A Practical Recipe for Structured Pruning of MotionGPT: Dependency-Graph Pruning and FFN Channel Reduction
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Human motion generation from natural language has become an important research challenge in computer vision and natural language processing, with applications in animation, robotics, and immersive VR/AR systems. Recent advances such as MotionGPT treat motion as a language sequence, enabling unified text-to-motion generation, captioning, and prediction across datasets like HumanML3D. However, the large size of such models limits their deployment in real-world settings. This paper presents a practical and reproducible pruning framework for MotionGPT, a state-of-the-art multimodal generative model that treats human motion as a foreign language. Two complementary structured pruning strategies are proposed: (1) dependency-graph pruning, a global, structure-aware method implemented with Torch-Pruning that removes channels consistently across residual connections, tied projections, and multi-head attention; and (2) FFN channel pruning, a local procedure tailored to T5-style DenseReluDense feed-forward blocks, which shrinks the dominant intermediate dimension. Experiments on HumanML3D show that the pruned model reduces parameters by ~14% while largely preserving semantic metrics such as Matching Score and R-Precision. Although FID increases (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.22 \rightarrow 0.86$</tex>) and trajectory errors (ADE/FDE) worsen moderately, motion diversity remains high. Compared to prior models such as MDM, the pruned MotionGPT maintains competitive or superior alignment scores while being more efficient. These findings demonstrate that structured pruning provides a viable path to making large motion-language models lighter and more accessible.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it