acceptodds
Withdrawn from ICLR 2027

FloodDiffusion 2: Efficient and Path Controllable Streaming Motion Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We present FloodDiffusion 2 (FD2), an efficient and controllable framework that builds upon FloodDiffusion (FD1), a state-of-the-art streaming motion generation model. While FD1 produces plausible motion, it suffers from low efficiency and limited controllability, as its attention design requires repeated computation over the entire history, and it lacks precise trajectory control for real-world applications. To address these limitations and improve generation quality, FD2 introduces three advances. First, Partial Attention makes finalized history representations independent of the active window, enabling KV-cached inference and shared-history packing for efficient training. Second, we establish a necessary-and-sufficient Bregman criterion for regression losses to preserve diffusion's conditional-mean velocity field. This criterion guides an FK-induced quadratic loss that incorporates motion geometry without online FK evaluation. Third, FD2 introduces precise path conditioning to control the character's root trajectory while preserving natural body motion. Experiments show that FD2 reduces training computation by 4.6 and accelerates denoising by 11.29, reaching 2.303 ms per update on long sequences. Alongside these efficiency gains, FD2 improves motion quality over FD1 and achieves state-of-the-art FID scores among streaming methods, with 0.048 on SEED and 0.053 on HumanML3D.

Primary area:
generative models
Keywords:
diffusion, human motion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bYQJOpenReview ↗
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
  1. 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.

  2. 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
  1. 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:

  1. 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?

  2. 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?

  3. 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?

1 comment

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 generation section 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:

  • Target-distribution preservation proof: Under frame-staggered diffusion, we prove Partial Attention preserves full bidirectional attention's exact target velocity field (Prop. A.3), whereas fully causal attention biases the optimum (Prop. A.4). Rolling Forcing lacks this proof.
  • Single-GPU from-scratch pre-training: Rolling Forcing uses distillation or self-rollouts from a pretrained teacher rather than training from scratch. With 8-window shared-history packing (Prop. A.6), isolated noisy windows reuse one clean reference encoding under direct velocity supervision, enabling single-GPU from-scratch pre-training and KV-cached inference.

2. FK loss ablations and dataset-averaged WW (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 WW 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)W(x) vs. dataset-averaged WW: since xtx_t is noisy, the geometry-consistent weight W(xt)=E[W(x0)∣xt]W(x_t) = \mathbb{E}[W(x_0) \mid x_t] cannot be evaluated cheaply online, whereas the offline dataset-averaged WW avoids online FK and Jacobians during training. On HumanML3D, WW 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 WW 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\text{std}(\text{pelvis})/\text{std}(\text{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:

HistoryFD1+VAEFD2 w/o cacheFD2 w/ cache
1551.08665.10841.90
101538.36470.78802.27
501323.25191.21719.68
1001293.91103.15607.39
450166.6019.00333.03

Whether untruncated (17.53×17.53\times speedup to 333.03 FPS at 4,501 frames, 5.00×5.00\times faster than FD1) or truncated at 101–1,001 frames, cached FD2 is faster end-to-end than FD1+VAE.

Reviewer_UJN4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FloodDiffusion 2 (FD2), designed to enhance the efficiency, generation quality, and trajectory controllability of streaming human motion synthesis. Building upon the existing FloodDiffusion framework, this method incorporates three key improvements: it reduces redundant computations during both training and inference through "Partial Attention" and a shared historical computation mechanism; it introduces a secondary loss function based on human kinematics, providing geometric supervision while preserving the target velocity field of the diffusion model; and it integrates trajectory-based conditional control, enabling generated motions to follow specified paths and adapt to time-varying text prompts.

The authors conducted experiments on the HumanML3D and SEED datasets. The results demonstrate that FD2 reduces the computational cost of training the denoiser by approximately 4.6 times. With a history of 4,500 frames, the use of a KV cache accelerates denoising updates by 11.29 times, resulting in a latency of 2.303 ms per update. In terms of generation quality, FD2 achieves an FID score of 0.053 on HumanML3D and 0.048 on SEED, outperforming other streaming motion synthesis methods compared in the study.

Critical strengths
  1. The efficiency improvements offer clear practical value. To address the issue of redundant computation of historical frames during streaming generation, the paper introduces "Partial Attention" and "Shared-History Packing," with theoretical analysis confirming their feasibility. Experiments demonstrate that these designs significantly reduce computational costs during training and accelerate inference for long sequences, all while largely maintaining generation quality. This contribution is quite robust.

  2. The design of the geometric loss is grounded in sound theoretical reasoning. The authors analyze the potential drawbacks of directly employing a non-linear Forward Kinematics (FK) loss and propose an alternative loss function that preserves the target velocity field. The theoretical derivations are comprehensive, and experimental results across two datasets show improvements in generation quality, further validating this design.

