a pan is not a zoom: badge size constant, M934) -- because the eye needs a carrier through a fast move (M922, M934).

6.5 Push and pull multipliers (MEAS)

Group n Multiplier p25 / median / p75 / p90 Duration median Ease
all pushes 256 1.19 / 1.46 / 2.0 / 4.0 10 f (p25-75 6-22) 48% out, 31% in
all pulls 280 1.15 / 1.32 / 1.72 / 3.7 13 f 50% out, 31% in
camera-wrapper pushes 21 1.17 / 1.63 / 2.0 / 4.3 18 f 52% out, 24% in
camera-wrapper pulls 21 1.15 / 1.82 / 3.3 / 4.4 17 f 62% in, 29% out

By length: punch <= 6 f -> 1.26x median; short 7-18 f -> 1.39x; medium 19-45 f -> 1.52x; long drift 46+ f -> 2.4x (p75 5.6x). 84% of moments contain a >= 1.5x scale move; the largest per moment has median 4x (includes scale-ins from small).

R6.12 WHEN you size a camera move -> DO read the multiplier off the duration: 1.25x for a 6-f punch, 1.4x for 12-18 f, 1.5x for 20-45 f, 2.4x only for slow drifts; for a reveal pull go 1.8x with an ease-in; the per-moment hero scale move is ~4x -- because these are the measured pairings (MEAS).

6.6 Dive zoom, constant octave rate

R6.13 WHEN the camera dives from a brand into the product -> DO scale from frame 0 with no hold at a constant octave rate: 1, 1.02, 1.15, 1.46, 2.28, 4.52, 8.98, 14.67, 21.9 at 0/.067/.133/.2/.267/.333/.4/.467/.533 s, i.e. x1.5-2 per 2 f@30 (per 4 f@60), accelerating 0.4 s then decelerating to ~1.1 s; implement by interpolating log(scale) (scale = exp(lerp(0, ln 21.9, e))); translate toward a feature of the logo (the underscore) not the centre so it exits the frame edge; let that feature become a near-full-frame mass, fade it in place in 3 f, show 3 blank frames, then zoom a container OUT to rest (0.05/0.98 -> 0.25/0.78 over 0.62 s, stroke 10 -> 4 px) -- because the logo is felt to contain the product (M974). Use vector or oversampled sources; a 22x raster zoom softens.

6.7 Camera micro-drift

R6.14 WHEN the camera appears to hold -> DO keep it moving at >= 0.003 H/f or ~1 px/f: grid drift 0.02 W per 2.9 s, scrubber 0.9 px/f, slow right drift 0.027 W/s, slow shrink 1.0 -> 0.82 per 0.6 s, badge growth 120 -> 208 px over 3 s (M934), title row -0.01 W per 0.44 s; where the camera must be still, move something else (M913, M914, M923, M934, M936, M964). R6.15 WHEN the camera ends a moment -> DO end it still accelerating with no settle (pan ends hard at 0.39 -> 0.31 W, M913; accelerating roll to -76 deg with 22 px smear, M944; accelerating card up to the last frame, M924).


