Siyu Zhu

Jiaxin Shi, Xidong Zhang, Fucai Zhu et al.

We propose a 3D-thinking-guided co-training framework that enables vision-language-action (VLA) models to perform 3D spatial reasoning implicitly during action prediction. Our core insight is that 3D geometry perception and 3D spatial reasoning are distinct capabilities that can be disentangled and injected at different feature hierarchies. During training, three tightly coupled components work in concert primarily within the latent space: (1) To gain geometric priors, a latent 3D geometry perception module aligns intermediate visual features with a 3D foundation model, acquiring low-level geometric cues without architectural modifications to the VLM backbone. (2) Complementing this, an online 3D reasoning distillation module mitigates the prompt-induced reasoning gap via a shared reasoning anchor token. During 3D VLM co-training, this anchor is emitted as the first output token to robustly encode spatial priors. During VLA training, it serves as an input token inserted between the task and action instructions, transferring high-level spatial thinking from explicit teacher reasoning prompts to student action prompts without chain-of-thought text generation. (3) These disentangled geometric and reasoning features are then united by a spatially augmented action integration, which jointly injects them into the action-query tokens as hierarchical spatial conditions to prevent action shortcuts. At deployment, our method retains only its lightweight adapters to perform implicit 3D reasoning, discarding the 3D foundation model and the teacher branch used for supervision. Consequently, it operates purely on 2D images without 3D sensors, external models, or explicit text generation while preventing catastrophic forgetting of the pretrained VLM, achieving state-of-the-art performance on LIBERO, LIBERO-PLUS, SimplerEnv, and real-world manipulation tasks.

Kaihui Cheng, Zhiqiang Cai, Wenkai Xiang et al.

Generative emulators of protein dynamics produce plausible trajectories at a fraction of the cost of molecular dynamics, but they inherit their training distribution and tend to revisit known states rather than reach rare ones under long-horizon extrapolation. Inspired by classical enhanced sampling, we introduce an implicit, history-dependent bias in the generative space of a pretrained emulator. Specifically, a history-aware score estimator augments the frozen emulator with a distance-weighted bias that steers reverse-time sampling away from previously generated structures, regularized by an environment-support term. To preserve structural validity at long horizons, a score-based refinement step re-projects drifted samples onto the data manifold using the frozen emulator. Our experiments demonstrate that the method (i) raises diversity by $35\%$ on DynamicPDB-80; (ii) on $12$ zero-shot Fast-Folding proteins, the learned bias alone reaches the unbiased emulator's coverage up to ${\sim}15\times$ faster, and pairing it with refinement reaches the coverage up to ${\sim}37\times$ faster while covering ${\sim}3\times$ as many low-energy states. Code will be released soon.

Haisheng Fu, Runyu Yang, Feng Ding et al.

State-of-the-art learned image compression (LIC) schemes are increasingly based on hybrid CNN-transformer architectures. To further improve rate-distortion performance, we introduce channel-wise wavelet transforms into both the transformer and entropy-coding components. First, we propose a channel-wise wavelet-domain transformer attention (ChWDTA) mechanism. ChWDTA keeps the efficient windowed spatial self-attention used in modern LIC backbones, but computes the Q/K/V projections on channel-wise wavelet-transformed features before mapping the attention output back with the inverse transform. The resulting Channel-wise Wavelet-Domain Transformer Block (ChWDTB) therefore preserves the spatial tokenization pattern of windowed attention while sparsifying the channel covariance seen by the attention projections. Second, in the entropy-coding stage, we introduce a channel-wise wavelet packet (ChWP) decomposition that produces four equal-sized subbands, which better fit channel-wise slice-based autoregressive entropy modeling. When each channel-wise subband is divided into two slices, we use eight slices for entropy coding. With this configuration, the proposed scheme obtains BD-rate reductions of -17.82%, -19.15%, and -22.56% on the Kodak, CLIC Professional Validation, and Tecnick test sets, respectively. Even when each channel-wise subband is coded as a single slice, the scheme still retains most of the coding gains with lower complexity. The results confirm the advantage of introducing wavelet transform in CNN-transformer-based LIC schemes.

Zhenghao Zhang, Junchao Liao, Menghao Li et al.

Recent advancements in Diffusion Transformer (DiT) have demonstrated remarkable proficiency in producing high-quality video content. Nonetheless, the potential of transformer-based diffusion models for effectively generating videos with controllable motion remains an area of limited exploration. This paper introduces Tora, the first trajectory-oriented DiT framework that concurrently integrates textual, visual, and trajectory conditions, thereby enabling scalable video generation with effective motion guidance. Specifically, Tora consists of a Trajectory Extractor (TE), a Spatial-Temporal DiT, and a Motion-guidance Fuser (MGF). The TE encodes arbitrary trajectories into hierarchical spacetime motion patches with a 3D motion compression network. The MGF integrates the motion patches into the DiT blocks to generate consistent videos that accurately follow designated trajectories. Our design aligns seamlessly with DiT's scalability, allowing precise control of video content's dynamics with diverse durations, aspect ratios, and resolutions. Extensive experiments demonstrate that Tora excels in achieving high motion fidelity compared to the foundational DiT model, while also accurately simulating the complex movements of the physical world. Code is made available at https://github.com/alibaba/Tora .

Weihao Yuan, Xiaodong Gu, Zuozhuo Dai et al.

Estimating the accurate depth from a single image is challenging since it is inherently ambiguous and ill-posed. While recent works design increasingly complicated and powerful networks to directly regress the depth map, we take the path of CRFs optimization. Due to the expensive computation, CRFs are usually performed between neighborhoods rather than the whole graph. To leverage the potential of fully-connected CRFs, we split the input into windows and perform the FC-CRFs optimization within each window, which reduces the computation complexity and makes FC-CRFs feasible. To better capture the relationships between nodes in the graph, we exploit the multi-head attention mechanism to compute a multi-head potential function, which is fed to the networks to output an optimized depth map. Then we build a bottom-up-top-down structure, where this neural window FC-CRFs module serves as the decoder, and a vision transformer serves as the encoder. The experiments demonstrate that our method significantly improves the performance across all metrics on both the KITTI and NYUv2 datasets, compared to previous methods. Furthermore, the proposed method can be directly applied to panorama images and outperforms all previous panorama methods on the MatterPort3D dataset. Project page: https://weihaosky.github.io/newcrfs.

Weijia Dou, Hui Li, Jiahao Cui et al.

Streaming video generation models typically rely on temporal-centric memory, which organizes historical context as raw frames, chunk segments, or unclustered tokens. This organization frequently leads to identity drift and semantic inconsistency when entities exit the frame or during interactive prompt transitions. To address these limitations, we propose SlotMemory, an object-centric Key-Value memory mechanism for streaming video diffusion. Our approach shifts the memory abstraction from "when" an event occurred to "what" is being represented by decomposing the transformer's key-value manifold into discrete, reusable semantic slots. By utilizing these slots as routing addresses to index and store high-fidelity key-value tokens, we enable entity-level persistence and prompt-aware retrieval across long horizons. Evaluated on 60-second interactive narratives using the Wan2.1-T2V-1.3B backbone, SlotMemory achieves a state-of-the-art quality score of 81.61 and a 22.8 percent relative improvement in dynamic consistency over the strongest existing streaming baseline. Our results demonstrate that structured semantic representation, rather than raw temporal capacity, is the essential primitive for persistent long-form video synthesis. Our codes and checkpoints are available at https://tj12323.github.io/SlotMemory/.

