Aug 2026· Computer graphics forum (Print)· 0 citations· 29 references
TL;DR
This work introduces a novel motion prior based on the sparsity of high‐order temporal derivatives, serving as a kinematic proxy for impulsive force generation and achieves linear complexity, enabling efficient processing of long sequences.
Abstract
Conventional human motion denoising methods often prioritize smoothness at the cost of dynamic fidelity. This aggressive smoothing tends to obliterate sharp transients, thereby diminishing the perceived forcefulness (or ‘Sense of Force') of the motion—a critical attribute for realistic animation. Inspired by physiological findings on RFD, we introduce a novel motion prior based on the sparsity of high‐order temporal derivatives. Specifically, we formulate denoising as an optimization problem that encourages jerk sparsity, serving as a kinematic proxy for impulsive force generation. Our solver achieves linear complexity, enabling efficient processing of long sequences. Extensive experiments on synthetic and real MoCap data demonstrate that our method effectively eliminates noise while preserving high‐frequency acceleration details more effectively compared to state‐of‐the‐art filtering and learning‐based approaches. User studies with professional animators confirm that our results are significantly preferred for their preserved dynamic impact. Furthermore, we showcase the versatility of our approach in applications such as forcefulness restoration and controllable editing. Our code will be made publicly available at
https://github.com/ChambinLee/sparse‐jerk‐mocap‐denoising
.
This work proposes FlashMo, a frequency-aware sparse motion diffusion model that prunes low-frequency tokens to enhance efficiency without custom kernel design, and introduces MotionSiT, a scalable diffusion transformer based on a joint-temporal factorized interpolant with Lie group geodesics over SO(3) manifolds, enabling principled generation of joint rotations.
Zeyu Zhang, Yiran Wang, Danning Li et al.· Advances in Neural Informati...· 11 citations
Recent advances in generative motion synthesis have enabled realtime autoregressive generation of diverse and realistic character animations conditioned on user inputs, as demonstrated by models such as the Conditional Autoregressive Motion Diffusion Model (CAMDM). However, real‐world applications (e.g., computer games) often demand faster‐than‐realtime performance for large numbers of characters. We introduce F3AMD (Fast FiLM‐conditioned Fourier Autoregressive Motion Diffusion), a framework that achieves an order of magnitude speedup over state‐of‐the‐art systems for multi‐character animation on both GPUs and CPUs while maintaining high motion quality. Our key insight is that autoregressive motion diffusion is primarily bottlenecked by architectural and sampling inefficiencies. To address this, F3AMD employs Fourier Neural Operators (FNOs) as encoder‐decoder modules, substitutes Transformer backbones with FNO blocks, replaces condition concatenation with lightweight Feature‐wise Linear Modulation (FiLM), and adopts a variance‐exploding noise schedule with a deterministic sampler. This design enables a substantially lower‐dimensional latent space, facilitates learning in both the spectral and temporal domains, and significantly improves sample efficiency. We conduct systematic ablations of key design factors, including latent dimension, backbone type, diffusion window length, and number of denoising steps. Our recommended configuration, F3AMD‐FNO‐96, achieves 20x speedup over the baseline CAMDM model, while maintaining comparable motion quality.
Calvin Qiao, Benjamin MacAdam, Mohammadarsh Khokhar et al.· Computer graphics forum (Pri...· 0 citations
Motion warping is a core technique in character animation that enables the adaptation of existing motion data to novel spatio-temporal constraints. Conventional motion warping methods often rely on heuristic modifications that can violate physical consistency or introduce visual artifacts. More recent learning-based editing approaches improve realism, but many of them encode motion into tightly entangled latent space, which makes them struggle to balance editing flexibility and content preservation. To address this, we propose a novel deep motion warping framework that explicitly disentangles the motion structure from global and stylistic attributes for intuitive motion editing. Our key insight is to leverage learned phase features as a continuous and robust representation of the underlying structure, and explicitly disentangle motion into root velocity, phase, and learned latent variables using a phase-conditioned diffusion autoencoder. This design supports a wide range of editing operations, including root motion warping, motion exaggeration, time warping, and style transfer by directly manipulating decoupled components, without requiring paired training data. Extensive experiments demonstrate that our approach enables high-level, flexible motion editing while strictly preserving the structural consistency and physical plausibility of the source motion
Bowen Zheng, Linjun Wu, Xinwei Jiang et al.· International Conference on...· 0 citations
This paper presents a motion editing framework for smooth motion stitching based on Laplacian optimization in Rodrigues vector space. By representing joint rotations as continuous Rodrigues vectors, motion stitching is formulated as a temporal Laplacian optimization problem, enabling smooth transitions between motion segments while preserving characteristic temporal variations of reference motions. The proposed approach supports both intra-category replacement and cross-category motion stitching without relying on learning-based models or complex manual tuning, and is computationally efficient for interactive editing. Through a series of stitching experiments and comparisons with linear interpolation, we demonstrate that Laplacian editing produces stable and visually coherent transitions under a wide range of motion differences. Furthermore, an analysis of rotational continuity clarifies that rotation-axis inversions are rare in real motion data and explains why numerical instabilities observed in synthetic axis-flipping scenarios do not arise in practical motion stitching. These results highlight the importance of rotational representation in stabilizing temporal optimization and suggest that the proposed framework is well suited not only for animation authoring but also for motion analysis and future extensions incorporating perceptual or physiological cues.
Ryosuke Higasayama, Hideki Todo, Jongseong Gwak· 0 citations
Diffusion-based super-resolution (SR) achieves strong perceptual quality but requires costly iterative denoising. Existing one-step distillation methods reduce inference time but depend on expensive pretrained teachers, whereas CTMSR avoids distillation through PF-ODE consistency training yet does not explicitly model the restoration dynamics from low-resolution (LR) inputs to high-resolution (HR) images. We propose MeanSR, a one-step perceptual SR method that learns an LR-conditioned average velocity field to directly capture the finite-time transition from degraded or noisy inputs to plausible HR outputs. We further reformulate distribution trajectory matching for average-velocity generation and introduce a Stage-Aware Temporal Sampling strategy to improve trajectory learning. Experiments on synthetic and real-world benchmarks show that MeanSR outperforms CTMSR on CLIPIQA, MUSIQ, and MANIQA while substantially reducing FLOPs and inference latency. MeanSR also reconstructs sharper structures and more realistic textures with fewer perceptual artifacts.
Axi Niu, Jiawei Kou, Kang Zhang et al.· 0 citations
We present PhysDiff-VTON, a diffusion-based framework for image-based virtual try-on that systematically addresses the dual challenges of garment deformation modeling and high-frequency detail preservation. The core innovation lies in integrating physics-inspired mechanisms into the diffusion process: a pose-guided deformable warping module simulates fabric dynamics by predicting spatial offsets conditioned on human pose semantics, while wavelet-enhanced feature decomposition explicitly preserves texture fidelity through frequency-aware attention. Further enhancing generation quality, a novel sampling strategy optimizes the de-noising trajectory via least action principles, enforcing temporal coherence, spatial smoothness, and multi-scale structural consistency. Comprehensive evaluations across multiple datasets demonstrate significant improvements in both geometric plausibility and perceptual quality compared to existing approaches. The framework establishes a new paradigm for synthesizing photorealistic try-on images that adhere to physical constraints while maintaining intricate garment details, advancing the practical applicability of diffusion models in fashion technology.
Shibin Mei, Bingbing Ni· Advances in Neural Informati...· 1 citation