- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper extends the streaming motion diffusion model FloodDiffusion (FD1) with three components: Partial Attention (causal history, bidirectional active window), enabling KV-cached inference and shared-history training; a fixed FK-induced quadratic loss, justified by a Bregman criterion for preserving the conditional-mean velocity; and root-path conditioning. Experiment results are given to support the effectiveness of the proposed method.
Critical strengths
-
Practical efficiency for streaming generation. Partial Attention enables KV caching and shared-history training with little loss in quality, reducing training cost and, for long histories, inference cost.
-
A clear argument about geometric supervision. Supported by a controlled 6-D rotation experiment (App. B.7), the paper shows that nonlinear FK losses can bias the diffusion target and reduce diversity, while the proposed loss avoids this.
Critical weaknesses
- The paper argues for streaming generation, but does not evaluate it where streaming matters most: long, continuous sequences.
The target setting is not tested. The motivation is interactive characters, where motion is generated continuously over open-ended horizons with changing prompts. However, all quantitative quality results are on individual benchmark clips (HumanML3D clips are at most about 10 s), and prompt switching is shown only qualitatively (Figs. 1 and 5). FD1 also reported results on BABEL, whose sequences contain multiple consecutive actions; FD2 does not.
Efficiency and quality are validated in different regimes. The 11.29× speedup is highlighted at 4,500 history frames (2.5 minutes at 30 fps); at 90 and 450 frames, it is 1.08× and 1.56× (Table 3). Quality, in contrast, is measured only on short clips. The abstract presents both together, but no experiment shows that the method is both fast and high-quality on long sequences. Long-sequence evaluation is important for streaming. Training uses ground-truth history (teacher forcing), and Table 2 uses 300-frame sequences. The Limitations section notes that finalized errors persist in history and that long generation requires extrapolation. I appreciate this acknowledgment. Since sustained long-horizon generation is central to streaming tasks, I encourage the authors to evaluate it directly.
2.The contribution of the two main technical components beyond FD1 and existing designs is not clearly established.
Partial Attention. FD1 already streams with fixed finalized frames. Partial Attention mainly removes FD1’s redundant recomputation of history states. Its attention pattern matches that of Rolling Forcing (cited by the paper), and the stated differences lie in training. The paper should clarify what is specific to its setting.
FK-induced loss. It shares the population optimum of MSE (Theorem B.6), so any benefit must come from finite-capacity effects, which are not analyzed. The gains are small: confidence intervals overlap on HumanML3D, and SEED reports none. The controlled rotation experiment (App. B.7) shows the loss performs on par with MSE, but not better.
Questions for authors
Questions that could change my assessment:
-
Please report quality on long, continuous multi-prompt streams, e.g., FID, foot skate, and jitter as a function of elapsed time, and transition quality (e.g., on BABEL). What sequence length is used in training, and is context noise augmentation used?
-
Please add ablations for the FK loss: (a) simple per-joint or per-channel-group reweighting, a natural baseline, compared with the full matrix with cross-joint terms; (b) a nonlinear FK loss on real data. Have the authors considered an input-dependent W(x), which the paper notes also preserves the conditional mean (App. B.4)? What motivates the dataset-averaged W?
-
The path-control comparisons are not matched. On HumanML3D, FD2-Path and GMD impose the root trajectory directly, while OmniControl and DART approach it via guidance or optimization. FD2-Path also receives heading, and on SEED, Position is measured at different points (CoM vs. pelvis). Could the authors compare under matched inputs and metrics?
Clarification: 4. Is the efficiency baseline FD1 with untruncated history? How does FD2 compare with FD1 using a truncated or sliding history window?
Thank you for recognizing Partial Attention's training and long-history inference efficiency and our geometric supervision analysis. We address W1, W2, and Q1–Q4 below:
1. Long-horizon streaming, training setup, and Rolling Forcing (W1, W2a & Q1)
For Q1: models are trained on official HumanML3D and SEED splits following FD1's protocol (200-frame max clips on HumanML3D, 300 on SEED, standard diffusion schedules, no extra motion or history-noise augmentation). Our supplementary webpage's
Long-form generationsection shows 48–60 s (1,440–1,800 frames) continuous streaming videos across multiple prompt and path transitions. Here we focus on streaming diffusion training/inference efficiency and exact path control in native motion space; under clean-history training, drift can still accumulate over open-ended rollouts, and adding context-noise or rollout training is left for future work.For W2(a), our work differs from Rolling Forcing in two ways:
2. FK loss ablations and dataset-averaged W (W2b & Q2)
Section 3.3 proves target-first Bregman divergences are the necessary and sufficient losses preserving the diffusion conditional-mean velocity, showing why online nonlinear FK losses bias samples toward high-density modes and motivating our offline FK quadratic loss, where W reweights channels (with cross-joint terms) by world-space pose impact.
For Q2(a), we ablated Full vs. Diagonal FK (per-channel/per-joint weights without off-diagonal cross-terms) on HumanML3D: FID is 0.053 for Full FK vs. 0.067 for Diagonal FK (0.058 for MSE without FK). Dropping off-diagonal terms hurts performance by ignoring cross-joint kinematic couplings. On 138-D SEED, the FK matrix cuts mean path error from 0.18 to 0.11 cm and P95 from 0.55 to 0.32 cm.
For Q2 on W(x) vs. dataset-averaged W: since xt is noisy, the geometry-consistent weight W(xt)=E[W(x0)∣xt] cannot be evaluated cheaply online, whereas the offline dataset-averaged W avoids online FK and Jacobians during training. On HumanML3D, W from 3,000 random training clips is nearly identical to the full 23,384-clip matrix (differing by 0.5005% in relative Frobenius norm, 0.3757% in relative operator norm on the normalized 64-D position block, and 0.5367% in directional response across 12,941 motion directions on 100 test clips), showing W is stable across the dataset. Adaptive mean-consistent losses are left for future work.
3. Path-control conventions and CoM vs. Pelvis scale on SEED (Q3)
In Table 4 (HumanML3D), FD2-Path verifies path conditioning preserves motion quality (0.050 FID vs. 0.053 without path), while path-tracking comparisons are on SEED (Table 5). For streaming generation, we condition on relative CoM increments to avoid gait-cycle pelvis sway and random global-translation augmentation; clamping pelvis coordinates directly in our explicit representation would make pelvis error trivially zero. Across all 142,220 SEED clips, per-frame planar speeds of pelvis and CoM have nearly identical scale: mean 0.30916 vs. 0.29690 m/s, and std 0.47978 vs. 0.47603 m/s. Scaling FD2-Path's CoM error by std(pelvis)/std(CoM)=1.00787 gives 0.1109 / 0.3225 cm (mean / P95), far below ARDY (1.60 / 3.91 cm) and Kimodo (2.37 / 7.44 cm).
4. Untruncated vs. sliding-window history in efficiency comparisons (Q4)
The efficiency baseline uses untruncated history by default. Sliding-window truncation applies to both: if FD1 truncates history (e.g., 100, 500, or 1,000 frames), FD2 truncates its KV cache identically. On an RTX 4090 (30 steps, with CFG and VAE for FD1), steady-state FPS across window lengths is:
Whether untruncated (17.53× speedup to 333.03 FPS at 4,501 frames, 5.00× faster than FD1) or truncated at 101–1,001 frames, cached FD2 is faster end-to-end than FD1+VAE.