Source confidence: medium
See also: Deeper Dive — One Dancer, a Thousand Bodies — Ensemble Choreography from Herd Physics →

One Dancer, a Thousand Bodies — Ensemble Choreography from Herd Physics

Insight class: Emergent Capability · Difficulty: Experimental · Quality: 3/5 Connected entries: 9 · Tools: After Effects, Arc, Blender, Claude, FFmpeg, JupyterLab, Microduck, Python, React, Unity, Unreal Engine, XInputReader, ppf-contact-solver

In short: Produces a 90-second ensemble piece in Blender where 200 to 1,000 bodies perform one dancer's captured phrases, and the choice of phrase and facing for each body is governed by an Ising or Potts herding model whose temperature the artist drops live through the critical point on a gamepad axis.

Key fact Detail
What it covers 3D, animation, open source, art
Tools Blender, Python, After Effects, Arc, Claude
Posts covered 9 posts (3 S-tier, 3 A-tier) · top: Blender rigid-body spaceship — motors… · PPF Contact Solver — ZOZO's open-source GPU… · Blender — • 500+ body, hand & face landmarks…
Creators @marc.kaz, @6ab3, @dima.easyart
Sources dated Jan 2026 – Oct 2026
Difficulty Experimental
Prompt system Run-it-yourself prompts
Last verified 2026-10-22

Thesis

Crystal Pite's ensemble pieces make dozens of dancers move like one organism, and that takes a company and months of rehearsal. Now webcam mocap captures one dancer, text-to-motion models spin out variations, and the quant collection's Ising herding model supplies the missing rule: one temperature knob that moves a crowd from noise to sudden collective turns. That last step is a market model used here as a choreographic score.

Why it matters

A solo dancer or animator can compose crowd choreography with real intention — not a boids screensaver — and render it with production cloth physics. It's a new kind of score: the choreographer sets coupling and temperature, and the dancer's signature motion spreads through the crowd.

The evidence — connected entries

Crystal Pite's The Seasons' Canon — contemporary ballet of mass ensemble motion

Creator: @dima.easyart · Tier: B

The target aesthetic — Pite's mass ensemble moving as a single organism

Crystal Pite's 'The Seasons' Canon' (Paris Opera Ballet, 2016, set to Max Richter's Vivaldi recomposition), back at Palais Garnier. Worth studying for how a large ensemble moves as one organism.

View original post

Blender — • 500+ body, hand & face landmarks per frame

Creator: @marc.kaz · Tier: S Tools: Blender, Unity, Unreal Engine

Capture — FreeMoCap turns ordinary webcams into markerless 3D mocap of the one dancer

FreeMoCap turns ordinary webcams into professional markerless 3D motion capture. — @marc.kaz; uses Blender, Unity, Unreal Engine; demonstrates motion capture.

View original post

HY-Motion 1.0 — Tencent's text-to-3D motion generation

Creator: @6ab3 · Tier: A

Variation — text-to-3D-motion fills the crowd with variants that still read as human

by Tencent Hunyuan — featured by @marc.kaz | Billion-parameter text-to-3D motion generation. Type "a person walks forward then starts jumping" and get skeleton-based character animation. 200+ action types, 700hrs of pro mocap training. Works on human + robot skeletons. Free on HuggingFace. | DiT + Flow Matching 200+ action types Free

View original post

Ising Model — modeling herd behavior and market phase transitions

Creator: @quant.dhawan · Tier: B Tools: Python

The score — Ising coupling and temperature give order-to-chaos transitions and symmetry breaking

3D lattice visualization of the Ising model applied to financial markets. Blue=sellers, orange=buyers. At high volatility the market is random; as social pressure increases, clusters form leading to spontaneous symmetry breaking (crashes/trends). Built in Python for agent-based modeling.

View original post

Shepherd dog flock simulation — pressure-field + boids rules

Creator: @simulateitnow · Tier: B Tools: Arc, React

Spatial layer — boids plus pressure-field herding keep bodies from colliding and let you steer the mass

@simulateitnow explains a flock-herding simulation combining a pressure-field force F=(1-d/R)^1.2 with classic boids (cohesion/alignment/separation). Great reference for swarm/multi-agent simulation math.

