Compute-Optimal Scaling Laws for Human Motion Generation

ListenMotion Gets Its Chinchilla Moment9:45

Show notes

Cartwheel fits the first Chinchilla-style compute-optimal scaling laws for 3D human motion generation, training hundreds of models on a purpose-built ~12,000-hour corpus. Both autoregressive and flow-matching frameworks land near N* proportional to C to the one-half — matching language — and a controlled repetition study finds flow matching tolerates roughly 30× more reuse of the same data before held-out loss degrades. We dig into the data-curation funnel, the extrapolation test, and what the AR-versus-flow repetition gap means for video model teams. **Sources** - Compute-Optimal Scaling Laws for Human Motion Generation (Param Hanji, Joakim Olsson, Aalok Patwardhan, Katie Lowen, Andrew Carr)

Subscribe to the channel: https://podcast.lumalabs.link/27a25f51ea81d41a8253f46c6133a35e/feed.xml

TL;DR

Cartwheel fits the first Chinchilla-style compute-optimal scaling laws for 3D human motion generation, on a purpose-built ~12,000-hour motion corpus that is ~40× the size of the only prior motion scaling-law study and ~4× the next-largest motion dataset. Hundreds of models are trained across compute budgets in both autoregressive and flow-matching frameworks; both fit independently and both land near $N^* \propto C^{1/2}$ , matching Chinchilla’s language exponent. The paper also runs a controlled data-repetition study at matched size (49.2M) and matched compute ( $\sim 10^{18}$ FLOPs), reporting that flow matching tolerates ~30× more epochs of the same data than autoregression before held-out loss degrades ~2%. The framing claim is that human motion is now a “predictable and scalable frontier” — no longer a data-constrained modality — and joins text/image/audio/video as the fifth modality with a known scaling law.

Key claims

Both autoregressive and flow-matching motion generators fit independently, and both give compute-optimal exponents near $N^* \propto C^{1/2}$ (i.e. the Chinchilla exponent for language); losses are in different units so only the exponents are directly comparable [Fig. 1].
Uncertainties on the fitted exponents are reported spanning all three Chinchilla estimators (parametric loss fit / IsoFLOP / FLOP-derivative) and the analysis choices each leaves open — the exponent agreement is not an artifact of any one methodology [contributions §1].
The extrapolation test is done properly: fits are frozen on cheaper budgets, then compared against held-out larger budgets that the fit never saw; both frameworks’ frontiers extrapolate cleanly [Fig. 6].
The corpus is built, not scraped: quality gates reduce a ~423k-hour raw video catalog through prefiltering (→ ~91k h), visual gates (→ 24.3k h), “Comic” 3D reconstruction with physical QA (→ 12.1k h, plus mocap + public corpora), and a final safety screen to 11,969 hours / 5.56M clips — a ~35× total reduction [Fig. 3].
At matched size (49.2M) and matched compute (~ $10^{18}$ FLOPs), the “+2% held-out loss” repetition wall is reached at ~16 epochs for the autoregressive framework and ~498 epochs for flow matching — flow matching tolerates ~30× more reuse of the same data [Fig. 7].
Explicit framing: “the choice of framework determines how long a fixed corpus lasts” — for a data-constrained modality this is a first-class design lever, not a downstream detail [contributions §3].
The dataset is ~40× larger than the corpus behind the one prior motion scaling-law study and ~4× the next-largest motion corpus [Fig. 1 caption].

Method

The setup is deliberately Chinchilla-shaped. For each of two generative frameworks — autoregressive and flow matching — Cartwheel trains a family of models at several sizes for each of several compute budgets; the × marker on each isoFLOP curve is the best model size for that budget, and the slope of the optima across budgets is the compute-optimal exponent. The two frameworks are fit independently; the paper’s headline agreement (both near $C^{1/2}$ ) is empirical, not enforced by a shared parameterization.

The data pipeline is the load-bearing methodological contribution alongside the scaling fit. A ~423k-hour raw video catalog is passed through a five-stage funnel: prefiltering (source metadata, licensing, coarse content) → visual gates (person presence, occlusion, resolution) → 3D human reconstruction via Cartwheel’s “Comic” system with physical QA (foot-slide, penetration, kinematic plausibility) → mocap + public-corpus join → safety screen. The final 11,969 hours / 5.56M clips is the corpus all scaling runs are fit on. The repetition study runs at fixed model size (49.2M params) and fixed compute ( $\sim 10^{18}$ FLOPs), sweeping the number of epochs and observing when held-out loss crosses +2% relative to the best-epoch checkpoint.

Results

Scaling exponents. Autoregressive and flow-matching motion models both fit near $N^* \propto C^{1/2}$ . The frozen small-compute fits extrapolate correctly to held-out large-compute runs in both frameworks [Fig. 6].
Data-repetition wall. Autoregression: ~16 epochs to +2% loss. Flow matching: ~498 epochs to +2% loss. Ratio ~30× at matched size and matched compute [Fig. 7].
Corpus. Final size 11,969 hours / 5.56M clips after ~35× quality-gate reduction from a 423k-hour raw catalog [Fig. 3].
No downstream benchmark numbers (e.g. text-to-motion instruction-following, kinematic quality) are reported in the GitHub README at filing time; the paper’s contribution is the scaling curves themselves and the corpus-construction methodology, not a competing generation model.

Why it’s interesting

This is the first motion-side Chinchilla-style scaling law filed on the wiki, and it complements Scaling Behavior Foundation Model for Humanoid Robots‘s recent whole-body-control scaling result and EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data‘s log-linear egocentric-hours fit — three independent embodied-substrate scaling laws now agree that the data-axis dependency looks Kaplan/Chinchilla-like across mocap-adjacent modalities. The 40× dataset ratio versus HY-Motion 1.0: Scaling Flow Matching Models for Text-To-Motion Generation (HY-Motion’s ~3,000-hour corpus) is the numerically load-bearing claim: HY-Motion was the wiki’s prior “biggest motion training run” datapoint and reported that a 1B-parameter flow-matching DiT works, but it did not fit a scaling law; Cartwheel’s paper is the actual scaling fit that HY-Motion left open. The ~30× flow-matching-over-AR repetition-tolerance gap is also a new datapoint in the AR-vs-flow debate the wiki has been tracking on the video side (e.g. Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation on the AR side, Self-Forcing++: Towards Minute-Scale High-Quality Video Generation on the self-forcing side) — a domain-independent AR-vs-flow property (data-repetition tolerance) is a rare cross-framework comparison at controlled compute.