- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes Timo, which applies the MMDiT architecture with flow matching to text-to-motion generation. Text and motion tokens are processed jointly by a fully shared transformer. Since directly applying MMDiT leads to jerky motion, the authors add rotational kinematics and geometric losses, and train the model with a two-stage coarse-to-fine curriculum. On a six-dataset benchmark built by the authors, Timo outperforms existing methods such as Kimodo and HY-Motion.
Critical strengths
- The rotational kinematics loss is simple and effective, and it clearly improves motion smoothness as shown in Table 2.
- The paper is well-written and easy to follow.
Critical weaknesses
-
Limited technical contribution and novelty. The core components, namely MMDiT-style joint text–motion attention and flow matching, have already been explored in prior text-to-motion methods such as HY-Motion and MotionHiFlow [1], so the paper offers limited new insight to the field. Moreover, the abstract's first sentence claims that cross-attention limits text comprehension and that bidirectional attention addresses this. However, Table 2 shows that DDIM+RoPE and MMDiT+flow achieve the same Text-Motion Match score (80.8), indicating that bidirectional attention brings no measurable gain in text alignment. Instead, the improvement in Text-Motion Match comes mainly from mirroring and the body-rotation loss.
-
The main comparison in Table 1 is unconvincing. The evaluation protocol is constructed by the authors, and several important details are missing.
- The test clips are held out from the same six corpora used in Timo's second training stage, so Timo is evaluated in distribution, whereas the compared methods were trained on different datasets with different motion representations. The data split is also created by the authors. Notably, the DDIM baseline in Table 2, trained only on these public corpora, already outperforms all released methods in Table 1 by a large margin (e.g., Naturalness 75.6 vs. at most 44.2), suggesting that most of the reported gain comes from data alignment rather than from the proposed method.
- The paper does not explain how the evaluator is trained or which data it is trained on.
- The paper does not explain how the baselines, which use different motion representations, are converted and evaluated.
- All results come from a single generation seed, with no confidence intervals over multiple runs.
- No experiments are conducted on commonly used benchmarks such as HumanML3D and KIT-ML under the standard protocol.
-
The ablation studies are heavily confounded, leaving the individual effect of each contribution unclear.
- Table 2 presents cumulative ablations without leave-one-out experiments. The first step simultaneously changes the architecture and the generation process, and possibly also the prediction target and the sampler.
- L_geo, the geometric loss with weights as high as 500/500/1000, is never ablated, despite being presented as part of the contributions.
- Table 3 compares the full model against a single-stage MMDiT+flow model, so the comparison is confounded by differences in data and training curriculum.
- In Table 4, the two-stage configuration simultaneously introduces in-house data and increases the number of training steps.
-
Insufficient evaluation of physical plausibility. The authors claim that the generated motions are coordinated and temporally coherent, but do not report standard physical metrics such as foot skating ratio, floating, and ground penetration. Moreover, the proposed method performs considerably worse than Kimodo on limb interpenetration (4.93 vs. 1.42, Table S4).
[1]: Heng Li, Xiaotong Lin, Ling-An Zeng, et al. MotionHiFlow: Text-to-Motion via Hierarchical Flow Matching. In CVPR, 2026.
Questions for authors
See Weaknesses