Zhu Yu, Zhengyi Zhao, Runmin Zhang et al.

This work presents the Large Depth Completion Model (LDCM), a simple, effective, and robust framework for single-view metric depth estimation with sparse observations. Without relying on complex architectural designs, LDCM generates metric-accurate dense depth maps using a transformer. It outperforms existing approaches across diverse datasets and sparse observations. We achieve this from two key perspectives: (1) leveraging existing monocular foundation models to improve the quality of sparse depth inputs, and (2) reformulating training objectives to better capture geometric structure and metric consistency. Specifically, a Poisson-based depth initialization strategy is first introduced to generate a uniform coarse dense depth map from diverse sparse observations, providing a strong structural prior for the network. Regarding the training objective, we replace the conventional depth head with a point map head that regresses per-pixel 3D coordinates in camera space, enabling the model to directly learn the underlying 3D scene structure instead of performing pixel-wise depth map restoration. Moreover, this design eliminates the need for camera intrinsic parameters, allowing LDCM to naturally produce metric-scaled 3D point maps. Extensive experiments demonstrate that LDCM consistently outperforms state-of-the-art methods across multiple benchmarks and varying sparsity levels in both depth completion and point map estimation, showcasing its effectiveness and strong generalization to unseen data distributions.

Zhu Yu, Jingnan Gao, Runmin Zhang et al.

This work presents ViGeo, a feed-forward foundation model for recovering spatially dense and temporally consistent geometry from video sequences. Built upon a plain transformer architecture without task-specific architectural modifications, ViGeo supports streaming, full-sequence, and long-video inference within a unified model. The key design is dynamic chunking attention, which exposes the model to both bidirectional and causal temporal contexts during training and allows it to adapt its attention pattern at test time without retraining. To improve supervision quality, we further introduce a completion-based data refinement framework. This framework trains a video depth completion teacher that conditions on sparse and noisy annotations and exploits video/multi-view context to produce dense, temporally coherent, and geometrically reliable training targets. Beyond depth and point maps, ViGeo also predicts surface normals within the same framework. Trained solely on public datasets, ViGeo achieves state-of-the-art performance across online, offline, and long-video depth estimation, surface normal estimation, and video point map estimation.

Weihao Yuan, Xiaodong Gu, Heng Li et al.

Monocular scene reconstruction from posed images is challenging due to the complexity of a large environment. Recent volumetric methods learn to directly predict the TSDF volume and have demonstrated promising results in this task. However, most methods focus on how to extract and fuse the 2D features to a 3D feature volume, but none of them improve the way how the 3D volume is aggregated. In this work, we propose an SDF transformer network, which replaces the role of 3D CNN for better 3D feature aggregation. To reduce the explosive computation complexity of the 3D multi-head attention, we propose a sparse window attention module, where the attention is only calculated between the non-empty voxels within a local window. Then a top-down-bottom-up 3D attention network is built for 3D feature aggregation, where a dilate-attention structure is proposed to prevent geometry degeneration, and two global modules are employed to equip with global receptive fields. The experiments on multiple datasets show that this 3D transformer network generates a more accurate and complete reconstruction, which outperforms previous methods by a large margin. Remarkably, the mesh accuracy is improved by 41.8%, and the mesh completeness is improved by 25.3% on the ScanNet dataset. Project page: https://weihaosky.github.io/sdfformer.

Yingxin Lai, Yafei Zhou, Fucai Zhu et al.

While rule-based reinforcement learning has recently catalyzed explicit reasoning in multimodal models, tactile reasoning remains largely underexplored. Existing tactile-language models primarily rely on supervised or contrastive objectives, which limits their capacity to ground predictions in physical evidence or rectify misleading visual priors. Tactile reasoning introduces two modality-specific challenges: the ordinal nature of physical attributes (e.g., hardness, roughness) and the cross-sensor distribution shifts inherent in optical tactile hardware. In this work, we introduce TouchReason-1M, a large-scale multimodal dataset comprising over 1M synchronized tactile pairs across four distinct sensors, and TouchReason-Bench, a rigorous framework for evaluating tactile perception and visual-tactile conflict resolution. Building upon these, we propose Touch-R1, a tactile reasoning MLLM based on Qwen2.5-VL-7B. Touch-R1 is trained via a tactile-grounded GRPO objective that combines ordinal-aware accuracy, cross-sensor physical consistency, structured-format control, and an input-side tactile grounding objective. Specifically, the tactile-use reward assigns credit only when authentic tactile inputs yield superior correctness relative to counterfactual controls where the tactile stream is removed, shuffled, or noise-masked. On TouchReason-Bench, Touch-R1-7B outperforms Octopi-13B by 18.4\% and GPT-4o by 24.7\% on average. Its structured reasoning traces reveal emergent behaviors of probing, comparison, and revision, demonstrating that R1-style reasoning can be effectively grounded in physical contact.

Kayhan Behdin, Ayan Acharya, Aman Gupta et al.

With the rising popularity of Large Language Models (LLMs), there has been an increasing interest in compression techniques that enable their efficient deployment. This study focuses on the Post-Training Quantization (PTQ) of LLMs. Drawing from recent advances, our work introduces QuantEase, a layer-wise quantization framework where individual layers undergo separate quantization. The problem is framed as a discrete-structured non-convex optimization, prompting the development of algorithms rooted in Coordinate Descent (CD) techniques. These CD-based methods provide high-quality solutions to the complex non-convex layer-wise quantization problems. Notably, our CD-based approach features straightforward updates, relying solely on matrix and vector operations, circumventing the need for matrix inversion or decomposition. We also explore an outlier-aware variant of our approach, allowing for retaining significant weights (outliers) with complete precision. Our proposal attains state-of-the-art performance in terms of perplexity and zero-shot accuracy in empirical evaluations across various LLMs and datasets, with relative improvements up to 15% over methods such as GPTQ. Leveraging careful linear algebra optimizations, QuantEase can quantize models like Falcon-180B on a single NVIDIA A100 GPU in $\sim$3 hours. Particularly noteworthy is our outlier-aware algorithm's capability to achieve near or sub-3-bit quantization of LLMs with an acceptable drop in accuracy, obviating the need for non-uniform quantization or grouping techniques, improving upon methods such as SpQR by up to two times in terms of perplexity.

Jiayang Liu, Siyu Zhu, Siyuan Liang et al.

