Text-to-motion models turn a sentence such as "run forward, then stop, then wave" into a 3D skeleton animation. They fail in a recognizable way: once a prompt asks for more than three or four actions, they drop, merge, or reorder them. Event-T2M documented this "action collapse" by splitting HumanML3D into subsets by event count (Event-T2M). Nobody knows whether the collapse depends on how many actions are requested or on how densely they are packed into the time available. Nobody knows whether a given model's failure curve could be predicted before running it on harder prompts. This note proposes to measure text-to-motion generators the way information theory measures noisy channels. Each model gets two numbers. The first is a capacity: the highest rate of distinguishable actions per second it can render reliably. The second is a dispersion: how quickly reliability degrades near that rate as the clip gets longer.
The idea, in two sentences: treat the requested event sequence as a message, the generated motion as the channel output, and an independent motion-to-text retrieval model as the decoder. Then test whether the classical laws for coding near capacity predict how sequence-level success depends on clip duration and event density, extrapolated from regimes fitted on short clips. If they do, a model's failure on any new prompt length becomes predictable from two fitted constants. Comparing the same constants measured on real motion-capture data separates what is an architectural limit from what is a limit of human movement itself.
Core Claim
In information theory, three regimes describe how the error of a code behaves as blocklength $n$ grows. In the second-order (central-limit) regime, rate sits within $1/\sqrt{n}$ of capacity and error tends to a constant. In the moderate-deviation regime, the gap to capacity shrinks more slowly than $1/\sqrt{n}$, and error vanishes sub-exponentially at a speed set entirely by the dispersion $V$. In the strong-converse regime, rate sits above capacity and success probability itself goes to zero. Tan showed for lossy source coding that one and the same dispersion governs both the central-limit and moderate-deviation behavior (Moderate-Deviations of Lossy Source Coding). Cheng et al. showed the analogous fact for classical-quantum channels via a Taylor expansion of the error exponent around capacity (Moderate Deviation Analysis for Classical-Quantum Channels).
The claim is that this cross-regime consistency is a falsifiable empirical signature, and that it can be checked on neural generators. Let $n$ be clip duration in seconds and $k$ the number of requested events, each drawn from a vocabulary of $M$ action descriptions. Define the rate $R = k\log_2 M / n$ in bits per second. Let $\epsilon(n,R)$ be the probability that the decoder fails to recover the full ordered event list. The normal approximation predicts
$$\epsilon(n,R) \approx Q\!\left(\frac{\sqrt{n}\,(C-R)}{\sqrt{V}}\right),$$
where $Q$ is the standard Gaussian tail function, $C$ is the model's capacity in bits per second, and $V$ is its dispersion. Its moderate-deviation counterpart, for $C-R=a_n$ with $a_n\to 0$ and $a_n\sqrt{n}\to\infty$, is
$$-\log \epsilon(n,R) \approx \frac{n\,a_n^2}{2V}.$$
Strong-converse behavior means $\epsilon \to 1$ for every fixed $R>C$ as $n$ grows, rather than plateauing at a constant below one.
We do not claim to prove a coding theorem for diffusion models; the generator is not an optimal code. The claim is narrower and testable. Fit $(C,V)$ on short clips near the threshold. If the two-parameter law then predicts failure on longer clips and on rates far below and above threshold better than generic curve fits, the "capacity and dispersion" of a generator are meaningful, transferable quantities. The real open question in the field is whether collapse is an artifact or an inevitable limit. That question becomes a measurable comparison: the generator's $C$ against the $C$ of real human motion.
Research Program
The first component is a controlled prompt suite that varies event count and duration independently. Event-T2M's HumanML3D-E stratifies by count alone, so density and count are confounded there. We build prompts by concatenating single-action descriptions from HumanML3D and KIT-ML (KIT Motion-Language Dataset) with "then". We fix a vocabulary of $M$ well-separated actions, chosen so that a retrieval model distinguishes them reliably on isolated real clips. We sweep $k$ from 1 to 8 and $n$ over the range current generators accept. Most HumanML3D-trained models cap length near ten seconds, so the reachable blocklengths are short. We say so plainly, and we use streaming generators such as FloodDiffusion and AMD to reach longer clips.
