CVPR 2026 Papers

Training a unified model integrating video-to-audio (V2A), text-to-audio (T2A), and joint video-text-to-audio (VT2A) generation offers significant application flexibility, yet faces two unexplored foundational challenges: (1) the scarcity of high-quality audio captions with tight V-A-T alignment, leading to severe semantic conflict between multimodal conditions, and (2) cross-task and intra-task competition, manifesting as an adverse V2A-T2A performance trade-off and modality bias in the VT2A task. First, to address data scarcity, we introduce SoundAtlas, a large-scale dataset (470k pairs) that significantly outperforms existing benchmarks and even human experts in quality. Powered by a novel agentic pipeline, it integrates Vision-to-Language Compression to mitigate visual bias of MLLMs, a Junior-Senior Agent Hand-off for a five times cost reduction, and rigorous Post-hoc Filtering to ensure fidelity. Consequently, SoundAtlas delivers semantically rich and temporally detailed captions with tight V-A-T alignment. Second, we propose Omni2Sound, a unified VT2A diffusion model supporting flexible input modalities. To resolve the inherent cross-task and intra-task competition, we design a three-stage multi-task progressive training schedule that converts cross-task competition into joint optimization and mitigates modality bias in the VT2A task, maintaining both audio-visual alignment and off-screen audio generation faithfulness. Finally, we construct VGGSound-Omni, a comprehensive benchmark for unified evaluation, including challenging off-screen tracks. With a standard DiT backbone, Omni2Sound achieves unified SOTA performance across all three tasks within a single model, demonstrating strong generalization across benchmarks with heterogeneous input conditions.

Diffusion transformers have achieved remarkable progress in high-quality image and video generation, but their computational overhead remains a significant challenge. Existing token reduction-based acceleration techniques, such as caching and merging, attempt to reduce this cost from both temporal and spatial perspectives, but often compromise generation quality by introducing non-updated or non-self denosing directions. In this paper, we propose Residual Caching (ResCa), a novel, training-free framework that introduces a proxy denoising perspective to overcome these limitations. ResCa achieves acceleration while maintaining a denoising trajectory that is both self and updated. The core idea is to perform true denoising on only one proxy token within each trajectory-based cluster, and use its computed multi-order residuals to guide the simulated denoising of all other tokens. ResCa can be seamlessly integrated into various diffusion models, including DiT, FLUX, and HunyuanVideo. Extensive quantitative and qualitative experiments demonstrate the effectiveness of our method, achieving up to a 5.5 times acceleration in GFLOPs while maintaining near-lossless generation quality on FLUX. Project page: https://fanghaipeng.github.io/ResCa.

Synthesis of diverse driving scenes serves as a crucial data augmentation technique for validating the robustness and generalizability of autonomous driving systems. Current methods aggregate high-definition (HD) maps and 3D bounding boxes as geometric conditions in diffusion models for conditional scene generation. However, implicit inter-condition dependency causes generation failures when control conditions change independently. Additionally, these methods suffer from insufficient details in both semantic and structural aspects. Specifically, brief and view-invariant captions restrict semantic contexts, resulting in weak background modeling. Meanwhile, the standard denoising loss with uniform spatial weighting neglects foreground structural details, causing visual distortions and blurriness. To address these challenges, we propose DrivePTS, which incorporates three key innovations. Firstly, our framework adopts a progressive learning strategy to mitigate inter-dependency between geometric conditions, reinforced by an explicit mutual information constraint. Secondly, a Vision-Language Model is utilized to generate multi-view hierarchical descriptions across six semantic aspects, providing fine-grained textual guidance. Thirdly, a frequency-guided structure loss is introduced to strengthen the model's sensitivity to high-frequency elements, improving foreground structural fidelity. Extensive experiments demonstrate that our DrivePTS achieves state-of-the-art fidelity and controllability in generating diverse driving scenes. Notably, DrivePTS successfully generates rare scenes where prior methods fail, highlighting its strong generalization ability.