Deep neural networks (DNNs) are susceptible to adversarial examples, which introduce imperceptible perturbations to benign samples, deceiving DNN predictions. While some attack methods excel in the white-box setting, they often struggle in the black-box scenario, particularly against models fortified with defense mechanisms. Various techniques have emerged to enhance the transferability of adversarial attacks for the black-box scenario. Among these, input transformation-based attacks have demonstrated their effectiveness. In this paper, we explore the potential of leveraging data generated by Stable Diffusion to boost adversarial transferability. This approach draws inspiration from recent research that harnessed synthetic data generated by Stable Diffusion to enhance model generalization. In particular, previous work has highlighted the correlation between the presence of both real and synthetic data and improved model generalization. Building upon this insight, we introduce a novel attack method called Stable Diffusion Attack Method (SDAM), which incorporates samples generated by Stable Diffusion to augment input images. Furthermore, we propose a fast variant of SDAM to reduce computational overhead while preserving high adversarial transferability. Our extensive experimental results demonstrate that our method outperforms state-of-the-art baselines by a substantial margin. Moreover, our approach is compatible with existing transfer-based attacks to further enhance adversarial transferability.

Zuozhuo Dai, Zhenghao Zhang, Yao Yao et al.

Image animation is a key task in computer vision which aims to generate dynamic visual content from static image. Recent image animation methods employ neural based rendering technique to generate realistic animations. Despite these advancements, achieving fine-grained and controllable image animation guided by text remains challenging, particularly for open-domain images captured in diverse real environments. In this paper, we introduce an open domain image animation method that leverages the motion prior of video diffusion model. Our approach introduces targeted motion area guidance and motion strength guidance, enabling precise control the movable area and its motion speed. This results in enhanced alignment between the animated visual elements and the prompting text, thereby facilitating a fine-grained and interactive animation generation process for intricate motion sequences. We validate the effectiveness of our method through rigorous experiments on an open-domain dataset, with the results showcasing its superior performance. Project page can be found at https://animationai.github.io/AnimateAnything.

Xiaodong Gu, Chengzhou Tang, Weihao Yuan et al.

3D motion estimation including scene flow and point cloud registration has drawn increasing interest. Inspired by 2D flow estimation, recent methods employ deep neural networks to construct the cost volume for estimating accurate 3D flow. However, these methods are limited by the fact that it is difficult to define a search window on point clouds because of the irregular data structure. In this paper, we avoid this irregularity by a simple yet effective method.We decompose the problem into two interlaced stages, where the 3D flows are optimized point-wisely at the first stage and then globally regularized in a recurrent network at the second stage. Therefore, the recurrent network only receives the regular point-wise information as the input. In the experiments, we evaluate the proposed method on both the 3D scene flow estimation and the point cloud registration task. For 3D scene flow estimation, we make comparisons on the widely used FlyingThings3D and KITTIdatasets. For point cloud registration, we follow previous works and evaluate the data pairs with large pose and partially overlapping from ModelNet40. The results show that our method outperforms the previous method and achieves a new state-of-the-art performance on both 3D scene flow estimation and point cloud registration, which demonstrates the superiority of the proposed zero-order method on irregular point cloud data.

Jiahui Zhang, Shitao Tang, Kejie Qiu et al.

Visual relocalization has been a widely discussed problem in 3D vision: given a pre-constructed 3D visual map, the 6 DoF (Degrees-of-Freedom) pose of a query image is estimated. Relocalization in large-scale indoor environments enables attractive applications such as augmented reality and robot navigation. However, appearance changes fast in such environments when the camera moves, which is challenging for the relocalization system. To address this problem, we propose a virtual view synthesis-based approach, RenderNet, to enrich the database and refine poses regarding this particular scenario. Instead of rendering real images which requires high-quality 3D models, we opt to directly render the needed global and local features of virtual viewpoints and apply them in the subsequent image retrieval and feature matching operations respectively. The proposed method can largely improve the performance in large-scale indoor environments, e.g., achieving an improvement of 7.1\% and 12.2\% on the Inloc dataset.

Pin-Lun Hsu, Yun Dai, Vignesh Kothapalli et al.

Training Large Language Models (LLMs) efficiently at scale presents a formidable challenge, driven by their ever-increasing computational demands and the need for enhanced performance. In this work, we introduce Liger-Kernel, an open-sourced set of Triton kernels developed specifically for LLM training. With kernel optimization techniques like kernel operation fusing and input chunking, our kernels achieve on average a 20% increase in training throughput and a 60% reduction in GPU memory usage for popular LLMs compared to HuggingFace implementations. In addition, Liger-Kernel is designed with modularity, accessibility, and adaptability in mind, catering to both casual and expert users. Comprehensive benchmarks and integration tests are built in to ensure compatibility, performance, correctness, and convergence across diverse computing environments and model architectures. The source code is available under a permissive license at: github.com/linkedin/Liger-Kernel.

Zhenghao Zhang, Fangtao Shao, Zuozhuo Dai et al.

Most existing transformer based video instance segmentation methods extract per frame features independently, hence it is challenging to solve the appearance deformation problem. In this paper, we observe the temporal information is important as well and we propose TAFormer to aggregate spatio-temporal features both in transformer encoder and decoder. Specifically, in transformer encoder, we propose a novel spatio-temporal joint multi-scale deformable attention module which dynamically integrates the spatial and temporal information to obtain enriched spatio-temporal features. In transformer decoder, we introduce a temporal self-attention module to enhance the frame level box queries with the temporal relation. Moreover, TAFormer adopts an instance level contrastive loss to increase the discriminability of instance query embeddings. Therefore the tracking error caused by visually similar instances can be decreased. Experimental results show that TAFormer effectively leverages the spatial and temporal information to obtain context-aware feature representation and outperforms state-of-the-art methods.

Qianyun He, Xinya Ji, Yicheng Gong et al.

We present a novel approach for synthesizing 3D talking heads with controllable emotion, featuring enhanced lip synchronization and rendering quality. Despite significant progress in the field, prior methods still suffer from multi-view consistency and a lack of emotional expressiveness. To address these issues, we collect EmoTalk3D dataset with calibrated multi-view videos, emotional annotations, and per-frame 3D geometry. By training on the EmoTalk3D dataset, we propose a \textit{`Speech-to-Geometry-to-Appearance'} mapping framework that first predicts faithful 3D geometry sequence from the audio features, then the appearance of a 3D talking head represented by 4D Gaussians is synthesized from the predicted geometry. The appearance is further disentangled into canonical and dynamic Gaussians, learned from multi-view videos, and fused to render free-view talking head animation. Moreover, our model enables controllable emotion in the generated talking heads and can be rendered in wide-range views. Our method exhibits improved rendering quality and stability in lip motion generation while capturing dynamic facial details such as wrinkles and subtle expressions. Experiments demonstrate the effectiveness of our approach in generating high-fidelity and emotion-controllable 3D talking heads. The code and EmoTalk3D dataset are released at https://nju-3dv.github.io/projects/EmoTalk3D.

Yuxiao He, Yiyu Zhuang, Yanwen Wang et al.