View original post

Microduck robot — reinforcement learning with PPO and reward shaping

Creator: @summerr.exe · Tier: A Tools: Microduck

Wildcard — RL reward hacking shows how agents invent unintended movement; a source of strange gestures

Training Pollen Robotics' open-source Microduck to stand: define a reward function (e.g. points for target body height), simulate 4,096 parallel robots, and let PPO update the policy toward higher-reward actions. The result 'stands' in an unintended way, a clean demo of reward hacking and why reward design is iterative.

View original post

PPF Contact Solver — ZOZO's open-source GPU cloth and soft-body physics

Creator: @marc.kaz · Tier: S Tools: Blender, JupyterLab, ppf-contact-solver

Finish — GPU cloth solver gives a thousand costumes penetration-free fabric

ZOZO open-sourced ppf-contact-solver, its internal virtual try-on physics engine: penetration-free fabric, rope and soft-body simulation with strain limits, up to 180M contacts, full GPU acceleration, a Blender plugin and a JupyterLab interface.

View original post

Blender rigid-body spaceship — motors, constraints and drivers for flight

Creator: @ianhubertz · Tier: S Tools: Blender, XInputReader

Live control — Blender drivers and gamepad input let you conduct temperature in real time

Ian Hubert shows physics-driven vehicle animation in Blender: active/passive rigid bodies, hinge and spring constraints, motor constraints pushing against the ground as thrusters, gravity off, and custom properties wired via drivers with scripted expressions for throttle and lights. Mentions Erindale's XInputReader addon for gamepad control.

View original post

Claude Opus 5.5 to Blender — 3D animation built entirely as code

Creator: @nateherkai · Tier: A Tools: After Effects, Blender, Claude, FFmpeg, Python

Glue — the agent writes the Blender Python that wires capture, crowd and render together

Claude Opus 5.5 wrote everything: a web mock of the After Effects UI, a Python script that builds the Blender scene, the character animation, plus the music and sound effects. Blender rendered 360 frames on GPU in about 15 minutes and FFmpeg stitched the video. Full prompt is comment-gated.

View original post

Action sketch — how to implement this

Capture 5 minutes of one dancer with FreeMoCap (2–3 webcams), retarget to a Blender rig, and generate 20 variants of each phrase with HY-Motion prompts Have Claude write a Blender script: 200–1,000 instances on a lattice, each with an Ising spin (which phrase or facing to adopt) updated by neighbor coupling, plus boids spacing; expose temperature as a driver on a gamepad axis Render a 90-second piece where you drop temperature through the critical point live, then add PPF cloth to a 50-body subset for the hero shot Falsifier: if viewers describe it as a screensaver or particle effect rather than a crowd of people with intention, the herd score isn't carrying meaning

Prerequisites

Key tools and resources

Creators to follow

Prompt system — run this yourself

Everything below is pasteable into Claude Code, Claude, or any agent. Use the one-shot when you want the whole result in one run; use the step prompts when you want to steer. Every step carries its own audit, and the gap audit at the end makes the agent find and fill what it missed.

Mission. Produces a 90-second ensemble piece in Blender where 200 to 1,000 bodies perform one dancer's captured phrases, and the choice of phrase and facing for each body is governed by an Ising or Potts herding model whose temperature the artist drops live through the critical point on a gamepad axis. The pipeline runs from FreeMoCap capture, through HY-Motion text-to-motion variants, a verified Ising simulation, Blender instancing with boids spacing, to a PPF cloth hero shot. Done means a rendered piece, a validated physics core, and a blind viewer test answering the falsifier: do viewers see a crowd of people with intention, or a screensaver?

1 · Deeper, not bigger

Before running anything, make the task deeper rather than larger. Ask: what would a top professional do here that I am not asking for; what is the rawest material; which methods and tools; what, other than the model or me, proves it right. Then ask for exactly that.