The second component is the decoder. A motion-text retrieval encoder of the TMR type, the same family Event-T2M uses for conditioning, scores sliding windows of the generated motion against all $M$ actions. A dynamic-programming alignment then recovers the most likely ordered sequence of $k$ events. Success means exact recovery of the ordered list. A softer per-event accuracy is reported alongside it. The evaluator could share biases with a TMR-conditioned generator, so we use a second decoder trained separately, an action classifier on BABEL segments, and report agreement between the two.
The third component is a source-side reference. BABEL provides frame-level action labels on real motion capture. We can therefore cut real clips containing $k$ labeled actions within $n$ seconds and run the same decoder on them. This yields an empirical capacity of human motion as seen by this decoder: the rate beyond which even real people's motion cannot be read back reliably, because actions physically need time. A generator whose fitted $C$ approaches the motion-capture $C$ is near the true limit. A large gap means the collapse is architectural. This is the cleanest available answer to the tension Gap 3 of the research file identifies.
The fourth component is fitting and extrapolation. For each generator we fit $(C,V)$ by maximum likelihood on the grid cells with $n \le n_0$ and rate within a band around the empirical 50% point. We then predict $\epsilon$ on held-out cells: longer durations, far-below-threshold rates (the moderate-deviation tail), and above-threshold rates (the strong-converse side). The strong-converse question is answered directly. It asks whether $\epsilon$ above threshold keeps rising toward one with duration at fixed rate, or saturates. Saturation would mean a model that degrades gracefully, the weak-converse picture that the file's synthesis glosses over.
A secondary, clearly optional thread tests Zhou and Tan's result on streaming coding. In their analysis a decoding delay $T$ multiplies the moderate-deviation constant (Achievable Moderate Deviations for Streaming Compression). In a streaming generator we vary the lookahead window and ask whether the fitted $1/V$ grows roughly linearly with it. Their theorem concerns optimal codes, so a mismatch would be informative rather than damning.
Baselines
The naive reference point is per-event independence. It assumes each event is rendered correctly with a fixed probability $p$, so sequence success is $p^k$ whatever the duration. This is the "do nothing" model behind reading HumanML3D-E as a function of event count. It is fitted on the same short-clip cells with the same likelihood. The capacity law beats it if, at fixed $k$, failure depends strongly on duration in the direction the law predicts, and if held-out log-likelihood improves. If independence fits as well, the density framing is wrong and we report that.
The strongest generic competitor is a flexible predictor with the same inputs: a logistic regression on $k$, $n$, $k/n$ and their interactions, plus a Gaussian-process classifier over $(k,n)$. Both get the same training cells and a hyperparameter budget matched to the two-parameter fit via held-out cross-validation on the training region only. Inside the training region they will likely fit as well or better. The test is extrapolation to longer clips and to the tails, where the law imposes a specific $\sqrt{n}$ interaction and these models do not. Beating them means lower held-out negative log-likelihood and calibration error on the extrapolation cells, with the gap reported along with its uncertainty. No reported result was found on Lacuna for any predictive model of text-to-motion failure, so there is no prior number to beat here.
The generators themselves serve as subjects, and the strongest one anchors the comparison. On the four-or-more-event subset of HumanML3D-E, Event-T2M reports FID 0.265, against 0.418 for MoMask and 1.077 for AttT2M (Event-T2M). On full HumanML3D, Fg-T2M reports Top-3 R-precision 0.783 (Fg-T2M). We profile Event-T2M, MoMask, AttT2M, MotionDiffuse (MotionDiffuse) and GraphMotion (jpthu17/GraphMotion). Each uses its released weights, the same prompt suite, the same seeds per cell, and the same decoder.