Diffusion-native watermarking provides a more secure and reliable way to trace images from latent diffusion models (LDMs) by embedding information directly into the generative process. However, existing methods suffer from a fundamental limitation: their embedding capacity is extremely small. We introduce MaxMark, a high-capacity watermarking framework that supports embed rich watermark messages into generated images. MaxMark uses two components: a robust watermark embedding module that enhance the secret message and places them into reliable regions of the latent noise, and a distribution transformation module that maps the watermarked latent back to an approximate Gaussian, ensuring compatibility with the diffusion process and preserving image fidelity. The distribution transformation is implemented with an invertible neural network (INN), whose exactly reversible structure enables precise recovery and efficient training. Experiments show that MaxMark surpasses prior methods in capacity, robustness, and imperceptibility, achieving up to a 46% improvement in bit accuracy for large watermark payloads.

Diffusion Transformers (DiTs) have achieved state-of-the-art image and video generation performance, but sampling remains expensive due to repeated transformer forward passes over many timesteps. Feature caching offers a training-free way to accelerate inference by reusing or forecasting hidden representations, yet recent forecasting-based methods derive their coefficients from hand-crafted formulas (e.g., Taylor expansion), which ultimately reduce to fixed linear combinations of a few historical features. Such fixed coefficients are suboptimal and fragile under aggressive skipping. In this paper, we first show that existing forecasting-based caching methods can be unified in a common linear form, and then analyze DiT feature trajectories, finding that for most denoising steps the current feature can be reconstructed from past features with projection fidelity above 0.95, indicating that accurate linear prediction is feasible. Motivated by this, we propose L^2P (Learnable Linear Predictor), a simple data-driven caching framework that replaces hand-designed coefficients with learnable per-timestep weights trained on a small set of cached trajectories using a mean-squared error loss, converging in about 20 seconds on a single GPU. Extensive experiments on state-of-the-art DiTs demonstrate that L^2P consistently outperforms existing caching baselines: on FLUX.1-dev, L^2P achieves a 4.55xFLOPs reduction and 4.15xlatency speedup with a PSNR of 31.459, and on Qwen-Image and Qwen-Image-Lightning, it maintains high visual fidelity even under up to 7.18xacceleration, where prior methods suffer from noticeable quality degradation. These results show that learning linear predictors is a practical and effective alternative to designing increasingly complex forecasting formulas for efficient diffusion model inference. The implementation code is available at https://github.com/Aredstone/L2P-Cache.

Infrared imaging is essential for perception in harsh environments. However, dynamically coupled degradation factors severely impair visual quality and downstream semantic accuracy. Although generative diffusion models provide strong image restoration priors, high computational cost and physical inconsistency limit their application in infrared sensing. To bridge these gaps, we reformulate infrared imaging as a single-step diffusion process, aligning degraded observations with trajectory latent states via dynamic timestep estimation to leverage timestep-specific diffusion priors for high-fidelity reconstruction. Meanwhile, we introduce a spectral regularization term to enforce thermal radiation constraints and ensure physical consistency. Subsequently, a task-aware low-rank adaptation mechanism is devised through dynamic prompting to enable efficient transfer across downstream infrared tasks. Experiments demonstrate our method surpasses existing approaches in restoration quality, semantic structure preservation, and task generalization. The code is available at https://github.com/csmty/InfraredIR.

Benefiting from the powerful generative priors of diffusion models, diffusion-based real-world image super-resolution (Real-ISR) methods have demonstrated impressive performance.To achieve efficient Real-ISR, several recent works have designed one-step diffusion-based models.Howerver, unmediatedly feeding LR into a diffusion model creates a distributional gap with the model's original input.A straightforward approach to reduce the distribution gap is to introduce noise to the LR latents. However, directly adding noise inevitably corrupts the content of the LR images.In this study, we propose DNF-SR, a Dual-input and Negative-aware Feature fine-tuning method for Real-ISR.Specifically, we use a dual-input strategy that concatenates the original LR image with the noisy LR input and feeds them into a diffusion-based image editing model, ensuring both high-fidelity one-step super-resolution and improved perceptual and content consistency.Additionally, the noise present in the noisy LR input introduces randomness and diversity into the outputs. We exploit this property and propose a post-training optimization method, Negative-aware Feature Fine-Tuning (NF2T), which guides the model toward producing higher-quality results.NF^2T classifies multiple outputs into positive and negative subsets and then defines implicit policy improvement directions in both the image and feature spaces, thereby further enhancing the stability of the optimization.Extensive experiments show that DNF-SR outperforms other methods.Code will be released.