Rewritten prompt — ask for exactly that:

Act as a technical director and choreographer's assistant. Build a 90-second Blender ensemble piece from one dancer. Step one: give me a FreeMoCap capture plan for 5 minutes of four distinct phrases with 2 to 3 webcams and ChArUco calibration, then a retarget into a Blender rig with foot-contact cleanup. Generate 20 variants per phrase with HY-Motion 1.0 and reject variants with foot sliding over 2 cm or joint angles out of range. Write a NumPy q-state Potts model on a 2D lattice with Glauber dynamics; verify the q=2 case reproduces Onsager magnetization within 5% and a susceptibility peak near T of 2.27 before using it. Map each lattice site to a body: the spin chooses which phrase and facing the body performs, synchronized to musical counts. Add boids separation with pressure falloff (1-d/R)^1.2 for spacing. In Blender, instance 200 to 1,000 bodies with Geometry Nodes using baked actions and per-instance time offsets; drive temperature from a gamepad axis via a modal operator and a custom property, recording the curve as keyframes. Render a 90-second piece dropping temperature through the critical point at a chosen musical moment, plus a boids-only control. Add ppf-contact-solver cloth to a 50-body subset for one hero shot. Then run a blind viewer test of 15 or more people on both renders and code their descriptions.

2 · One-shot prompt

ROLE AND STANCE
You are a technical director working with a choreographer. You will turn one dancer's captured motion into ensemble choreography for hundreds of bodies, using a herding model from statistical physics as the score. Source claims to test: Crystal Pite's ensembles move 'as a single organism'; FreeMoCap gives professional markerless capture from ordinary webcams; HY-Motion 1.0 generates skeleton animation from text across 200+ action types; the Ising market visualization shows random behavior at high temperature and sudden cluster formation as social pressure rises; ppf-contact-solver gives penetration-free cloth at scale with a Blender plugin; an LLM can write the whole Blender scene as Python.

CONTEXT
- Capture: FreeMoCap, 2 to 3 webcams, ChArUco board, Blender export add-on.
- Motion variants: HY-Motion 1.0 on Hugging Face.
- Score: q-state Potts (q equals number of phrases) or Ising for two facings, Glauber dynamics; boids spacing with pressure falloff (1-d/R)^1.2 from the shepherd simulation.
- Scene: Blender Python, Geometry Nodes instancing, drivers and scripted expressions in the Ian Hubert style, gamepad via a modal operator.
- Hero shot: ppf-contact-solver on a 50-body subset.
- Output: EEVEE drafts, Cycles finals, FFmpeg assembly.

HARD CONSTRAINTS
1. The physics core must reproduce known Ising results before any art direction.
2. The dancer consents in writing to capture and to use of their motion; credit them.
3. Every body's motion derives from the captured dancer or a filtered variant; no stock mocap.
4. Spin updates are quantized to musical counts so phrase switches land on beats.
5. A boids-only control render with identical bodies, phrases and camera is produced for the viewer test.

DELIVERABLES IN ORDER
1. capture/: protocol, calibration files, raw FreeMoCap output, retargeted actions per phrase.
2. variants/: HY-Motion variants with the filter report (foot slide, joint limits).
3. ising/: simulation code, validation plots (magnetization versus T against Onsager, susceptibility peak), and a function returning spin fields over time for a temperature curve.
4. blender/: scene-building script, Geometry Nodes setup, gamepad modal operator, recorded temperature curve.
5. renders/: the 90-second piece, the boids-only control, and the cloth hero shot, with frame-time logs.
6. viewer_test.md: protocol, raw responses, coding scheme, results.
7. verdict.md with GAPS.

BUILT-IN GAP ANALYSIS
Write GAPS: phrases with capture artifacts; HY-Motion variants rejected and why; whether the critical slowing near Tc made transitions too slow on screen; render compromises (body count, resolution); claims you could not test. Fill gaps you can, for example adjusting lattice size or update rate so the transition reads in under four bars, and note the change.

SELF-QUESTIONS
- Which moments in the music must the transition land on, and does the temperature curve put them there?
- Do four phrases read as distinct from a wide shot, or do they blur?
- Does the lattice neighborhood match how dancers actually cue each other (sight lines, nearest neighbors in front)?
- Is the transition near Tc too slow or too noisy to read as a decision?
- Would viewers see intention because of the dancer's motion alone, which the control render would also show?

STOP CONDITIONS
Stop and report if the Ising validation fails, if capture yields fewer than three usable phrases, or if 200 bodies cannot render under 2 minutes per frame in Cycles on the available GPU.

RETURN FORMAT
Falsifier verdict from the viewer test, file list, validation plot paths, render paths, GAPS, self-question answers.