Creating a 360° parametric model of a human head is a very challenging task. While recent advancements have demonstrated the efficacy of leveraging synthetic data for building such parametric head models, their performance remains inadequate in crucial areas such as expression-driven animation, hairstyle editing, and text-based modifications. In this paper, we build a dataset of artist-designed high-fidelity human heads and propose to create a novel parametric 360° renderable parametric head model from it. Our scheme decouples the facial motion/shape and facial appearance, which are represented by a classic parametric 3D mesh model and an attached neural texture, respectively. We further propose a training method for decompositing hairstyle and facial appearance, allowing free-swapping of the hairstyle. A novel inversion fitting method is presented based on single image input with high generalization and fidelity. To the best of our knowledge, our model is the first parametric 3D full-head that achieves 360° free-view synthesis, image-based fitting, appearance editing, and animation within a single model. Experiments show that facial motions and appearances are well disentangled in the parametric space, leading to SOTA performance in rendering and animating quality. The code and SynHead100 dataset are released at https://nju-3dv.github.io/projects/Head360.

Zhichao Wei, Xiaohao Chen, Mingqiang Chen et al.

Referring image segmentation aims to segment the image region of interest according to the given language expression, which is a typical multi-modal task. Existing methods either adopt the pixel classification-based or the learnable query-based framework for mask generation, both of which are insufficient to deal with various text-image pairs with a fix number of parametric prototypes. In this work, we propose an end-to-end framework built on transformer to perform Linguistic query-Guided mask generation, dubbed LGFormer. It views the linguistic features as query to generate a specialized prototype for arbitrary input image-text pair, thus generating more consistent segmentation results. Moreover, we design several cross-modal interaction modules (\eg, vision-language bidirectional attention module, VLBA) in both encoder and decoder to achieve better cross-modal alignment.

Yuwei Sun, Yuxuan Yao, Hui Li et al.

Diffusion models have achieved success in high-fidelity data synthesis, yet their capacity for more complex, structured reasoning like text following tasks remains constrained. While advances in language models have leveraged strategies such as latent reasoning and recursion to enhance text understanding capabilities, extending these to multimodal text-to-image generation tasks is challenging due to the continuous and non-discrete nature of visual tokens. To tackle this problem, we draw inspiration from modular human cognition and propose a recursive, sparse mixture-of-experts framework integrated into conventional diffusion models. Our approach introduces a recursive component within joint attention layers that iteratively refines visual tokens over multiple latent steps while efficiently sharing parameters via sparse selection of neural modules. At each step, a gating network is devised to dynamically select specialized neural modules, conditioned on the current visual tokens, the diffusion timestep, and the conditioning information. Comprehensive evaluation on class-conditioned ImageNet image generation tasks and additional studies on the GenEval and DPG benchmark demonstrate the superiority of the proposed method in enhancing model image generation performance.

Zuozhuo Dai, Fangtao Shao, Qingkun Su et al.

State-of-the-art text-video retrieval (TVR) methods typically utilize CLIP and cosine similarity for efficient retrieval. Meanwhile, cross attention methods, which employ a transformer decoder to compute attention between each text query and all frames in a video, offer a more comprehensive interaction between text and videos. However, these methods lack important fine-grained spatial information as they directly compute attention between text and video-level tokens. To address this issue, we propose CrossTVR, a two-stage text-video retrieval architecture. In the first stage, we leverage existing TVR methods with cosine similarity network for efficient text/video candidate selection. In the second stage, we propose a novel decoupled video text cross attention module to capture fine-grained multimodal information in spatial and temporal dimensions. Additionally, we employ the frozen CLIP model strategy in fine-grained retrieval, enabling scalability to larger pre-trained vision models like ViT-G, resulting in improved retrieval performance. Experiments on text video retrieval datasets demonstrate the effectiveness and scalability of our proposed CrossTVR compared to state-of-the-art approaches.

Ce Liu, Jun Wang, Zhiqiang Cai et al.

Despite significant progress in static protein structure collection and prediction, the dynamic behavior of proteins, one of their most vital characteristics, has been largely overlooked in prior research. This oversight can be attributed to the limited availability, diversity, and heterogeneity of dynamic protein datasets. To address this gap, we propose to enhance existing prestigious static 3D protein structural databases, such as the Protein Data Bank (PDB), by integrating dynamic data and additional physical properties. Specifically, we introduce a large-scale dataset, Dynamic PDB, encompassing approximately 12.6K proteins, each subjected to all-atom molecular dynamics (MD) simulations lasting 1 microsecond to capture conformational changes. Furthermore, we provide a comprehensive suite of physical properties, including atomic velocities and forces, potential and kinetic energies of proteins, and the temperature of the simulation environment, recorded at 1 picosecond intervals throughout the simulations. For benchmarking purposes, we evaluate state-of-the-art methods on the proposed dataset for the task of trajectory prediction. To demonstrate the value of integrating richer physical properties in the study of protein dynamics and related model design, we base our approach on the SE(3) diffusion model and incorporate these physical properties into the trajectory prediction process. Preliminary results indicate that this straightforward extension of the SE(3) model yields improved accuracy, as measured by MAE and RMSD, when the proposed physical properties are taken into consideration. https://fudan-generative-vision.github.io/dynamicPDB/ .

Yizheng Song, Yiyu Zhuang, Qipeng Xu et al.

Single-view 3D human reconstruction has garnered significant attention in recent years. Despite numerous advancements, prior research has concentrated on reconstructing 3D models from clear, close-up images of individual subjects, often yielding subpar results in the more prevalent multi-person scenarios. Reconstructing 3D human crowd models is a highly intricate task, laden with challenges such as: 1) extensive occlusions, 2) low clarity, and 3) numerous and various appearances. To address this task, we propose CrowdGaussian, a unified framework that directly reconstructs multi-person 3D Gaussian Splatting (3DGS) representations from single-image inputs. To handle occlusions, we devise a self-supervised adaptation pipeline that enables the pretrained large human model to reconstruct complete 3D humans with plausible geometry and appearance from heavily occluded inputs. Furthermore, we introduce Self-Calibrated Learning (SCL). This training strategy enables single-step diffusion models to adaptively refine coarse renderings to optimal quality by blending identity-preserving samples with clean/corrupted image pairs. The outputs can be distilled back to enhance the quality of multi-person 3DGS representations. Extensive experiments demonstrate that CrowdGaussian generates photorealistic, geometrically coherent reconstructions of multi-person scenes.

Jiahao Cui, Yan Chen, Mingwang Xu et al.