The ablation that isolates the key ingredient fixes $k$ and varies only $n$. If the collapse is about density, success at $k=4$ should recover as the clip lengthens, at the rate the fitted $V$ implies. If it is about count alone, it should not.
Why not just…?
Why not just plot success against event count, as HumanML3D-E already does? That plot cannot distinguish "too many actions" from "too little time per action", because standard HumanML3D clips with more events also tend to have them packed differently. Our duration sweep at fixed count is the experiment that separates the two. If the count-only plot turns out to be sufficient, the per-event independence baseline will win and we will say so.
Why not just fit any smooth curve to success over count and duration? This is the real threat. Inside the measured grid a flexible classifier will match or beat a two-parameter law. The information-theoretic form earns its place only if it extrapolates: to longer clips from streaming models, to the deep tail where moderate-deviation scaling applies, and above threshold where strong-converse behavior is tested. The held-out-region comparison against logistic and Gaussian-process fits is designed to settle this. A loss there would reduce the contribution to the motion-capture-relative capacity comparison, which still stands on its own.
Why not just decompose prompts with an LLM and move on? Event-T2M shows decomposition helps a lot. Its FID on the four-or-more-event subset is 0.265 against MoMask's 0.418. Yet the file notes that it still breaks beyond four to seven events and keeps physical artifacts such as foot skating. Decomposition is a design choice whose effect our measurement can quantify, as a change in $C$, in $V$, or in both. It does not tell you where the remaining limit lies or whether real motion shares it.
Why not just rely on spatial guidance instead of text? OmniControl shows that guidance without a realism prior stretches limbs beyond human reach, and that it costs about 121 seconds per sequence (OmniControl). More to the point, spatial guidance answers a different question. Here the message is semantic, and the limit we measure is how many distinguishable actions language plus a generator can deliver per second.
Proposed Experiments
The first decisive check is the fixed-count duration sweep on all five generators and on BABEL motion capture, with $k$ from 2 to 6. Per-event independence predicts flat curves in $n$. The capacity law predicts rising success with a slope set by $V$.
The second check is cross-regime prediction. We fit $(C,V)$ on clips up to five seconds near the 50% point. We then score predictions of $\epsilon$ on six-to-ten-second clips, and on streaming-generated clips up to thirty seconds, separately in the low-rate tail and the above-capacity region. We report held-out log-likelihood and calibration against the generic baselines. As a consistency test in the spirit of Tan's result, we also check whether $V$ fitted from the tail alone agrees with $V$ fitted from the threshold band within bootstrap intervals.
The third check is the strong-converse question. At three rates above each model's fitted $C$, does sequence failure keep rising toward one as duration grows, or plateau? We also report the empirical exponent if it rises.
The fourth check is the architecture-versus-physics comparison. We put each generator's $C$ next to the BABEL motion-capture $C$, under both decoders. A generator within the bootstrap interval of motion capture is effectively at the human limit for this action vocabulary. A large gap is a quantified, architecture-specific deficit that future conditioning methods can be scored against.
Two outcomes would undercut the proposal, and both are worth reporting. The first is that the generic fits extrapolate as well as the law. The second is that the two decoders disagree on $C$ by more than the model-to-model differences.
References
- Event-T2M: Event-level Conditioning for Complex Text-to-Motion Synthesis
- Moderate-deviations of lossy source coding for discrete and Gaussian sources
- Moderate Deviation Analysis for Classical-Quantum Channels and Quantum Hypothesis Testing
- The KIT Motion-Language Dataset
- FloodDiffusion: Tailored Diffusion Forcing for Streaming Motion Generation
- AMD: Autoregressive Motion Diffusion
- Achievable Moderate Deviations Asymptotics for Streaming Compression of Correlated Sources
- Fg-T2M: Fine-Grained Text-Driven Human Motion Generation via Diffusion Model
- MotionDiffuse: Text-Driven Human Motion Generation with Diffusion Model
- OmniControl: Control Any Joint at Any Time for Human Motion Generation