3 · Step-by-step prompts

Step 1 — Ground truth: capture the dancer and study the reference

Goal: Get clean motion of four phrases and the structural grammar of the reference work.

Write a FreeMoCap capture protocol: 2 to 3 webcams at 1080p 30 fps or higher, placed about 90 to 120 degrees apart, a printed ChArUco board for calibration, even lighting, fitted clothing, and a 3 by 3 m floor area marked with tape. The dancer performs four distinct phrases of 8 counts each to a click track, three takes each, plus a T-pose. Process with FreeMoCap, export to Blender via its add-on, retarget to one rig, clean foot contacts, and save one action per phrase. In parallel, analyze a recording of Crystal Pite's The Seasons' Canon: log in a table every moment of unison, canon (ripple across the group), dissent (a subgroup moving differently), and floor-pattern change, with timestamps. This table becomes the target grammar for the score. Get written consent from the dancer.

Expect: capture/ with calibration, raw output, four cleaned actions, and a reference grammar table of 20 or more events.

Audit before moving on: Play each action on the rig and fail any with foot sliding over 2 cm or limb flips; recapture or clean. Confirm the consent note exists.

Step 2 — Generate and filter motion variants

Goal: Multiply the dancer's vocabulary without losing their signature.

For each phrase, write HY-Motion 1.0 prompts that describe the phrase in plain movement language (for example 'a person sweeps the right arm overhead, turns half left, and drops into a low lunge') and generate 20 variants. Retarget variants to the same rig. Filter automatically: foot sliding over 2 cm during contact, joint angles beyond human range, root drift over 0.5 m from the original's path, and duration mismatch over 10% against 8 counts at the piece's tempo. Score each surviving variant by dynamic time warping distance to the captured phrase on joint rotations, keeping variants close enough to read as the same dancer but different enough to vary. Write a filter report with counts kept and rejected per reason.

Expect: variants/ with kept actions and a filter report, ideally 5 to 10 kept variants per phrase.

Audit before moving on: View a grid of kept variants beside the original; fail the phrase if fewer than three variants survive, and rewrite prompts using the dancer's own movement qualities.

Step 3 — Build and validate the Ising or Potts score

Goal: A physics core proven correct before it is art-directed.

Write ising/ in NumPy: a q-state Potts model on an L by L periodic lattice with coupling J equals 1 and Glauber or heat-bath updates, with q equals 2 for Ising. Validate q equals 2 on L equals 128: sweep T from 1.5 to 3.5, equilibrate 5,000 sweeps, measure 5,000 more, and plot magnitude of magnetization against Onsager's exact result (1 - sinh(2/T)^-4)^(1/8) below Tc, plus susceptibility with its peak near T of 2.27. Then write score(temperature_curve, counts, L, q, seed) returning a spin field per musical count, with sweeps per count adjustable so transitions read on screen within two to four bars. Add an external field term so the choreographer can bias toward a chosen phrase at climaxes. Save validation plots.

Expect: ising/ code, validation plots, a table of measured versus exact magnetization, and the score function with tests.

Audit before moving on: Fail if magnetization deviates from Onsager by more than 5% at any T below 2.1 or if the susceptibility peak lies outside 2.2 to 2.35. Fix the update rule before continuing.

Step 4 — Blender scene, spacing and gamepad temperature

Goal: Put hundreds of bodies on stage, driven by the score and a live temperature knob.

Write a Blender Python script that builds the scene: a lattice of 200 to 1,000 points, each a body instanced with Geometry Nodes from a collection of baked phrase actions, with per-instance action choice and time offset read from the spin field and facing from a second Ising layer. Add boids separation in a pre-pass that offsets positions with force (1-d/R)^1.2 inside radius R so bodies never overlap. Write a modal operator that reads a gamepad axis (evdev or inputs on Linux, XInputReader on Windows) into a scene custom property 'temperature', and wire it with a driver into the score; record the live curve as keyframes so renders are reproducible. Light and stage camera to read floor pattern from above and unison from a low angle. Render an EEVEE draft at 720p.

Expect: blender/ with the build script, Geometry Nodes group, modal operator, and a draft render plus frame-time log.

Audit before moving on: Check no two bodies intersect in any frame by a bounding-sphere test. Fail if phrase switches land between counts or if draft frame time exceeds 10 s at 500 bodies.