Generating highly dynamic and photorealistic portrait animations driven by audio and skeletal motion remains challenging due to the need for precise lip synchronization, natural facial expressions, and high-fidelity body motion dynamics. We propose a human-preference-aligned diffusion framework that addresses these challenges through two key innovations. First, we introduce direct preference optimization tailored for human-centric animation, leveraging a curated dataset of human preferences to align generated outputs with perceptual metrics for portrait motion-video alignment and naturalness of expression. Second, the proposed temporal motion modulation resolves spatiotemporal resolution mismatches by reshaping motion conditions into dimensionally aligned latent features through temporal channel redistribution and proportional feature expansion, preserving the fidelity of high-frequency motion details in diffusion-based synthesis. The proposed mechanism is complementary to existing UNet and DiT-based portrait diffusion approaches, and experiments demonstrate obvious improvements in lip-audio synchronization, expression vividness, body motion coherence over baseline methods, alongside notable gains in human preference metrics. Our model and source code can be found at: https://github.com/xyz123xyz456/hallo4.

Kaihui Cheng, Ce Liu, Qingkun Su et al.

Protein structure prediction is pivotal for understanding the structure-function relationship of proteins, advancing biological research, and facilitating pharmaceutical development and experimental design. While deep learning methods and the expanded availability of experimental 3D protein structures have accelerated structure prediction, the dynamic nature of protein structures has received limited attention. This study introduces an innovative 4D diffusion model incorporating molecular dynamics (MD) simulation data to learn dynamic protein structures. Our approach is distinguished by the following components: (1) a unified diffusion model capable of generating dynamic protein structures, including both the backbone and side chains, utilizing atomic grouping and side-chain dihedral angle predictions; (2) a reference network that enhances structural consistency by integrating the latent embeddings of the initial 3D protein structures; and (3) a motion alignment module aimed at improving temporal structural coherence across multiple time steps. To our knowledge, this is the first diffusion-based model aimed at predicting protein trajectories across multiple time steps simultaneously. Validation on benchmark datasets demonstrates that our model exhibits high accuracy in predicting dynamic 3D structures of proteins containing up to 256 amino acids over 32 time steps, effectively capturing both local flexibility in stable states and significant conformational changes. URL: https://fudan-generative-vision.github.io/AlphaFolding/#/

Lin Zhong, Siyu Zhu, Zizhen Yuan et al.

Modeling human cognitive states is essential for advanced artificial intelligence. Existing Large Language Models (LLMs) mainly address isolated tasks such as emotion analysis or stance detection, and fail to capture interactions among cognitive dimensions defined in psychology, including emotion, thinking style, stance, and intention. To bridge this gap, we construct CognitiveBench, the first benchmark with unified annotations across the above four dimensions. Experiments on CognitiveBench show that although LLMs perform well on single dimension tasks, their performance drops sharply in joint multi-dimensional modeling. Using Gromov $δ$-hyperbolicity analysis, we find that CognitiveBench exhibits a strong hierarchical structure. We attribute the performance bottleneck to ``Cognitive Crowding'', where hierarchical cognitive states require exponential representational space, while the Euclidean space of LLMs grows only polynomially, causing representation overlap and degraded performance. To address this mismatch, we propose HyCoLLM, which models cognitive states in hyperbolic space and aligns LLM representations via Hyperbolic Guided Alignment Tuning. Results show that HyCoLLM substantially improves multi-dimensional cognitive understanding, allowing 8B parameter model to outperform strong baselines, including GPT-4o.

Junchao Liao, Zhenghao Zhang, Xiangyu Meng et al.

Audio-video (AV) generation has recently made strong progress in perceptual quality and multimodal coherence, yet generating content with plausible motion-sound relations remains challenging. Existing methods often produce object motions that are visually unstable and sounds that are only loosely aligned with salient motion or contact events, largely because they lack an explicit motion-aware structure shared by video and audio generation. We present Tora3, a trajectory-guided AV generation framework that improves physical coherence by using object trajectories as a shared kinematic prior. Rather than treating trajectories as a video-only control signal, Tora3 uses them to jointly guide visual motion and acoustic events. Specifically, we design a trajectory-aligned motion representation for video, a kinematic-audio alignment module driven by trajectory-derived second-order kinematic states, and a hybrid flow matching scheme that preserves trajectory fidelity in trajectory-conditioned regions while maintaining local coherence elsewhere. We further curate PAV, a large-scale AV dataset emphasizing motion-relevant patterns with automatically extracted motion annotations. Extensive experiments show that Tora3 improves motion realism, motion-sound synchronization, and overall AV generation quality over strong open-source baselines.

Baoyou Chen, Hanchen Xia, Peng Tu et al.

Autoregressive vision-language models (VLMs) deliver strong multimodal capability, but their token-by-token decoding imposes a fundamental inference bottleneck. Diffusion VLMs offer a more parallel decoding paradigm, yet directly converting a pretrained autoregressive VLM into a large-block diffusion VLM (dVLM) often leads to substantial quality degradation. In this work, we present BARD, a simple and effective bridging framework that converts a pretrained autoregressive VLM into a same-architecture, decoding-efficient dVLM. Our approach combines progressive supervised block merging, which gradually enlarges the decoding block size, with stage-wise intra-dVLM distillation from a fixed small-block diffusion anchor to recover performance lost at larger blocks. We further incorporate a mixed noise scheduler to improve robustness and token revision during denoising, and memory-friendly training to enable efficient training on long multimodal sequences. A key empirical finding is that direct autoregressive-to-diffusion distillation is poorly aligned and can even hurt performance, whereas distillation within the diffusion regime is consistently effective. Experimental results show that, with $\leq 4.4M$ data, BARD-VL transfers strong multimodal capability from Qwen3-VL to a large-block dVLM. Remarkably, BARD-VL establishes a new SOTA among comparable-scale open dVLMs on our evaluation suite at both 4B and 8B scales. At the same time, BARD-VL achieves up to \textbf{3$\times$} decoding throughput speedup compared to the source model.

Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai et al.

In this study, we introduce a methodology for human image animation by leveraging a 3D human parametric model within a latent diffusion framework to enhance shape alignment and motion guidance in curernt human generative techniques. The methodology utilizes the SMPL(Skinned Multi-Person Linear) model as the 3D human parametric model to establish a unified representation of body shape and pose. This facilitates the accurate capture of intricate human geometry and motion characteristics from source videos. Specifically, we incorporate rendered depth images, normal maps, and semantic maps obtained from SMPL sequences, alongside skeleton-based motion guidance, to enrich the conditions to the latent diffusion model with comprehensive 3D shape and detailed pose attributes. A multi-layer motion fusion module, integrating self-attention mechanisms, is employed to fuse the shape and motion latent representations in the spatial domain. By representing the 3D human parametric model as the motion guidance, we can perform parametric shape alignment of the human body between the reference image and the source video motion. Experimental evaluations conducted on benchmark datasets demonstrate the methodology's superior ability to generate high-quality human animations that accurately capture both pose and shape variations. Furthermore, our approach also exhibits superior generalization capabilities on the proposed in-the-wild dataset. Project page: https://fudan-generative-vision.github.io/champ.

Yuxuan Yao, Yuxuan Chen, Hui Li et al.