Recently, vision foundation models have been shown to boost the efficiency of flow-based generative models by revealing the intrinsic union-of-manifold structures and lowering the complexity of the latent/target distribution. In this paper, we exploit the multimodality aspect of the union-of-manifold structures, and aim to further improve the learning and inference efficiency for flow-matching models. To this end, we propose an efficient source and coupling co-design method termed Mixture-Modeling Flow Matching (MM-FM), by integrating a data-adaptive multimodal source distribution (implemented as Gaussian mixture models) and mode-dependent data coupling. The former shortens the distance between the source and the target, and the latter promotes local and straighter flows. We also derive theoretical results to confirm our intuition in a quantitative sense. In our experiments on ImageNet256x256 with multimodal DINOv2-B latents, MM-FM exhibits superior learning efficiency and state-of-the-art unconditional generation quality: FID=2.74 with autoguidance in only 80 epochs.

Generative models have shown great potential in trajectory planning. Recent studies demonstrate that anchor-guided generative models are effective in modeling the uncertainty of driving behaviors and improving overall performance. However, these methods rely on discrete anchor vocabularies that must sufficiently cover the trajectory distribution during testing to ensure robustness, inducing an inherent trade-off between vocabulary size and model performance.To overcome this limitation, we propose MeanFuser, an end-to-end autonomous driving method that enhances both efficiency and robustness through three key designs. (1) We introduce Gaussian Mixture Noise (GMN) to guide generative sampling, enabling a continuous representation of the trajectory space and eliminating the dependency on discrete anchor vocabularies. (2) We introduce "MeanFlow Identity", which models the mean velocity field between GMN and data distribution instead of the instantaneous velocity field used in naive flow-matching methods, effectively eliminating numerical errors from ODE solvers and significantly accelerating inference. (3) We design a lightweight Adaptive Reconstruction Module (ARM) that enables the model to consider all sampled proposals and adaptively decide whether to reconstruct a trajectory when none of the proposals is satisfactory. Experiments on the NAVSIM closed-loop benchmark demonstrate that MeanFuser achieves outstanding performance and exceptional inference efficiency, offering a robust and efficient solution for end-to-end autonomous driving.

Monocular human motion capture in occlusion scenarios presents significant challenges. Although a few works have explicitly considered the occlusion problem, image-based methods are unreliable due to the lack of temporal constraints while video-based approaches cannot gain sufficient knowledge from time domain motion priors to address long-term occlusions. However, occluded human motion typically exhibits periodic patterns and consistent momentum. Inspired by this observation, we exploit reliable image observations in frequency domain and formulate the motion capture task as a wavelet coefficients selection process. Specifically, we first construct probabilistic distributions for the occluded 2D keypoints, and then introduce a frequency domain diffusion model to refine the distributions by learning long-term periodic information and physical momentum with Discrete Wavelet Transform (DWT). Consequently, the learned denoising prior can select valid wavelet components to facilitate the 3D motion capture with a 3D decoder. By employing a joint reprojection strategy, we can also use the same diffusion process to train the 3D decoder. To further promote human occlusion-related tasks, we also present the first 3D occluded motion dataset, OcMotion, which serves as a new benchmark for both training and evaluation. Experimental results demonstrate that our method can produce accurate and coherent human motions from occluded videos. More information is available at https://github.com/boycehbz/FreqMotion.

While recent text-to-video (T2V) diffusion models have achieved impressive quality and prompt alignment, they often produce low-diversity outputs when sampling multiple videos from a single text prompt. We tackle this challenge by formulating it as a set-level policy optimization problem, with the goal of training a policy that can cover the diverse range of plausible outcomes for a given prompt. To address this, we introduce DPP-GRPO, a novel framework for diverse video generation generation that combines Determinantal Point Processes (DPPs) and Group Relative Policy Optimization (GRPO) theories to enforce explicit reward on diverse generations. Our objective turns diversity into an explicit signal by imposing diminishing returns on redundant samples (via DPP) while supplies groupwise feedback over candidate sets (via GRPO). Our framework is plug-and-play and model-agnostic, and encourages diverse generations across visual appearance, camera motions, and scene structure without sacrificing prompt fidelity or perceptual quality. We implement our method on WAN and CogVideoX, and show that our method consistently improves video diversity on state-of-the-art benchmarks such as VBench, VideoScore, and human preference studies. Moreover, we release our code and a new benchmark dataset of 30,000 diverse prompts to support future research.