Step 5 — Perform, render the piece, the control and the cloth hero shot

Goal: Produce final renders including the comparison arm.

Perform the temperature curve live on the gamepad against the music, aiming the critical drop at the moment you chose from the reference grammar table, and record it. Render the 90-second piece in Cycles at 1080p headless with blender -b -a, logging frame times. Render the control: same bodies, phrases, camera and music, but phrase choice from boids neighbor-averaging with no temperature or phase transition. Then take a 50-body subset for the hero shot, apply garments with ppf-contact-solver through its Blender plugin, simulate, and render 8 to 10 seconds. Assemble with FFmpeg, including the control as a separate file with identical length and audio. Log any penetration warnings from the solver. Keep the piece and control renders at identical resolution, samples and color management so viewers judge motion only.

Expect: renders/ with the piece, the control, the hero shot, frame-time logs and solver logs.

Audit before moving on: Fail if the control differs from the piece in anything but the choice rule. Inspect the hero shot for cloth penetration; fail on visible clipping.

Step 6 — Validate against oracles and write the gap report

Goal: Run the blind viewer test and answer the falsifier.

Show at least 15 viewers the piece and the control in random order, unlabeled. Ask each to describe each video in one or two sentences before any other question. Code responses blind into categories: crowd of people with intention, particles or screensaver, unclear. Report counts per video and a sign test or McNemar test on paired responses. Ask the dancer and, if possible, a choreographer to review the piece for whether the dancer's signature survives and whether the transitions read as choices. Write verdict.md: viewer results, reviewer notes, Ising validation summary, render stats, and the falsifier verdict. Then GAPS: each source claim (single-organism ensembles, FreeMoCap quality, HY-Motion variety, Ising herd transitions, PPF no clipping, LLM-written Blender scenes) marked verified, unverified or false with evidence.

Expect: viewer_test.md with raw responses and coding, verdict.md with the falsifier line and GAPS.

Audit before moving on: Have a second coder recode responses blind and report agreement; fail if under 80% and reconcile. Fail if viewers saw labels or the hypothesis before describing.

4 · Gap analysis and audit

Run this after the build. It makes the agent enumerate its own claims, mark what is verified, rank the gaps and fix the top ones.

You are an independent reviewer of a herd-physics ensemble choreography project. Open capture/, variants/ and the filter report, ising/ with validation plots, blender/, renders/ with logs, viewer_test.md and verdict.md. List every claim in verdict.md and mark each verified (cite plot, log, or response data), unverified (say what would prove it), or false. Check: Ising validation matches Onsager within 5% and the susceptibility peak is near 2.27; the score quantizes switches to counts; the control render differs only in the choice rule; no bodies intersect; cloth has no visible penetration; viewers saw videos unlabeled and described before being questioned; coding was blind with a second coder; the dancer's consent and credit exist. List missing pieces: phrases dropped, variants rejected, render compromises, untested body counts. Rank problems by whether they could flip the falsifier verdict, fix the top three now (for example recoding responses or re-rendering a mismatched control), and report before and after. End with what needs more viewers, a second dancer, or more GPU time to settle.

Questions the agent must ask itself before it says done:

5 · What proves it right

6 · Failure modes

7 · Agent direction (unattended runs)

Validate the Ising core first in plain NumPy; it is fast, and nothing else matters if it is wrong. Then build the Blender pipeline headless with blender -b and Python scripts, caching baked actions and spin fields to disk so renders are reproducible. Spawn one sub-agent for HY-Motion generation and filtering while another builds the Blender scene against placeholder actions; review the filter report and an intersection test yourself. Capture, the live gamepad performance and the viewer test involve people: write protocols and wait for the user's data, never invent viewer responses or the dancer's consent. Stop and ask if the GPU cannot meet the render budget at 200 bodies, if capture yields fewer than three usable phrases, or before showing the dancer's motion publicly. Return verdict.md as the entry point with capture/, variants/, ising/, blender/, renders/ and viewer_test.md.

Prompt system updated 2026-10-22 · all prompt systems · JSON

Browse more deep dives at DejaViewed — 846 entries across the full catalog.

Generated 2026-10-22 · curated by @6ab3 for DejaViewed