Critical weaknesses
  1. The specific design of the FK Loss lacks key ablation studies. While the paper demonstrates that a fixed, positive-definite quadratic loss preserves the conditional mean, this property is not unique to the proposed "FK Matrix." The current experiments lack comparisons against a "Diagonal FK" approach or simpler channel-weighting methods; consequently, it is difficult to ascertain whether the performance gains truly stem from the kinematic structure. Given that the improvements in FID are relatively modest, this omission undermines the persuasiveness of the paper’s contributions.

  2. Cross-method comparisons for path control lack a unified evaluation framework. Although path control is a primary contribution of the paper, the methods compared in Table 5 utilize different reference points for calculating "Position Error": FD2-Path uses the center of mass (CoM) of the generated body, whereas Kimodo and ARDY use the pelvis. Even though the same test paths are employed, the variation in control-point definitions renders a direct, absolute comparison of position errors invalid. While the results demonstrate that FD2-Path achieves high trajectory accuracy within its own specific evaluation setup, they are insufficient to determine whether it would still outperform other methods under a unified control-point definition. This weakens the persuasiveness of the experiments concerning path control.

Questions for authors

Q1. Regarding the necessity of the FK-induced quadratic loss: The paper demonstrates that a fixed, positive-definite quadratic loss preserves the target conditional mean; however, this is not a property unique to the FK matrix. Could the authors compare the performance of the full FK matrix, a diagonal FK matrix, and a simple channel-weighted loss to clarify whether the off-diagonal kinematic couplings provide any additional benefit? Specifically, given that the 263-D geometric response matrix for HumanML3D contains only 67 non-zero response channels, how can the respective contributions of channel reweighting and the full kinematic structure be disentangled?

Q2. Regarding the fairness of the path-control evaluation: In Table 5, while all methods are evaluated using the same 917 paths, FD2-Path calculates the position error based on the center of mass (CoM) of the generated body, whereas Kimodo and ARDY use the pelvis. Could the authors clarify the impact of this difference in control points on the results and provide a comparison of trajectory errors using a unified reference point? This would help establish the actual precision advantage of FD2-Path over existing methods.

2 comments

Thank you for recognizing the practical efficiency of Partial Attention and shared-history packing, as well as the theoretical grounding of our geometric loss. We address your two points (W1/Q1 and W2/Q2) below:

1. Role of the FK quadratic loss and diagonal ablation (Weakness 1 & Question 1)

We prove that target-first Bregman divergences (up to a target-only term) are the necessary and sufficient class of regression losses that preserve the diffusion conditional-mean velocity target, whereas general nonlinear losses (such as online nonlinear FK loss) shift the target and pull samples toward high-density regions. Within this valid class, we use a fixed positive-definite quadratic form.

Regarding why this differs in practice from standard MSE despite sharing the same population optimum: for a finite-capacity network regressing multi-channel velocities, the matrix acts as a cross-channel weighting over outputs, assigning larger gradient weights to joint dimensions that have a stronger effect on world-space pose geometry. While any fixed positive-definite matrix preserves the conditional mean, we parameterize it offline using FK sensitivity so no online FK is needed during training. Its effect is especially clear on the rotationally accurate 138-D SEED representation, where adding the FK matrix reduces FD2-Path's mean tracking error from 0.18 to 0.11 cm, P95 from 0.55 to 0.32 cm, and foot skate from 4.07 to 3.24 cm/s.

To directly answer Q1, we ran an ablation on HumanML3D comparing the full FK matrix against its diagonal-only version (Diagonal FK, which keeps per-channel reweighting but drops off-diagonal cross-terms): Full FK achieves 0.053 FID, whereas Diagonal FK gets 0.067 FID (compared with 0.058 for standard MSE without FK). Purely diagonal reweighting performs worse than the full matrix because it ignores kinematic couplings across joints along the skeletal chain. Exploring better designs within the mean-consistent loss family is left for future work.

2. Fairness of Table 5 path evaluation and Pelvis vs. CoM speed scale (Weakness 2 & Question 2)

Prior path-conditioned models define different control points, and Table 5 evaluates each method under its published control convention. In our explicit native representation, if we conditioned directly on pelvis coordinates and clamped those channels during sampling, the pelvis tracking error would trivially be zero by construction. Instead, we condition on relative CoM increments—which avoids gait-cycle pelvis sway and fits streaming training—and recompute whole-body CoM from the generated articulated pose to measure tracking error.

