CVPR 2026 论文列表

We introduce MoS (Mixture of States), a novel fusion paradigm for multimodal diffusion models that merges modalities using flexible, state-based interactions. The core of MoS is a learnable, token-wise router that creates denoising timestep- and input-dependent interactions between modalities' hidden states, precisely aligning token-level features with the diffusion trajectory. This router sparsely selects the top-k hidden states and is trained with an epsilon-greedy strategy, efficiently selecting contextual features with minimal learnable parameters and negligible computational overhead. We validate our design with text-to-image generation (MoS-Image) and editing (MoS-Editing), which achieve state-of-the-art results. With only 3B to 5B parameters, our models match or surpass counterparts up to 4xlarger. These findings establish MoS as a flexible and compute-efficient paradigm for scaling multimodal diffusion models. The project homepage is available at https://haozheliu-st.github.io/mos-homepage/

Generative models excel at motion synthesis for a fixed number of agents but struggle to generalize with variable agents. Based on limited, domain-specific data, existing methods employ autoregressive models to generate motion recursively, which suffer from inefficiency and error accumulation. We propose Unified Motion Flow (UMF), which consists of Pyramid Motion Flow (P-Flow) and Semi-Noise Motion Flow (S-Flow). UMF decomposes the number-free motion generation into a single-pass motion prior generation stage and multi-pass reaction generation stages. Specifically, UMF utilizes a unified latent space to bridge the distribution gap between heterogeneous motion datasets, enabling effective unified training. For motion prior generation, P-Flow operates on hierarchical resolutions conditioned on different noise levels, thereby mitigating computational overheads. For reaction generation, S-Flow learns a joint probabilistic path that adaptively performs reaction transformation and context reconstruction, alleviating error accumulation. Extensive results and user studies demonstrate UMF's effectiveness as a generalist model for multi-person motion generation from text. Project page: https://githubhgh.github.io/umf/.

Recent unified models integrate understanding experts (e.g., LLMs) with generative experts (e.g., diffusion models), achieving strong multimodal performance. However, recent advanced methods such as BAGEL and LMFusion follow the Mixture-of-Transformers (MoT) paradigm, adopting a symmetric design that mirrors one expert to another for convenient initialization and fusion, which remains suboptimal due to inherent modality discrepancies. In this work, we propose HBridge, an asymmetric H-shaped architecture that enables heterogeneous experts to optimally leverage pretrained priors from their respective Recent unified models integrate understanding experts (e.g., LLMs) with generative experts (e.g., diffusion models), achieving strong multimodal performance. However, recent advanced methods such as BAGEL and LMFusion follow the Mixture-of-Transformers (MoT) paradigm, adopting a symmetric design that mirrors one expert to another for convenient initialization and fusion, which remains suboptimal due to inherent modality discrepancies. modality domains. Unlike prior dense fusion strategies that straightforwardly connect all layers between experts via shared attention, HBridge selectively bridges intermediate layers, reducing over 40% attention sharing, which improves efficiency and enhances generation quality. Shallow and deep layers, which capture modality-specific representations, are decoupled, while mid-layer bridging promotes semantic alignment. To further strengthen cross-modal coherence, we introduce semantic reconstruction tokens that explicitly guide the generative expert to reconstruct visual semantic tokens of the target image. Extensive experiments across multiple benchmarks demonstrate the effectiveness and superior performance of HBridge, establishing a new paradigm for unified multimodal generation.

Recent large vision-language models (LVLMs) have been applied to diverse VQA tasks. However, achieving practical performance typically requires task-specific fine-tuning with large numbers of image-text pairs, which are costly to collect. In this work, we study text-centric training, a setting where only textual descriptions are available and no real images are provided, as a paradigm for low-cost data scaling. Unlike images, whose collection is often restricted by privacy constraints and scarcity in niche domains, text is widely available. Moreover, text is easily editable, enabling automatic diversification and expansion with LLMs at minimal human effort. While this offers clear advantages over image collection in terms of scalability and cost, training on raw text without images still yields limited gains on VQA tasks because of the image-text modality gap. To address this issue, we propose a Text-Printed Image (TPI), which generates synthetic images by directly rendering the given textual description on a plain white canvas. This simple rendering projects text into the image modality and can be integrated into many existing LVLM training pipelines at low cost. Moreover, TPI preserves the semantics of the text, whereas text-to-image models often fail to do. Across four models and seven benchmarks, our systematic experiments show that TPI enables more effective text-centric training than synthetic images generated by a diffusion model. We further explore TPI as a low-cost data-augmentation strategy and demonstrate its practical utility. Overall, our findings highlight the significant potential of text-centric training and, more broadly, chart a path toward fully automated data generation for LVLMs.