Multimodal Diffusion Transformers (MMDiTs) for text-to-image generation maintain separate text and image branches, with bidirectional information flow between text tokens and visual latents throughout denoising. In this setting, we observe a prompt forgetting phenomenon: the semantics of the prompt representation in the text branch is progressively forgotten as depth increases. We further verify this effect on three representative MMDiTs--SD3, SD3.5, and FLUX.1 by probing linguistic attributes of the representations over the layers in the text branch. Motivated by these findings, we introduce a training-free approach, prompt reinjection, which reinjects prompt representations from early layers into later layers to alleviate this forgetting. Experiments on GenEval, DPG, and T2I-CompBench++ show consistent gains in instruction-following capability, along with improvements on metrics capturing preference, aesthetics, and overall text--image generation quality.

Hui Li, Jiayue Lyu, Fu-Yun Wang et al.

This paper studies the training-testing discrepancy (a.k.a. exposure bias) problem for improving the diffusion models. During training, the input of a prediction network at one training timestep is the corresponding ground-truth noisy data that is an interpolation of the noise and the data, and during testing, the input is the generated noisy data. We present a novel training approach, named MixFlow, for improving the performance. Our approach is motivated by the Slow Flow phenomenon: the ground-truth interpolation that is the nearest to the generated noisy data at a given sampling timestep is observed to correspond to a higher-noise timestep (termed slowed timestep), i.e., the corresponding ground-truth timestep is slower than the sampling timestep. MixFlow leverages the interpolations at the slowed timesteps, named slowed interpolation mixture, for post-training the prediction network for each training timestep. Experiments over class-conditional image generation (including SiT, REPA, and RAE) and text-to-image generation validate the effectiveness of our approach. Our approach MixFlow over the RAE models achieve strong generation results on ImageNet: 1.43 FID (without guidance) and 1.10 (with guidance) at 256 x 256, and 1.55 FID (without guidance) and 1.10 (with guidance) at 512 x 512.

Youtian Lin, Zuozhuo Dai, Siyu Zhu et al.

We introduce Gaussian-Flow, a novel point-based approach for fast dynamic scene reconstruction and real-time rendering from both multi-view and monocular videos. In contrast to the prevalent NeRF-based approaches hampered by slow training and rendering speeds, our approach harnesses recent advancements in point-based 3D Gaussian Splatting (3DGS). Specifically, a novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model attribute deformations of each Gaussian point, where the time-dependent residual of each attribute is captured by a polynomial fitting in the time domain, and a Fourier series fitting in the frequency domain. The proposed DDDM is capable of modeling complex scene deformations across long video footage, eliminating the need for training separate 3DGS for each frame or introducing an additional implicit neural field to model 3D dynamics. Moreover, the explicit deformation modeling for discretized Gaussian points ensures ultra-fast training and rendering of a 4D scene, which is comparable to the original 3DGS designed for static 3D reconstruction. Our proposed approach showcases a substantial efficiency improvement, achieving a $5\times$ faster training speed compared to the per-frame 3DGS modeling. In addition, quantitative results demonstrate that the proposed Gaussian-Flow significantly outperforms previous leading methods in novel view rendering quality. Project page: https://nju-3dv.github.io/projects/Gaussian-Flow

Feng Ding, Haisheng Fu, Jie Liang et al.

We study full-reference image quality assessment from a machine-centric perspective, where images are evaluated by how well they preserve information for downstream models. We formulate machine-oriented quality as a latent machine utility and approximate it through pairwise predictive-consistency comparisons. To this end, we construct PCMP, a dataset of PSNR-matched distortion pairs labeled by consistency votes from multiple pretrained models. We further propose ML-CLIPSim, a differentiable quality metric built on a frozen CLIP visual encoder, which aggregates intermediate patch-token similarities and global image embeddings. Experiments on machine-preference benchmarks, human-IQA datasets, and learned image compression show that ML-CLIPSim better aligns with machine-oriented preferences than conventional fidelity and perceptual metrics, while remaining competitive for human quality prediction. Used as a compression distortion term, it improves rate--task trade-offs across multiple downstream tasks.

Hui Li, Baoyou Chen, Liwei Zhang et al.

Diffusion transformers (DiTs) adopt Patchify, mapping patch representations to token representations through linear projections, to adjust the number of tokens input to DiT blocks and thus the computation cost. Instead of a single patch size for all the timesteps, we introduce a Pyramidal Patchification Flow (PPFlow) approach: Large patch sizes are used for high noise timesteps and small patch sizes for low noise timesteps; Linear projections are learned for each patch size; and Unpatchify is accordingly modified. Unlike Pyramidal Flow, our approach operates over full latent representations other than pyramid representations, and adopts the normal denoising process without requiring the renoising trick. We demonstrate the effectiveness of our approach through two training manners. Training from scratch achieves a $1.6\times$ ($2.0\times$) inference speed over SiT-B/2 for 2-level (3-level) pyramid patchification with slightly lower training FLOPs and similar image generation performance. Training from pretrained normal DiTs achieves even better performance with small training time. The code and checkpoint are at https://github.com/fudan-generative-vision/PPFlow.

Yan Chen, Hanlin Shang, Ce Liu et al.

Video face restoration faces a critical challenge in maintaining temporal consistency while recovering fine facial details from degraded inputs. This paper presents a novel approach that extends Vector-Quantized Variational Autoencoders (VQ-VAEs), pretrained on static high-quality portraits, into a video restoration framework through variational latent space modeling. Our key innovation lies in reformulating discrete codebook representations as Dirichlet-distributed continuous variables, enabling probabilistic transitions between facial features across frames. A spatio-temporal Transformer architecture jointly models inter-frame dependencies and predicts latent distributions, while a Laplacian-constrained reconstruction loss combined with perceptual (LPIPS) regularization enhances both pixel accuracy and visual quality. Comprehensive evaluations on blind face restoration, video inpainting, and facial colorization tasks demonstrate state-of-the-art performance. This work establishes an effective paradigm for adapting intensive image priors, pretrained on high-quality images, to video restoration while addressing the critical challenge of flicker artifacts. The source code has been open-sourced and is available at https://github.com/fudan-generative-vision/DicFace.

Zhenghao Zhang, Shengfan Zhang, Zhichao Wei et al.

The current state-of-the-art methods for unsupervised video object segmentation (UVOS) require extensive training on video datasets with mask annotations, limiting their effectiveness in handling challenging scenarios. However, the Segment Anything Model (SAM) introduces a new prompt-driven paradigm for image segmentation, offering new possibilities. In this study, we investigate SAM's potential for UVOS through different prompt strategies. We then propose UVOSAM, a mask-free paradigm for UVOS that utilizes the STD-Net tracker. STD-Net incorporates a spatial-temporal decoupled deformable attention mechanism to establish an effective correlation between intra- and inter-frame features, remarkably enhancing the quality of box prompts in complex video scenes. Extensive experiments on the DAVIS2017-unsupervised and YoutubeVIS19\&21 datasets demonstrate the superior performance of UVOSAM without mask supervision compared to existing mask-supervised methods, as well as its ability to generalize to weakly-annotated video datasets. Code can be found at https://github.com/alibaba/UVOSAM.