Generating human motion with precise spatial control is a challenging problem. Existing approaches often require task-specific training or slow optimization, and enforcing hard constraints frequently disrupts motion naturalness. Building on the observation that many animation tasks can be formulated as a linear inverse problem, we introduce ProjFlow, a training-free sampler that achieves zero-shot, exact satisfaction of linear spatial constraints while preserving motion realism. Our key advance is a novel kinematics-aware metric that encodes skeletal topology. This metric allows the sampler to enforce hard constraints by distributing corrections coherently across the entire skeleton, avoiding the unnatural artifacts of naive projection. Furthermore, for sparse inputs, such as filling in long gaps between a few keyframes, we introduce a time-varying formulation using pseudo-observations that fade during sampling. Extensive experiments on representative applications, motion inpainting, and 2D-to-3D lifting, demonstrate that ProjFlow achieves exact constraint satisfaction and matches or improves realism over zero-shot baselines, while remaining competitive with training-based controllers.

Training-free image editing with large diffusion models has become practical, yet faithfully performing complex non-rigid edits (e.g., pose or shape changes) remains highly challenging. We identify a key underlying cause: attention collapse in existing attention sharing mechanisms, where either positional embeddings or semantic features dominate visual content retrieval, leading to over-editing or under-editing.To address this issue, we introduce SynPS, a method that Synergistically leveragesPositional embeddings and Semantic information for faithful non-rigid image editing. We first propose an editing measurement that quantifies the required editing magnitude at each denoising step. Based on this measurement, we design an attention synergy pipeline that dynamically modulates the influence of positional embeddings, enabling SynPS to balance semantic modifications and fidelity preservation.By adaptively integrating positional and semantic cues, SynPS effectively avoids both over- and under-editing. Extensive experiments on public and newly curated benchmarks demonstrate the superior performance and faithfulness of our approach. Our code will be publicly released.

Current semantic segmentation models are very data-hungry and require massive costly pixel-wise human annotations. Generative data augmentation, which scales the train set using generative models, provides a potential remedy. In this paper, we propose MatchMask, a novel mask-centric generative data augmentation approach tailored for label-scarce semantic segmentation. By leveraging a limited set of labeled semantic masks, MatchMask generates diverse, realistic, and well-aligned image-mask pairs, thereby enhancing the performance of semantic segmentation models. Specifically, to adapt existing text-to-image models for semantic image synthesis in the few-shot setting, we first propose a Gradient Probe Method to investigate the role of each layer in the diffusion model. On this basis, a lightweight LoRA-style adapter is designed for critical layers to enable efficient adaptation, coupled with a Layer-adaptive Cross-attention Fusion mechanism. Meanwhile, we present a robust relative filtering principle to suppress incorrectly synthesized regions. Moreover, the proposed approach is extended to MatchMask++ in the semi-supervised setting to take advantage of additional unlabeled data. Experimental results on PASCAL VOC, COCO and ADE20K demonstrate that MatchMask remarkably enhances the performance of segmentation models, surpassing prior data augmentation techniques in various benchmarks, e.g., 67.5%->74.3% mIoU on PASCAL VOC.

Sign Language Production (SLP) aims to translate spoken language into sign sequences, where the main challenge lies in generating coherent and natural poses from discrete glosses (G2P). Existing G2P methods typically treat each pose as an indivisible unit, limiting their ability to capture fine-grained joint-level dependencies and thus degrading pose quality. To address this, we propose the Focal-General Diffusion Model (FGDM), characterized by a pioneering two-stage denoising framework that harmonizes local joint-level dependencies and global coherence. Specifically, in the Focal stage, a novel Adaptive Sign GCN (ASGCN) adaptively models each pose based on contextual correlations, skeletal topology, and semantic conditions, ensuring precise generation of local details. In the General stage, a Transformer-based module refines the entire pose sequence to enhance global coherence and naturalness. Moreover, we introduce a Semantic Consistent Guidance (SCG) mechanism that seamlessly integrates semantic supervision into diffusion training, enforcing tighter alignment between generated pose sequences and their intended gloss semantics. Extensive experiments on PHOENIX14T and USTC-CSL demonstrate that FGDM achieves SOTA performance. Our project page is available at https://yuyiheng-eu.github.io/fgdm/.