Recent advances in generative modeling have substantially enhanced novel view synthesis, yet maintaining consistency across viewpoints remains challenging. Diffusion-based models rely on stochastic noise-to-data transitions, which obscure deterministic structures and yield inconsistent view predictions. We advocate a Data-to-Data Flow Matching framework that learns deterministic transformations between paired views, enhancing view-consistent synthesis through explicit data coupling. Building on this, we propose Probability Density Geodesic Flow Matching (PDG-FM), which aligns interpolation trajectories with density-based geodesics of a data manifold. To enable tractable geodesic estimation, we employ a teacher-student framework that distills density-based geodesic interpolants into an efficient ambient-space predictor. Empirically, our method surpasses diffusion-based baselines on Objaverse and GSO30 datasets, demonstrating improved structural coherence and smoother transitions across views. These results highlight the advantages of incorporating data-dependent geometric regularization into deterministic flow matching for consistent novel view generation.

RAW images preserve superior fidelity and rich scene information compared to RGB, making them essential for tasks in challenging imaging conditions. To alleviate the high cost of data collection, recent RGB-to-RAW conversion methods aim to synthesize RAW images from RGB. However, they overlook two key challenges: (i) the reconstruction difficulty varies with pixel intensity, and (ii) multi-camera conversion requires camera-specific adaptation. To address these issues, we propose SpiralDiff, a diffusion based framework tailored for RGB-to-RAW conversion with a signal-dependent noise weighting strategy that adapts reconstruction fidelity across intensity levels. In addition, we introduce CamLoRA, a camera-aware lightweight adaptation module that enables a unified model to adapt to different camera-specific ISP characteristics. Extensive experiments on four benchmark datasets demonstrate the superiority of SpiralDiff in RGB-to-RAW conversion quality and its downstream benefits in RAW-based object detection. Our code is available at https://github.com/Chuancy-TJU/SpiralDiff.

Diffusion models achieve strong generative performance but remain slow at inference due to the need for repeated full-model denoising passes. We present Token-Adaptive Predictor (TAP), a training-free, probe-driven framework that adaptively selects a predictor for each token at every sampling step. TAP uses a single full evaluation of the model's first layer as a low-cost probe to compute proxy losses for a compact family of candidate predictors (instantiated primarily with Taylor expansions of varying order and horizon), then assigns each token the predictor with the smallest proxy error. This per-token "probe-then-select" strategy exploits heterogeneous temporal dynamics, requires no additional training, and is compatible with various predictor designs. TAP incurs negligible overhead while enabling large speedups with little or no perceptual quality loss. Extensive experiments across multiple diffusion architectures and generation tasks show that TAP substantially improves the accuracy-efficiency frontier compared to fixed global predictors and caching-only baselines.

Diffusion Models have emerged as a leading class of generative models, yet their iterative sampling process remains computationally expensive. Timestep distillation is a promising technique to accelerate generation, but it often requires extensive training and leads to image quality degradation. Furthermore, fine-tuning these distilled models for specific objectives, such as aesthetic appeal or user preference, using Reinforcement Learning (RL) is notoriously unstable and easily falls into reward hacking. In this work, we introduce Flash-DMD, a novel framework that enables fast convergence with distillation and joint RL-based refinement.Specifically, we first propose an efficient timestep-aware distillation strategy that significantly reduces training cost with enhanced realism, outperforming DMD2 with only 2.1% its training cost. Second, we introduce a joint training scheme where the model is fine-tuned with an RL objective while the timestep distillation training continues simultaneously. We demonstrate that the stable, well-defined loss from the ongoing distillation acts as a powerful regularizer, effectively stabilizing the RL training process and preventing policy collapse. Extensive experiments on score-based and flow matching models show that our proposed Flash-DMD not only converges significantly faster but also achieves state-of-the-art generation quality in the few-step sampling regime, outperforming existing methods in visual quality, human preference, and text-image alignment metrics. Our work presents an effective paradigm for training efficient, high-fidelity, and stable generative models. Codes are attached in the supplementary.