Shitao Tang, Jiahui Zhang, Siyu Zhu et al.

Transformers have been successful in many vision tasks, thanks to their capability of capturing long-range dependency. However, their quadratic computational complexity poses a major obstacle for applying them to vision tasks requiring dense predictions, such as object detection, feature matching, stereo, etc. We introduce QuadTree Attention, which reduces the computational complexity from quadratic to linear. Our quadtree transformer builds token pyramids and computes attention in a coarse-to-fine manner. At each level, the top K patches with the highest attention scores are selected, such that at the next level, attention is only evaluated within the relevant regions corresponding to these top K patches. We demonstrate that quadtree attention achieves state-of-the-art performance in various vision tasks, e.g. with 4.0% improvement in feature matching on ScanNet, about 50% flops reduction in stereo matching, 0.4-1.5% improvement in top-1 accuracy on ImageNet classification, 1.2-1.8% improvement on COCO object detection, and 0.7-2.4% improvement on semantic segmentation over previous state-of-the-art transformers. The codes are available at https://github.com/Tangshitao/QuadtreeAttention.

Lizhe Liu, Xiaohao Chen, Siyu Zhu et al.

Modern deep-learning-based lane detection methods are successful in most scenarios but struggling for lane lines with complex topologies. In this work, we propose CondLaneNet, a novel top-to-down lane detection framework that detects the lane instances first and then dynamically predicts the line shape for each instance. Aiming to resolve lane instance-level discrimination problem, we introduce a conditional lane detection strategy based on conditional convolution and row-wise formulation. Further, we design the Recurrent Instance Module(RIM) to overcome the problem of detecting lane lines with complex topologies such as dense lines and fork lines. Benefit from the end-to-end pipeline which requires little post-process, our method has real-time efficiency. We extensively evaluate our method on three benchmarks of lane detection. Results show that our method achieves state-of-the-art performance on all three benchmark datasets. Moreover, our method has the coexistence of accuracy and efficiency, e.g. a 78.14 F1 score and 220 FPS on CULane. Our code is available at https://github.com/aliyun/conditional-lane-detection.

Shitao Tang, Chengzhou Tang, Rui Huang et al.

Camera localization aims to estimate 6 DoF camera poses from RGB images. Traditional methods detect and match interest points between a query image and a pre-built 3D model. Recent learning-based approaches encode scene structures into a specific convolutional neural network (CNN) and thus are able to predict dense coordinates from RGB images. However, most of them require re-training or re-adaption for a new scene and have difficulties in handling large-scale scenes due to limited network capacity. We present a new method for scene agnostic camera localization using dense scene matching (DSM), where a cost volume is constructed between a query image and a scene. The cost volume and the corresponding coordinates are processed by a CNN to predict dense coordinates. Camera poses can then be solved by PnP algorithms. In addition, our method can be extended to temporal domain, which leads to extra performance boost during testing time. Our scene-agnostic approach achieves comparable accuracy as the existing scene-specific approaches, such as KFNet, on the 7scenes and Cambridge benchmark. This approach also remarkably outperforms state-of-the-art scene-agnostic dense coordinate regression network SANet. The Code is available at https://github.com/Tangshitao/Dense-Scene-Matching.

Xiaodong Gu, Weihao Yuan, Zuozhuo Dai et al.

There are increasing interests of studying the video-to-depth (V2D) problem with machine learning techniques. While earlier methods directly learn a mapping from images to depth maps and camera poses, more recent works enforce multi-view geometry constraints through optimization embedded in the learning framework. This paper presents a novel optimization method based on recurrent neural networks to further exploit the potential of neural networks in V2D. Specifically, our neural optimizer alternately updates the depth and camera poses through iterations to minimize a feature-metric cost, and two gated recurrent units iteratively improve the results by tracing historical information. Extensive experimental results demonstrate that our method outperforms previous methods and is more efficient in computation and memory consumption than cost-volume-based methods. In particular, our self-supervised method outperforms previous supervised methods on the KITTI and ScanNet datasets. Our source code is available at https://github.com/aliyun/dro-sfm.

Zuozhuo Dai, Guangyuan Wang, Weihao Yuan et al.

State-of-the-art unsupervised re-ID methods train the neural networks using a memory-based non-parametric softmax loss. Instance feature vectors stored in memory are assigned pseudo-labels by clustering and updated at instance level. However, the varying cluster sizes leads to inconsistency in the updating progress of each cluster. To solve this problem, we present Cluster Contrast which stores feature vectors and computes contrast loss at the cluster level. Our approach employs a unique cluster representation to describe each cluster, resulting in a cluster-level memory dictionary. In this way, the consistency of clustering can be effectively maintained throughout the pipline and the GPU memory consumption can be significantly reduced. Thus, our method can solve the problem of cluster inconsistency and be applicable to larger data sets. In addition, we adopt different clustering algorithms to demonstrate the robustness and generalization of our framework. The application of Cluster Contrast to a standard unsupervised re-ID pipeline achieves considerable improvements of 9.9%, 8.3%, 12.1% compared to state-of-the-art purely unsupervised re-ID methods and 5.5%, 4.8%, 4.4% mAP compared to the state-of-the-art unsupervised domain adaptation re-ID methods on the Market, Duke, and MSMT17 datasets. Code is available at https://github.com/alibaba/cluster-contrast.

Hualie Jiang, Zhe Sheng, Siyu Zhu et al.

Learning depth from spherical panoramas is becoming a popular research topic because a panorama has a full field-of-view of the environment and provides a relatively complete description of a scene. However, applying well-studied CNNs for perspective images to the standard representation of spherical panoramas, i.e., the equirectangular projection, is suboptimal, as it becomes distorted towards the poles. Another representation is the cubemap projection, which is distortion-free but discontinued on edges and limited in the field-of-view. This paper introduces a new framework to fuse features from the two projections, unidirectionally feeding the cubemap features to the equirectangular features only at the decoding stage. Unlike the recent bidirectional fusion approach operating at both the encoding and decoding stages, our fusion scheme is much more efficient. Besides, we also designed a more effective fusion module for our fusion scheme. Experiments verify the effectiveness of our proposed fusion strategy and module, and our model achieves state-of-the-art performance on four popular datasets. Additional experiments show that our model also has the advantages of model complexity and generalization capability.The code is available at https://github.com/alibaba/UniFuse-Unidirectional-Fusion.

Rakesh Shrestha, Zhiwen Fan, Qingkun Su et al.

Deep learning based 3D shape generation methods generally utilize latent features extracted from color images to encode the semantics of objects and guide the shape generation process. These color image semantics only implicitly encode 3D information, potentially limiting the accuracy of the generated shapes. In this paper we propose a multi-view mesh generation method which incorporates geometry information explicitly by using the features from intermediate depth representations of multi-view stereo and regularizing the 3D shapes against these depth images. First, our system predicts a coarse 3D volume from the color images by probabilistically merging voxel occupancy grids from the prediction of individual views. Then the depth images from multi-view stereo along with the rendered depth images of the coarse shape are used as a contrastive input whose features guide the refinement of the coarse shape through a series of graph convolution networks. Notably, we achieve superior results than state-of-the-art multi-view shape generation methods with 34% decrease in Chamfer distance to ground truth and 14% increase in F1-score on ShapeNet dataset.Our source code is available at https://git.io/Jmalg

Sicong Tang, Feitong Tan, Kelvin Cheng et al.