7. Physical bodies, pulses and springs

R7.1 Bodies. WHEN several shapes leave or arrive as "physical" -> DO simulate each as its own body (own path, rotation, one small bounce 0.03 H, slight rotation overshoot -95 -> -90 deg), bring at least one in late (0.60 s) so the choreography is not uniform, settle the pile by 1.8-2.2 s and HOLD 0.8-1.0 s motionless as the weight beat, then bring the next idea; shapes pass behind words without covering ink -- because a group translation reads as a slide, individual bodies read as consequence (M938). R7.2 Pulses. WHEN something is pressed -> DO an asymmetric press/release: fast squeeze (cursor 2 f to 0.71-0.9; button to 0.76-0.80 over 11-13 f@60 = 0.2 s), slower release (19 f@60 = 0.32 s, or a +5% rebound, 1.055), the target reacts 0-1 f after the cursor (button 0.63-0.80, arrow 0.75, avatar 0.77, card 0.97), the 391 raw pop-and-return pulses in MEAS run median 23 f total (0.77 s). Compression reference: arrow/avatar 71-77%, button 97% (close-up) to 63-80%, cursor close-up 88%, cursor body 0.81-0.92. Brightening of the target (#4C4DF5 -> #7FA7F5) beats a ripple ring. NEVER a ripple ring or white flash as the default (M914, M919, M922, M931, M953, M956, M959, M939-M946). R7.3 Springs. WHEN you add overshoot -> DO keep leans to R1.10 and spend the visible spring once per film (R1.11). popOver for the lean, bounceHard (peak 1.48 at 0.62, 22 f) only for a ty drop or scale pop in a film that is otherwise ease-out (MEAS 1.5% of moves; M948, M957, M973). R7.4 Anticipation. WHEN a move needs wind-up -> DO use a 3-f drift opposite to travel (+196 -> +234 px before sliding left), or a 0.1 s still before a press (hand stops 0.10 s, violet shadow grows 0.1 s before the burst) -- because wind-up makes the next move read as intended; only 5% of moves carry it, so use it on the key one (M929, M930, M943). R7.5 Stepped-then-smooth objects. WHEN a hand-posed thing meets smooth motion -> DO keep it on its own clock (R5.1) and let a physical bit (tail, loop, ring) exit as new motion (M933, M934).


Slop tells for this topic

An AI default would Do instead
Use one ease (power2.out, 0.6 s) on every tween Pick per role from the 12 eases: softLand 12 f, glide 10 f, cruise 6 f fades, accelExit 13 f exits; CTA moves 6 f
Fade every element in with opacity 0 -> 1 over 0.5 s Move it: offset/scale/blur decaying over 12 f, opacity only as a 4-6 f helper; hard-cut 16% of entries
Ease-out everything including exits Landing 58% out; leaving 37% in; scale and blur exits accelerate; exit across the frame ends at peak speed
Bounce every pop (back.out(1.7)) Overshoot on < 9% of moves, late lean (popOver), once per film as a payoff spring; CTA/proof ~5%
Constant stagger (0.1 s for all) 1-3 f uneven gaps overlapping at ~0.5 of the move; CTA 8 f
Grow a logo from scale 0 and fade Start 1.4-5x and scale DOWN, with excess halving every 2 f; snapSettle 18-30 f
Smooth global zoom-in One-frame snap 0.56-0.77x + 0.3-0.6 s tail; two-push pans; hard scale cut; constant-octave dive
Blur everything the same 20 px or smear on some beats and not others One blur grammar per film: per-element smear on frame 1 of fast moves decaying over 3 f, OR crisp steps, OR render-time shutter blur (R4.8); backdrop blur capped ~11 px
Hold a settled frame for 1-2 s No dead frame: one living thing (1-3 px/f drift, a creep, a sweep, moving light) on every hold; a designed breath of 0.6-1.5 s only before a payoff; only the payoff holds still, < 1 s
Drive every film toward "motion on 99% of frames" and constant energy That figure is for 3-7 s moments; human whole films run a mean motion median 6.8 (IQR 4.4-9.8) and breathe; use the band as a question, not a dial
Start big on every logo, word and cursor in every film Arrival decelerates from oversize, once; spend scale once and let the ending get smaller (R3.6b)
Press with ripple ring and glow 2-6 f squeeze to 0.75, target reacts 0-1 f later, brightens, slower release; no ring
Mix 24, 30 and 12 fps motion in one beat One clock per moment; on-twos quantised to even frames; drawn things at 12 fps on a smooth world
Move every part with its own tween One camera group; cursor on its own layer; text never zoomed alone
Treat a whip as a fast ease 8-f whip, peak speed at t 0.25, cut at peak, blur-matched both sides (or content changes inside it in a crisp film)