In recent times, large datasets hinder efficient model training while also containing redundant concepts. Dataset distillation aims to synthesize compact datasets that preserve the knowledge of large-scale training sets while drastically reducing storage and computation. Recent advances in diffusion models have enabled training-free distillation by leveraging pre-trained generative priors; however, existing guidance strategies remain limited. Current score-based methods either perform unguided denoising or rely on simple mode-based guidance toward instance prototype centroids (IPC centroids), which often are rudimentary and suboptimal. We propose Manifold-Guided Distillation (ManifoldGD), a training-free diffusion-based framework that integrates manifold consistent guidance at every denoising timestep. Our method employs IPCs computed via a hierarchical, divisive clustering of VAE latent features--yielding a multi-scale coreset of IPCs that captures both coarse semantic modes and fine intra-class variability. Using a local neighborhood of the extracted IPC centroids, we create the latent manifold for each diffusion denoising timestep. At each denoising step, we project the mode-alignment vector onto the local tangent space of the estimated latent manifold, thus constraining the generation trajectory to remain manifold-faithful while preserving semantic consistency. This formulation improves representativeness, diversity, and image fidelity without requiring any model retraining. Empirical results demonstrate consistent gains over existing training-free and training-based baselines in terms of FID, l_2 distance among real and synthetic dataset embeddings, and classification accuracy, establishing ManifoldGD as the first geometry-aware training-free data distillation framework.

Human motions are compositional: complex behaviors can be described as combinations of simpler primitives. However, existing approaches primarily focus on forward modeling, e.g., learning holistic mappings from text to motion or composing a complex motion from a set of motion concepts. In this paper, we consider the inverse perspective: decomposing a holistic motion into semantically meaningful sub-components. We propose DeMoGen, a compositional training paradigm for decompositional learning that employs an energy-based diffusion model. This energy formulation directly captures the composed distribution of multiple motion concepts, enabling the model to discover them without relying on ground-truth motions for individual concepts. Within this paradigm, we introduce three training variants to encourage a decompositional understanding of motion: 1. DeMoGen-Exp explicitly trains on decomposed text prompts; 2. DeMoGen-OSS performs orthogonal self-supervised decomposition; 3. DeMoGen-SC enforces semantic consistency between original and decomposed text embeddings. These variants enable our approach to disentangle reusable motion primitives from complex motion sequences. We also demonstrate that the decomposed motion concepts can be flexibly recombined to generate diverse and novel motions, generalizing beyond the training distribution. Additionally, we construct a text-decomposed dataset to support compositional training, serving as an extended resource to facilitate text-to-motion generation and motion composition.

We present DuoMo, a generative method that recovers human motion in world-space coordinates from unconstrained videos with noisy or incomplete observations. Reconstructing such motion requires solving a fundamental trade-off: generalizing from diverse and noisy video inputs while maintaining global motion consistency. Our approach addresses this problem by factorizing motion learning into two diffusion models. The camera-space model first estimates motion from videos in camera coordinates. The world-space model then lifts this initial estimate into world coordinates and refines it to be globally consistent. Together, the two models can reconstruct motion across diverse scenes and trajectories, even from highly noisy or incomplete observations. Moreover, our formulation is general, generating the motion of mesh vertices directly (bypassing parametric models). DuoMo achieves state-of-the-art performance. On EMDB, our method obtains a 16% reduction in world-space MPJPE error while maintaining low foot skating. On RICH, it obtains a 30% reduction in world-space MPJPE error. Project page: yufu-wang.github.io/duomo/

