CVPR 2026 Papers

Unrestricted adversarial attacks based on generative models typically operate either directly in image space or through diffusion-style denoising and re-noising, which limits transferability and robustness against defenses. We revisit this problem through the lens of flow matching and continuous-time velocity fields, and propose AdvFM, a velocity-field attack that injects adversarial signals into the flow-matching dynamics instead of the pixel space. Given a noisy state x_t, AdvFM perturbs the reconstruction at t = 1 and converts this perturbation into a change of the velocity field, yielding a state update that amplifies the inner PGD step in the noisy space. We further introduce a lookahead variant that optimizes a two-point objective over the current and rolled-out reconstructions, reducing temporal mismatch along the ODE trajectory. From a theoretical perspective, we show that compared to diffusion-based attacks, AdvFM enjoys: (i) larger single-step increases in the black-box loss via step amplification, (ii) reduced gradient variance and stronger surrogate-target alignment due to Gaussian smoothing, enhancing its transferability, and (iii) perturbations that concentrate in robust-tangent directions, thereby aligning with robust gradients of adversarially trained models and surviving purification more effectively; the lookahead variant further lowers gradient noise for a two-point robust objective. Extensive experiments demonstrate that AdvFM achieves promising performance in both black-box transferability and a suite of adversarial training and purification defenses.

Diffusion models have recently emerged as the dominant approach in visual generation tasks. However, the lengthy denoising chains and the computationally intensive noise estimation networks hinder their applicability in low-latency and resource-limited environments. Previous research has endeavored to address these limitations in a decoupled manner, utilizing either advanced samplers or efficient model quantization techniques. In this study, we uncover that quantization-induced noise disrupts directional estimation at each sampling step, further distorting the precise directional estimations of higher-order samplers when solving the sampling equations through discretized numerical methods, thereby altering the optimal sampling trajectory. To attain dual acceleration with high fidelity, we propose a sampling-aware quantization strategy, wherein a Mixed-Order Trajectory Alignment technique is devised to impose a more stringent constraint on the error bounds at each sampling step, facilitating a more linear probability flow. Extensive experiments on sparse-step fast sampling across multiple datasets demonstrate that our approach preserves the rapid convergence characteristics of high-speed samplers while maintaining superior generation quality. Code is publicly available at: https://github.com/TaylorJocelyn/Sampling-aware-Quantization.

Existing autoregressive (AR) methods for generating artist-designed meshes struggle to balance global structural consistency with high-fidelity local details, and are susceptible to error accumulation. To address this, we propose PartDiffuser, a novel semi-autoregressive diffusion framework for point-cloud-to-mesh generation. The method first performs semantic segmentation on the mesh and then operates in a "part-wise" manner: it employs autoregression between parts to ensure global topology, while utilizing a parallel discrete diffusion process within each semantic part to precisely reconstruct high-frequency geometric features. PartDiffuser is based on the DiT architecture and introduces a part-aware cross-attention mechanism, using point clouds as hierarchical geometric conditioning to dynamically control the generation process, thereby effectively decoupling the global and local generation tasks. Experiments demonstrate that this method significantly outperforms state-of-the-art (SOTA) models in generating 3D meshes with rich detail, exhibiting exceptional detail representation suitable for real-world applications.

Recently proposed pyramidal models decompose the conventional forward and backward diffusion processes into multiple stages operating at varying resolutions. These models handle inputs with higher noise levels at lower resolutions, while less noisy inputs are processed at higher resolutions. This hierarchical approach significantly reduces the computational cost of inference in multi-step denoising models. However, existing open-source pyramidal video models have been trained from scratch and tend to underperform compared to state-of-the-art systems in terms of visual plausibility.In this work, we present a pipeline that converts a pretrained diffusion model into a pyramidal one through low-cost finetuning, achieving this transformation without degradation in quality of output videos.Furthermore, we investigate and compare various strategies for step distillation within pyramidal models, aiming to further enhance the inference efficiency.