This paper presents a neural network to estimate a detailed depth map of the foreground human in a single RGB image. The result captures geometry details such as cloth wrinkles, which are important in visualization applications. To achieve this goal, we separate the depth map into a smooth base shape and a residual detail shape and design a network with two branches to regress them respectively. We design a training strategy to ensure both base and detail shapes can be faithfully learned by the corresponding network branches. Furthermore, we introduce a novel network layer to fuse a rough depth map and surface normals to further improve the final result. Quantitative comparison with fused `ground truth' captured by real depth cameras and qualitative examples on unconstrained Internet images demonstrate the strength of the proposed method. The code is available at https://github.com/sfu-gruvi-3dv/deep_human.

Tianwei Shen, Zixin Luo, Lei Zhou et al.

Convolutional Neural Networks (CNNs) have achieved superior performance on object image retrieval, while Bag-of-Words (BoW) models with handcrafted local features still dominate the retrieval of overlapping images in 3D reconstruction. In this paper, we narrow down this gap by presenting an efficient CNN-based method to retrieve images with overlaps, which we refer to as the matchable image retrieval problem. Different from previous methods that generates training data based on sparse reconstruction, we create a large-scale image database with rich 3D geometrics and exploit information from surface reconstruction to obtain fine-grained training data. We propose a batched triplet-based loss function combined with mesh re-projection to effectively learn the CNN representation. The proposed method significantly accelerates the image retrieval process in 3D reconstruction and outperforms the state-of-the-art CNN-based and BoW methods for matchable image retrieval. The code and data are available at https://github.com/hlzz/mirror.

Yifei Zeng, Yanqin Jiang, Siyu Zhu et al.

Recent progress in pre-trained diffusion models and 3D generation have spurred interest in 4D content creation. However, achieving high-fidelity 4D generation with spatial-temporal consistency remains a challenge. In this work, we propose STAG4D, a novel framework that combines pre-trained diffusion models with dynamic 3D Gaussian splatting for high-fidelity 4D generation. Drawing inspiration from 3D generation techniques, we utilize a multi-view diffusion model to initialize multi-view images anchoring on the input video frames, where the video can be either real-world captured or generated by a video diffusion model. To ensure the temporal consistency of the multi-view sequence initialization, we introduce a simple yet effective fusion strategy to leverage the first frame as a temporal anchor in the self-attention computation. With the almost consistent multi-view sequences, we then apply the score distillation sampling to optimize the 4D Gaussian point cloud. The 4D Gaussian spatting is specially crafted for the generation task, where an adaptive densification strategy is proposed to mitigate the unstable Gaussian gradient for robust optimization. Notably, the proposed pipeline does not require any pre-training or fine-tuning of diffusion networks, offering a more accessible and practical solution for the 4D generation task. Extensive experiments demonstrate that our method outperforms prior 4D generation works in rendering quality, spatial-temporal consistency, and generation robustness, setting a new state-of-the-art for 4D generation from diverse inputs, including text, image, and video.

Jiahao Cui, Hui Li, Yun Zhan et al.

Existing methodologies for animating portrait images face significant challenges, particularly in handling non-frontal perspectives, rendering dynamic objects around the portrait, and generating immersive, realistic backgrounds. In this paper, we introduce the first application of a pretrained transformer-based video generative model that demonstrates strong generalization capabilities and generates highly dynamic, realistic videos for portrait animation, effectively addressing these challenges. The adoption of a new video backbone model makes previous U-Net-based methods for identity maintenance, audio conditioning, and video extrapolation inapplicable. To address this limitation, we design an identity reference network consisting of a causal 3D VAE combined with a stacked series of transformer layers, ensuring consistent facial identity across video sequences. Additionally, we investigate various speech audio conditioning and motion frame mechanisms to enable the generation of continuous video driven by speech audio. Our method is validated through experiments on benchmark and newly proposed wild datasets, demonstrating substantial improvements over prior methods in generating realistic portraits characterized by diverse orientations within dynamic and immersive scenes. Further visualizations and the source code are available at: https://fudan-generative-vision.github.io/hallo3/.

Qi Zuo, Xiaodong Gu, Lingteng Qiu et al.

Generating multi-view images based on text or single-image prompts is a critical capability for the creation of 3D content. Two fundamental questions on this topic are what data we use for training and how to ensure multi-view consistency. This paper introduces a novel framework that makes fundamental contributions to both questions. Unlike leveraging images from 2D diffusion models for training, we propose a dense consistent multi-view generation model that is fine-tuned from off-the-shelf video generative models. Images from video generative models are more suitable for multi-view generation because the underlying network architecture that generates them employs a temporal module to enforce frame consistency. Moreover, the video data sets used to train these models are abundant and diverse, leading to a reduced train-finetuning domain gap. To enhance multi-view consistency, we introduce a 3D-Aware Denoising Sampling, which first employs a feed-forward reconstruction module to get an explicit global 3D model, and then adopts a sampling strategy that effectively involves images rendered from the global 3D model into the denoising sampling loop to improve the multi-view consistency of the final images. As a by-product, this module also provides a fast way to create 3D assets represented by 3D Gaussians within a few seconds. Our approach can generate 24 dense views and converges much faster in training than state-of-the-art approaches (4 GPU hours versus many thousand GPU hours) with comparable visual quality and consistency. By further fine-tuning, our approach outperforms existing state-of-the-art methods in both quantitative metrics and visual effects. Our project page is aigc3d.github.io/VideoMV.

Hui Li, Mingwang Xu, Yun Zhan et al.

Recent advancements in visual generation technologies have markedly increased the scale and availability of video datasets, which are crucial for training effective video generation models. However, a significant lack of high-quality, human-centric video datasets presents a challenge to progress in this field. To bridge this gap, we introduce OpenHumanVid, a large-scale and high-quality human-centric video dataset characterized by precise and detailed captions that encompass both human appearance and motion states, along with supplementary human motion conditions, including skeleton sequences and speech audio. To validate the efficacy of this dataset and the associated training strategies, we propose an extension of existing classical diffusion transformer architectures and conduct further pretraining of our models on the proposed dataset. Our findings yield two critical insights: First, the incorporation of a large-scale, high-quality dataset substantially enhances evaluation metrics for generated human videos while preserving performance in general video generation tasks. Second, the effective alignment of text with human appearance, human motion, and facial motion is essential for producing high-quality video outputs. Based on these insights and corresponding methodologies, the straightforward extended network trained on the proposed dataset demonstrates an obvious improvement in the generation of human-centric videos. Project page https://fudan-generative-vision.github.io/OpenHumanVid