While traditional and neural video codecs (NVCs) have achieved remarkable rate-distortion performance, improving perceptual quality at low bitrates remains challenging. Some NVCs incorporate perceptual or adversarial objectives but still suffer from artifacts due to limited generation capacity, whereas others leverage pretrained diffusion models to improve quality at the cost of high sampling complexity. To overcome these challenges, we propose S2VC, a Single Step diffusion-based Video Codec that integrates a conditional coding framework with an efficient single-step diffusion generator, enabling realistic reconstruction at low bitrates with reduced sampling cost. Recognizing the importance of semantic conditioning in single-step diffusion, we introduce Contextual Semantic Guidance to extract frame-adaptive semantics from buffered features. This guidance replaces text captions with efficient, fine-grained conditioning, thereby improving generation realism. In addition, Temporal Consistency Guidance is incorporated into the diffusion U-Net to enforce temporal coherence across frames and ensure stable generation. Extensive experiments show that S2VC delivers state-of-the-art perceptual quality with an average bitrate saving of 51.62% over prior perceptual method, underscoring the promise of single-step diffusion for efficient, high-quality video compression. Project: https://onedc-codec.github.io/s2vc/

As recent AI models have successfully generated high-resolution photorealistic images, it has also been socially important to detect whether an image is generated by AI. Since training data for the detection task is often not available due to the diversity of generative models, training-free detection approaches have been practically considered. A common approach is to utilize the image-level reconstruction error from the latent diffusion model (LDM). However, we find this score suffers from instance-specific biases, particularly in images with simple backgrounds. To this end, we propose a novel image-level debiasing score function that cancels out background contribution by normalizing the reconstruction error on the augmented images with similar background information. To be specific, we show that rotation and low-pass filtering are effective augmentation strategies. To promote generalization to broader generative models, we newly explore latent-level reconstruction error as an additional training-free signal. However, we observe that the latent-level score also suffers to latent-specific bias. To mitigate this, we introduce a rotation-based latent-level debiasing score based on the normalization of the rotated latent. We unify the aforementioned scores into a single unified debiasing score, RDD, which achieves state-of-the-art training-free detection performance across diverse generative models. Furthermore, our framework can be robust to corruption of the examined images.

Diffusion-based video editing has emerged as an important paradigm for high-quality and flexible content generation. However, despite their generality and strong modeling capacity, Diffusion Transformers (DiT) remain computationally expensive due to the iterative denoising process, posing challenges for practical deployment. Existing video diffusion acceleration methods primarily exploit denoising timestep-level feature reuse, which mitigates the redundancy in denoising process, but overlooks the architectural redundancy within the DiT that many attention operations over spatio-temporal tokens are redundantly executed, offering little to no incremental contribution to the model's output. This work introduces HetCache, a training-free diffusion acceleration framework designed to exploit the inherent heterogeneity in diffusion-based masked video-to-video (MV2V) generation and editing. Instead of uniformly reuse or randomly sampling tokens, HetCache assesses the contextual relevance and interaction strength among various types of tokens in designated computing steps. Guided by spatial priors, it divides the spatial-temporal tokens in DiT model into context and generative tokens, and selectively caches the context tokens that exhibit the strongest correlation and most representative semantics with generative ones. This strategy reduces redundant attention operations while maintaining editing consistency and fidelity. Experiments show that HetCache achieves a noticeable acceleration, including a 2.67 xlatency speedup and FLOPs reduction over commonly used foundation models, with negligible degradation in editing quality.

Current 3D human animation methods fail at photorealism: kinematics-based approaches lack non-rigid dynamics like clothing, while methods reconstructing from generated videos suffer from low-quality artifacts and identity loss. To overcome these limitations, we present Ani3DHuman, a framework that marries kinematics-based animation with video diffusion priors. We first introduce a layered motion representation that disentangles rigid motion and residual non-rigid motion. Then, we use a pretrained video diffusion model to restore a coarse rendering from the mesh-rigged animation, which provides supervision for the motion field. However, this restoration task, based on diffusion sampling, is highly challenging, as the initial renderings are out-of-distribution, causing standard deterministic ODE samplers to fail. Therefore, our core technical contribution is self-guided stochastic sampling, which effectively solves the out-of-distribution problem by combining stochastic sampling (for photorealistic quality) with self-guidance (for identity fidelity). These restored videos provide high-quality supervision, enabling the optimization of a realistic 4D motion field. Ani3DHuman achieves state-of-the-art results, and our ablations validate that both components of our sampler are essential for high-fidelity restoration.