To check how much the motion scales of the pelvis and CoM differ, we computed per-frame planar speed statistics over the entire prepared SEED dataset (142,220 clips, 31,049,862 adjacent-frame intervals at 30 fps):

MetricPelvisCoMDifference (Pelvis - CoM)
Mean speed (m/s)0.309160.29690+0.01226 (~4.0% difference)
Speed std (m/s)0.479780.47603+0.00375 (ratio 1.00787, <0.8% difference)

The per-frame speed mean and standard deviation of the two landmarks are nearly identical, with std(pelvis)/std(CoM)=1.00787\text{std}(\text{pelvis})/\text{std}(\text{CoM}) = 1.00787. Scaling FD2-Path's CoM tracking error by this std ratio gives:

MethodMean (cm)P95 (cm)
FD2-Path (CoM error scaled by std ratio)0.11090.3225
ARDY (pelvis error)1.603.91
Kimodo (pelvis error)2.377.44

Even after this scale adjustment, FD2-Path's mean error remains 14.4×14.4\times lower than ARDY and 21.4×21.4\times lower than Kimodo (and P95 remains 12.1×12.1\times and 23.1×23.1\times lower).

Thanks to the authors for their response and clarifications. I will keep my score.

Reviewer_kbzGOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

FloodDiffusion 2 extends FD1's frame-staggered text-to-motion diffusion with cache-compatible Partial Attention, shared-history training, a fixed FK-induced quadratic regression loss, and dense root-path conditioning. HumanML3D and SEED experiments report reduced denoiser computation, improved FID, and accurate trajectory tracking; static examples illustrate action and speed changes.

Critical strengths
  • Partial Attention addresses a concrete FD1 bottleneck
  • The demo video demonstrates a good streaming motion generation quality
Critical weaknesses

Overall, I don't think the technical contribution efforts of this paper are sufficient for qualified publication in ICLR, as the framework is basically retained from a previous work without novel technical contribution. The claimed contribution, i.e. partial attention, bregman characterization, root path control, are either well known standard practice or existing engineering fixes. Although the streaming motion quality is good, it's doubtful whether it mainly comes from the performance capability of FD1 or the contribution of this paper.

Some other detailed comments:

  • The benchmark rankings provide limited evidence for a decisive quality advantage over competing streaming approaches.
  • The motivating interactive streaming setting is weakly validated: Tables 4-5 measure clip quality and tracking, while the long-history experiment measures only computation (Table 3).
  • Comparative quality, control fidelity, and response after condition changes remain unestablished: Figures 1, 5, and 6 show FD2 alone, and user feedback confirms that the demos lack baseline comparisons.
  • No visual demo or figure for the generated results of baseline comparison.
  • The reported latency covers one denoiser branch without CFG, text encoding, or motion decoding (Table 3; Appendix D.3), leaving the practical real-time benefit incompletely quantified.
Questions for authors

see weaknesses

1 comment

Thank you for recognizing the efficiency bottleneck addressed by Partial Attention and the quality of our streaming demos. We address your comments below:

1. Technical novelty and relation to FD1 (Main Weakness & Comment 1)

Our goal is to remove the training and inference bottlenecks of streaming motion diffusion and enable exact path control, rather than chasing a large jump on single-clip unconstrained metrics. Keeping the base streaming schedule aligned with FD1, we remove the VAE and train directly in native motion space at a fraction of the compute cost, matching or improving upon FD1 (on HumanML3D, FD2 without FK gets 0.058 FID vs. FD1's 0.057; full FD2 reaches 0.053 on HumanML3D and 0.048 on SEED). This shows that our VAE-free streaming formulation is effective on its own.

None of the three components is standard practice in motion diffusion:

  • Partial Attention and single-GPU from-scratch training: We prove that under frame-staggered diffusion, Partial Attention preserves the exact same target distribution as full bidirectional attention, whereas fully causal attention biases the target (Prop. A.3–A.4). Video streaming methods such as Rolling Forcing neither prove this property nor train from scratch (relying on distillation or rollouts from a pretrained teacher). Combining Partial Attention with 8-window shared-history packing (K=8K=8) lets us train streaming diffusion from scratch on 1 GPU for a window budget that previously required 8 GPUs (4.6×4.6\times fewer denoiser FLOPs at 512 windows) and supports KV-cached inference.
  • Bregman characterization and FK quadratic loss: Proving that target-first Bregman divergences form the necessary and sufficient class of losses preserving the diffusion conditional-mean velocity—and showing that online nonlinear FK losses bias sampling toward high-density modes—has not been established in prior motion diffusion work. Guided by this result, we build a cross-joint quadratic weight matrix from offline FK sensitivity without online FK during training. On HumanML3D, Full FK (0.053 FID) outperforms Diagonal FK without cross-terms (0.067 FID); on 138-D SEED, it cuts mean path error from 0.18 to 0.11 cm.
  • VAE-free native diffusion and relative CoM path control: FD1 compresses motion into a coupled VAE latent space where root paths cannot be clamped directly. Partial Attention's training efficiency lets us diffuse directly on raw motion features and enforce exact path control by clamping clean trajectory channels. Relative CoM increments also filter out gait-cycle pelvis sway and avoid random global-translation augmentation during training.