Motion transfer from the driving to the source portrait remains a key challenge in the portrait animation. Current diffusion-based approaches condition only on the driving motion, which fails to capture source-to-driving correspondences and consequently yields suboptimal motion transfer. Although flow estimation provides an alternative, predicting dense correspondences from 2D input is ill-posed and often yields inaccurate animation. We address this problem by introducing 3D flows, a learning-free and geometry-driven motion correspondence directly computed from parametric 3D head models. To integrate this 3D prior into diffusion model, we introduce 3D flow encoding to query potential 3D flows for each target pixel to indicate its displacement back to the source location. To obtain 3D flows aligned with 2D motion changes, we further propose depth-guided sampling to accurately locate the corresponding 3D points for each pixel. Beyond high-fidelity portrait animation, our model further supports user-specified editing of facial expression and head pose. Extensive experiments demonstrate the superiority of our method on consistent driving motion transfer as well as faithful source identity preservation.

Deep learning models are well known to be susceptible to backdoor attacks, and text-to-image generation models are no exception. When a specific trigger is embedded in the input, a backdoored model can be manipulated to perform attacker-defined malicious behaviors, such as generating harmful or inappropriate images. Existing backdoor attacks on text-to-image generation models are largely limited to dirty-label attacks, where misaligned image-caption pairs are injected into the training data. While effective in controlled settings, such methods are often easily detectable, limiting their practicality in realistic applications. To address this limitation, we propose the first clean-label backdoor attack for text-to-image generative models, which preserves semantic consistency within poisoned image-caption pairs to evade detection. We design a dual-modality manipulation strategy that injects nearly imperceptible noise into images while embedding a composite semantic text trigger. The text trigger combines synonym substitution and syntactic restructuring, enabling stealthy yet effective backdoor implantation without compromising the visual-textual alignment. Experimental results demonstrate that our method achieves high attack success while effectively preserving model utility and evading mainstream defenses, including commercial content filters.

Compositional video generation aims to synthesize multiple instances with diverse appearance and motion. However, current approaches mainly focus on binding semantics, neglecting to understand diverse motion categories specified in prompts. In this paper, we propose a motion factorization framework that decomposes complex motion into three primary categories: motionlessness, rigid motion, and non-rigid motion. Specifically, our framework follows a planning before generation paradigm. (1) During planning, we reason about motion laws on the motion graph to obtain frame-wise changes in the shape and position of each instance. This alleviates semantic ambiguities in the user prompt by organizing it into a structured representation of instances and their interactions. (2) During generation, we modulate the synthesis of distinct motion categories in a disentangled manner. Conditioned on the motion cues, guidance branches stabilize appearance in motionless regions, preserve rigid-body geometry, and regularize local non-rigid deformations. Crucially, our two modules are model-agnostic, which can be seamlessly incorporated into various diffusion model architectures. Extensive experiments demonstrate that our framework achieves impressive performance in motion synthesis on real-world benchmarks. Code is available at https://github.com/ZixuanWang0525/MF-CVG.

High-resolution image editing is essential for professional and creative applications, yet existing multimodal diffusion-based editors remain computationally inefficient and constrained to relatively low resolutions. Current approaches redundantly process the entire image canvas or rely on large-scale high-resolution datasets, resulting in substantial training and inference costs. We introduce **HierEdit**, a region-aware hierarchical diffusion framework designed for efficient and scalable high-resolution image editing. Our method first performs edits on a low-resolution proxy using an off-the-shelf editing model to generate a reference and to localize the modified regions. A hierarchical local-window diffusion model (**Local-Window MMDiT**) that refines only edited regions within the original high-resolution image, while reusing the unaltered regions as conditioning inputs. The low-resolution proxy further provides structural guidance and intermediate denoising supervision (**Inference Acceleration**) , ensuring consistent global semantics and stable generation without the need for full-resolution attention computation. This targeted and hierarchical design enables fast, high-fidelity editing of images up to 4K resolution without requiring any specialized high-resolution training data. Extensive experiments demonstrate that **HierEdit** achieves competitive visual quality on commodity-resolution datasets while significantly accelerating inference and extending seamlessly to ultra-high-resolution 4K editing.