Chensheng Dai

Chensheng Dai, Shengjun Zhang, Yifan Li et al.

Video generation models based on Diffusion Transformers (DiTs) have achieved remarkable performance in video synthesis, yet they suffer from high inference latency and computational costs due to the quadratic complexity of 3D attention. Existing acceleration methods primarily reduce computational complexity within each individual denoising steps through techniques such as sparse attention and KV-caching. However, they rigidly adhere to the inherent constraint of the standard diffusion pipeline: every frame in the target video sequence must be subjected to a complete, dense denoising process across all diffusion timesteps. We observe that due to the corresponding contents and motions among adjacent frames, when keyframes with critical semantic transitions are anchored, the intermediate states of others often follow more predictable trajectories, which indicates that such uniform, dense denoising process is inherently redundant for natural video data. To this end, we introduce \textbf{RhymeFlow}, a training-free framework that decouples the denoising trajectories of different frames. Specifically, we first identify a sparse set of pivotal key frames that dominate the latent semantic evolution. Then, only these keyframes undergo dense, step-by-step denoising to ensure structural integrity, while non-keyframes progressively skip denoising steps to minimize computational cost. Since skipped intermediate states of non-keyframes break the temporal coherence in keyframe denoising steps, leading to visual degradation, we further introduce a latent trajectory projection module, which enables keyframes to interact with a complete and temporally consistent sequence representation. Extensive experiments on current DiT-based video generation models demonstrate our method outperforms existing baselines with higher inference speed and better visual quality.

Shengjun Zhang, Zhang Zhang, Simin Huang et al.

Recent advancements in video-based world models have demonstrated an unprecedented ability to synthesize high-fidelity visual sequences. However, a fundamental gap persists between visually plausible video generation and the functional requirements of a world model, particularly in maintaining a stable and reasonable internal state over extended temporal horizons. While existing benchmarks primarily emphasize visual quality, motion coherence, and text-video alignment, they largely overlook memory, the core capability of a world model to preserve consistency across long-term horizons and complex interactions. To address this gap, we present \textbf{MBench}, a comprehensive benchmark dedicated to quantifying and evaluating the memory capability of video world models. We systematically decompose the memory capability of video world models into three hierarchical and complementary core dimensions: entity consistency, environment consistency, and causal consistency, which are further refined into 12 quantifiable sub-dimensions for comprehensive characterization of long-term memory. Our benchmark is built upon rigorously curated real-captured long videos, and evaluated by rule-based quantitative matrices and VLM to enable objective and comprehensive consistency assessment. Extensive evaluations of mainstream state-of-the-art video world models reveal critical systemic limitations of existing methods in long-term state retention, providing a standardized benchmark and clear research direction to advance the field.

Shengjun Zhang, Zhang Zhang, Chensheng Dai et al.

Recent reinforcement learning has enhanced the flow matching models on human preference alignment. While stochastic sampling enables the exploration of denoising directions, existing methods which optimize over multiple denoising steps suffer from sparse and ambiguous reward signals. We observe that the high entropy steps enable more efficient and effective exploration while the low entropy steps result in undistinguished roll-outs. To this end, we propose E-GRPO, an entropy aware Group Relative Policy Optimization to increase the entropy of SDE sampling steps. Since the integration of stochastic differential equations suffer from ambiguous reward signals due to stochasticity from multiple steps, we specifically merge consecutive low entropy steps to formulate one high entropy step for SDE sampling, while applying ODE sampling on other steps. Building upon this, we introduce multi-step group normalized advantage, which computes group-relative advantages within samples sharing the same consolidated SDE denoising step. Experimental results on different reward settings have demonstrated the effectiveness of our methods.

Chensheng Dai, Shengjun Zhang, Min Chen et al.

3D Gaussian Splatting (3DGS) has demonstrated impressive performance in 3D scene reconstruction. Beyond novel view synthesis, it shows great potential for multi-view surface reconstruction. Existing methods employ optimization-based reconstruction pipelines that achieve precise and complete surface extractions. However, these approaches typically require dense input views and high time consumption for per-scene optimization. To address these limitations, we propose SurfelSplat, a feed-forward framework that generates efficient and generalizable pixel-aligned Gaussian surfel representations from sparse-view images. We observe that conventional feed-forward structures struggle to recover accurate geometric attributes of Gaussian surfels because the spatial frequency of pixel-aligned primitives exceeds Nyquist sampling rates. Therefore, we propose a cross-view feature aggregation module based on the Nyquist sampling theorem. Specifically, we first adapt the geometric forms of Gaussian surfels with spatial sampling rate-guided low-pass filters. We then project the filtered surfels across all input views to obtain cross-view feature correlations. By processing these correlations through a specially designed feature fusion network, we can finally regress Gaussian surfels with precise geometry. Extensive experiments on DTU reconstruction benchmarks demonstrate that our model achieves comparable results with state-of-the-art methods, and predict Gaussian surfels within 1 second, offering a 100x speedup without costly per-scene training.

Yifan Liu, Shengjun Zhang, Chensheng Dai et al.

Modeling animatable human avatars from videos is a long-standing and challenging problem. While conventional methods require per-instance optimization, recent feed-forward methods have been proposed to generate 3D Gaussians with a learnable network. However, these methods predict Gaussians for each frame independently, without fully capturing the relations of Gaussians from different timestamps. To address this, we propose Human Gaussian Graph to model the connection between predicted Gaussians and human SMPL mesh, so that we can leverage information from all frames to recover an animatable human representation. Specifically, the Human Gaussian Graph contains dual layers where Gaussians are the first layer nodes and mesh vertices serve as the second layer nodes. Based on this structure, we further propose the intra-node operation to aggregate various Gaussians connected to one mesh vertex, and inter-node operation to support message passing among mesh node neighbors. Experimental results on novel view synthesis and novel pose animation demonstrate the efficiency and generalization of our method.