Generating animated 3D objects is at the heart of many applications, yet most advanced works are typically difficult to apply in practice because of their limited setup, their long runtime, or their limited quality. We introduce ActionMesh, a generative model that predicts production-ready 3D meshes "in action" in a feed-forward manner. Drawing inspiration from early video models, our key insight is to modify existing 3D diffusion models to include a temporal axis, resulting in a framework we dubbed "temporal 3D diffusion". Specifically, we first adapt the 3D diffusion stage to generate a sequence of synchronized latents representing time-varying and independent 3D shapes. Second, we design a temporal 3D autoencoder that translates a sequence of independent shapes into the corresponding deformations of a pre-defined reference shape, allowing us to build an animation. Combining these two components, ActionMesh generates animated 3D meshes from different inputs like a monocular video, a text description, or even a 3D mesh with a text prompt describing its animation. Besides, compared to previous approaches, our method is fast and produces results that are rig-free and topology consistent, hence enabling rapid iteration and seamless applications like texturing and retargeting. We evaluate our model on standard video-to-4D benchmarks (Consistent4D, Objaverse) and report state-of-the-art performances on both geometric accuracy and temporal consistency, demonstrating that our model can deliver animated 3D meshes with unprecedented speed and quality.

We introduce WAM-Flow, a vision-language-action (VLA) model that casts ego-trajectory planning as discrete flow matching over a structured token space. In contrast to autoregressive decoders, WAM-Flow performs fully parallel, bidirectional denoising, enabling coarse-to-fine refinement with a tunable compute-accuracy trade-off. Specifically, the approach combines a metric-aligned numerical tokenizer that preserves scalar geometry via triplet-margin learning, a geometry-aware flow objective and a simulator-guided GRPO alignment that integrates safety, ego progress, and comfort rewards while retaining parallel generation. A multi-stage adaptation converts a pre-trained auto-regressive backbone (Janus-1.5B) from causal decoding to non-causal flow model and strengthens road-scene competence through continued multimodal pretraining. Thanks to the inherent nature of consistency model training and parallel decoding inference, WAM-Flow achieves superior closed-loop performance against autoregressive and diffusion-based VLA baselines, with 1-step inference attaining 88.7 PDMS and 5-step inference reaching 90.3 PDMS on NAVSIM v1 benchmark. These results establish discrete flow matching as a new promising paradigm for end-to-end autonomous driving.

Egocentric video generation with fine-grained control through body motion is a key requirement towards embodied AI agents that can simulate, predict, and plan actions. In this work, we propose EgoControl, a pose-controllable video diffusion model trained on egocentric data. We train a video prediction model to condition future frame generation on explicit 3D body pose sequences. To achieve precise motion control, we introduce a novel pose representation that captures both global camera dynamics and articulated body movements, and integrate it through a dedicated control mechanism within the diffusion process. Given a short sequence of observed frames and a sequence of target poses, EgoControl generates temporally coherent and visually realistic future frames that align with the provided pose control. Experimental results demonstrate that EgoControl produces high-quality, pose-consistent egocentric videos, paving the way toward controllable embodied video simulation and understanding.

Most existing 3D keypoint estimation methods rely on manual annotations or calibrated multi-view images, both of which are expensive to collect.This paper introduces KeyDiff3D, a framework that can accurately predict 3D keypoints from a single image, thus eliminating the need for such expensive data acquisitions.To achieve this, we leverage powerful geometric priors embedded in a pretrained multi-view diffusion model.In our framework, the diffusion model generates multi-view images from a single image, serving as supervision signals to provide 3D geometric cues to our model.We also introduce a 3D feature extractor that transforms implicit 3D priors embedded in the diffusion features into explicit 3D feature volumes.Beyond accurate keypoint estimation, we further introduce a pipeline that enables manipulation of 3D objects generated by the diffusion model.Experimental results on diverse datasets, including Human3.6M, CUB-200-2011, Stanford Dogs, and several in-the-wild and out-of-domain inputs, highlight the effectiveness of our method in terms of accuracy, generalization, and its ability to enable manipulation of 3D objects generated by the diffusion model from a single image.