2. Long-horizon streaming evaluation and error accumulation (Comment 2)

Our models are trained on the official HumanML3D and SEED splits following FD1's protocol (clips up to 200 frames on HumanML3D and 300 frames on SEED, standard diffusion schedules, no extra history-noise augmentation). The Long-form generation section of our supplementary webpage shows 48–60 s (1,440–1,800 frames) multi-prompt and path-controlled streaming rollouts. Under teacher-forced clean-history training, open-ended rollouts over even longer horizons can still accumulate drift; incorporating rollout or context-noise training is left for future work.

3. Side-by-side visual comparisons with baselines (Comments 3 & 4)

We added a Baseline comparisons section to our supplementary webpage with synchronized side-by-side videos comparing FD2 (SEED and HumanML3D) against ARDY and Kimodo under identical paths, speed profiles, and text prompts. These comparisons show that FD2 tracks the target trajectories closely and transitions cleanly across prompts, whereas the baselines exhibit noticeable path drift or posture artifacts under speed changes and action transitions.

4. End-to-end inference speed on RTX 4090 (Comment 5)

We measured end-to-end throughput (FPS) on an RTX 4090, including two-branch CFG and motion decoding (with VAE decoding for FD1). Both use 30 denoising steps and retain full history at matched frame counts:

History framesFD1 + VAEFD2 without cacheFD2 with cache
1551.08665.10841.90
101538.36470.78802.27
501323.25191.21719.68
1001293.91103.15607.39
450166.6019.00333.03

With KV caching, FD2 is faster end-to-end than FD1 (with VAE) at all history lengths. At 4,501 frames, KV caching speeds up FD2 by 17.53×17.53\times (to 333.03 FPS), running 5.00×5.00\times as fast as FD1 (66.60 FPS).

Comments on the paper

1 comment

We thank all the reviewers for their detailed comments and questions. We want to emphasize that the goal of this work is not to chase large gains on unconstrained benchmarks, but to make streaming motion diffusion much faster to train and run while enabling exact trajectory control at state-of-the-art quality. Our main points are summarized below:

1. Distribution-preserving Partial Attention and single-GPU from-scratch training: We prove that under frame-staggered diffusion, Partial Attention preserves the exact same target velocity field as full bidirectional attention, whereas fully causal attention introduces bias (Prop. A.3–A.4). Because finalized history states no longer depend on the active denoising window, inference supports standard KV caching and training can share clean history encodings across multiple isolated noisy windows (Prop. A.6). With 8-window packing (K=8K=8), a window budget that previously required 8 GPUs fits on a single GPU (cutting denoiser forward/backward FLOPs by 4.6×4.6\times at a matched 512-window budget). Unlike video streaming methods such as Rolling Forcing, which rely on distillation or self-rollouts from a pretrained teacher without a target-preservation proof, our formulation trains streaming diffusion directly from scratch. On an RTX 4090 (30 denoising steps, 4,501 history frames), end-to-end throughput is 66.60 FPS for FD1 (with VAE), 19.00 FPS for uncached FD2, and 333.03 FPS for cached FD2—a 17.53×17.53\times speedup from KV caching and 5.00×5.00\times faster than FD1.

2. VAE-free native diffusion and relative CoM path control: Latent diffusion with a VAE entangles trajectory and body posture in latent space, making exact trajectory clamping impractical. The compute savings from Partial Attention allow us to drop the VAE and diffuse directly in native motion space, matching or improving upon FD1 with minimal architectural changes (on HumanML3D, FD2 without FK gets 0.058 FID vs. FD1's 0.057; full FD2 reaches 0.053 on HumanML3D and 0.048 on SEED). Operating in native space lets us clamp clean trajectory channels directly, and representing paths as relative CoM increments removes gait-cycle pelvis sway while fitting streaming training without random global-translation augmentation.

3. Bregman characterization of diffusion losses and FK-weighted quadratic loss: We prove that target-first Bregman divergences are the necessary and sufficient class of regression losses that preserve the diffusion conditional-mean velocity field (Thm. 3.3), showing why standard online nonlinear FK losses shift the target and pull samples toward high-density modes. Within this class, we use a fixed positive-definite quadratic loss, which acts as a cross-channel weighting over output dimensions (giving larger weights to joints with stronger influence on world-space geometry), parameterized offline by FK sensitivity so no online FK is needed during training. On 138-D SEED, this reduces mean path error from 0.18 to 0.11 cm (P95 from 0.55 to 0.32 cm); on HumanML3D, the full FK matrix (0.053 FID) outperforms its diagonal-only counterpart without cross-joint terms (0.067 FID).

4. Mitigating drift over ultra-long horizons and designing adaptive mean-consistent losses are left for future work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.