thuann-vn/stick-crowd-mv

Stick crowd MV (painted 3D + code crowd)

Worked example in example/: "the lonely one in the crowd" — 2 shots in a Vietnamese book street (day, dusk), 4 shots in a rainy Japanese shopping street at night, and a time-lapse ending where the camera turns up to the sky. 30 s, 720 frames, 24 fps.

This skill sits on top of painted-3d-short (https://github.com/thuann-vn/claude-painted-3d-shorts): its p3d.py builds the blockout, renders the stills, projects the paintings and calls your setters every frame. Read that skill first. This one adds people.

Files

• tools/stick.py — the crowd. setter = stick.make(globals(), "<location>"), pass it to render(SHOTS, [setter]).
• tools/gait.py — the walk (numpy only, testable without Blender).
• tools/post.py — finishing pass over the rendered frames.
• example/scene.py, example/book.py — the two sets and their shots. example/jobs.json — the repaint prompts. example/paint/ — the paintings. example/m2m.json — idle clips (Mesh2Motion, CC0).

Order of work

1. Song first: cut the 30 s you want, find the lyric-line times, cut the shots on them (CUT in scene.py and stick.py).
2. Blockout + stills + repaint + projection exactly as in painted-3d-short. No people in the paintings.
3. Stage each shot in SHOTS of stick.py: hero start, direction, walk or stand, face pitch, lane half-width, bubble radius, people standing at the edges, side-street walkers (cross).
4. Test single frames: NTEST=2 STICK_OFF=<frame> blender -b --factory-startup -P scene.py -- --proj --shots 4. Look at them.
5. Full render (3 Blender processes at most), post.py, join with ffmpeg + the song clip.
6. Measure before you call it smooth: mean absolute difference between consecutive frames per shot; max ÷ median above ~2 means something popped (find the frame, look at it).

How the people are built (stick.py)

• Real geometry inside the set, so the painted street hides them correctly. Limbs of everybody = ONE mesh of 6-sided prisms whose vertices are rewritten every frame (numpy → foreach_set). Heads = spheres with a camera-space 3-tone cel shader + inverted hull. 200 people render in well under a second a frame in EEVEE.
• The head centre sits on the neck line (neck + dir × radius × 0.86), or the head floats off the body.
• Light: a list of lamps per location (position, colour, strength). Each head gets the sum as its object colour; normalise by the largest channel, do not clamp (clamping turns everything white). Cast shadow = the same limb mesh projected onto the ground from the nearest overhead lamp (length clamped), or along the sun in daylight; plus a soft contact blob.
• Umbrellas (rain): an 8-panel dome object per person, colour from the object colour, a third of them clear vinyl; the right arm is re-posed to hold the shaft; the canopy throws a round shadow. The hero has none.
• Phones: left hand to the ear, a thin slab against the cheek that glows at night.
• Rain: a box of thin camera-facing quads in front of the lens, at least 1.9 m away.

The walk (gait.py)

Mocap on a line figure looks messy (arms rooted at the neck spread wide, hands are noisy). The walk is procedural instead:

• Feet are planted: a stance foot stays on its spot while the body passes over it; position on the ground comes from the distance walked, so any speed is slide-free (measured: < 0.2 mm a frame).
• Hips ride high over the stance leg and low at double support; 2-bone leg IK with the knee forward.
• Arms hang from the neck close to the body and swing against the legs; the head stays calm.
• Specs: normal, stroll, brisk (stride, cadence, arm swing, lean differ per person) and hero (short steps, head down, arms hanging). Real cadences: stroll 80–105 steps/min, normal 105–125, brisk 125–140.

Crowd rules that were paid for

• One straight lane per walker for the whole shot. Steering during the shot (round the hero, round the lens) makes people dash sideways. Instead, before the shot starts, simulate each walker over the shot and move his lane aside once if he would enter the hero's empty circle, pass close to the lens, or stand between the lens and the hero.
• The loop must wrap out of sight. Fill the street from ~6 m behind the camera to ~50 m ahead and fade only at the wrap. A wrap inside the picture is a person jumping to another place.
• Never shrink or fade a person inside the picture.
• Density that reads as "crowded": 2–2.7 people per metre of a 6 m wide street.
• A pose clip with two keys left looping makes the figure shake: hold it (timeScale = 0).
• A standing person is never still: breathing, a slow shift of weight, the face turning.
• Couples walk abreast with the same spec, half a cycle apart.

Time-lapse (the whole film)

• In every shot the crowd's clock runs 10× faster (LAPSE, warp()); the hero alone lives in real time. That contrast is the idea of the film.
• Render 6 sub-frames per frame (SUB; the shot's frame count × 6 in scene.py); post.py averages them in linear light, so people become streaks and lights stay bright. 30 s = 4320 renders, about 25 minutes on three Blender processes.

Camera moves inside the painting

A projected painting only survives a small forward dolly. But the render camera may TURN, ZOOM IN and ROLL inside the painting freely: a pure rotation has no parallax. CAMK in stick.py holds keys per shot — (part of the shot, zoom, pan, tilt, roll, aim-at-hero) — applied after p3d's look(); pan and tilt are clamped to the room the zoom leaves, and a rolled frame is zoomed just enough to stay inside the picture. Give each shot one move with a reason:

• start tight on him, tilt up and pull back to reveal the crowded street;
• slow push-in on the face as he walks toward us;
• from the neon signs down into the street until we find him;
• a tilted horizon that settles while pushing in;
• a pan across passing umbrellas that lands on him;
• from above: tight on the bare head, turning, pulling back over the umbrellas;
• the ending: ONE very wide picture (14 mm, aimed ~20° up) and a 28 mm render camera that tilts from him up to the sky.

For an orbit or a real crane move, paint an extra wide picture for that shot.

Finishing pass (post.py)

Bloom on what is really bright, vignette, grain (stronger in the dark), a little colour fringe, slow handheld drift (not on the time-lapse), punch-in + flash on each cut, and a small pulse of zoom and glow on the song's kick (low band of the audio, frame by frame). Keep every effect small; together they are enough.

Music

Use music you have the rights to; no audio is in this repo. CLIP=/path/clip.m4a python post.py reads the song for the kick pulse. To fit a new track to shots that are already rendered, slide a 30 s window over the track and keep the offset where the onset strength at the cut times is highest. The example's final cut uses "Thunder & Rain Inst. & Construction Kits" by coruscate (https://ccmixter.org/files/Coruscate/61228), CC BY 3.0.