Generating realistic human-human interactions is a challenging task that requires not only high-quality individual body and hand motions, but also coherent coordination among all interactants. Due to limitations in available data and increased learning complexity, previous methods tend to ignore hand motions, limiting the realism and expressivity of the interactions. Additionally, current diffusion-based approaches generate entire motion sequences simultaneously, limiting their ability to capture the reactive and adaptive nature of human interactions. To address these limitations, we introduce Interact2Ar, the first end-to-end text-conditioned autoregressive diffusion model for generating full-body, human-human interactions. Interact2Ar incorporates detailed hand kinematics through dedicated parallel branches, enabling high-fidelity full-body generation. Furthermore, we introduce an autoregressive pipeline coupled with a novel memory technique that facilitates adaptation to the inherent variability of human interactions using efficient large context windows. The adaptability of our model enables a series of downstream applications, including temporal motion composition, real-time adaptation to disturbances, and extension beyond dyadic to multi-person scenarios. To validate the generated motions, we introduce a set of robust evaluators and extended metrics designed specifically for assessing full-body interactions. Through quantitative and qualitative experiments, we demonstrate the state-of-the-art performance of Interact2Ar.

Generative models and vision encoders have largely advanced on separate tracks, optimized for different goals and grounded in different mathematical principles. Yet, they share a fundamental property: latent space Gaussianity. Generative models map Gaussian noise to images, while encoders map images to semantic embeddings whose coordinates empirically behave as Gaussian. We hypothesize that both are views of a shared latent source, the Universal Normal Embedding (UNE): an approximately Gaussian latent space from which encoder embeddings and DDIM-inverted noise arise as noisy linear projections. To test our hypothesis, we introduce NoiseZoo, a dataset of per-image latents comprising DDIM-inverted diffusion noise and matching encoder representations (CLIP, DINO). On CelebA, linear probes in both spaces yield strong, aligned attribute predictions, indicating that generative noise encodes meaningful semantics along linear directions. These directions further enable faithful, controllable edits (e.g., smile, gender, age) without architectural changes, where simple orthogonalization mitigates spurious entanglements. Taken together, our results provide empirical support for the UNE hypothesis and reveal a shared Gaussian-like latent geometry that concretely links encoding and generation.

Action diffusion excels at high-fidelity action generation but incurs heavy computational costs owing to its iterative denoising nature. Despite current technologies showing promise in accelerating diffusion transformers by reusing the cached features, they struggle to adapt to policy dynamics arising from diverse perceptions and multi-round rollout iterations in open environments. We propose test-time sparsity to tackle this challenge, which aims to accelerate action diffusion by dynamically predicting prunable residual computations for each model forward at test time. However, two bottlenecks remain in this paradigm: 1) repetitive conditional encoding and pruning offset most potential speed gains, and 2) the features cached from previous denoising timesteps cannot constrain large pruning errors under aggressive sparsity. To address the first bottleneck, we design a highly parallelized inference pipeline that minimizes the non-decoder delay to milliseconds. Specifically, we first design a lightweight pruner that shares the encoder with the diffusion transformer. Then, we decouple the encoding and pruning from the autoregressive denoising loop by processing all denoising timesteps in parallel, and overlap the pruner with the decoder forward inference through asynchronism. To overcome the second bottleneck, we introduce an omnidirectional reusing strategy, which achieves 95% sparsity by selectively reusing the features cached from the current forward, previous denoising timesteps, and earlier rollout iterations. To learn the rollout-level reusing strategies, we sample a few action trajectories to supervise the sparsified diffusion step by step. Extensive experiments demonstrate that our method reduces FLOPs by 92% and accelerates action generation by 5x, achieving lossless performance with an inference frequency of 47.5 Hz. Our code is available at https://github.com/ky-ji/Test-time-Sparsity.

Understanding and predicting motion is a fundamental component of visual intelligence. Although modern video models exhibit strong comprehension of scene dynamics, exploring multiple possible futures through full video synthesis remains prohibitively inefficient. We model scene dynamics orders of magnitude more efficiently by directly operating on a long-term motion embedding that is learned from large-scale trajectories obtained from tracker models.This enables efficient generation of long, realistic motions that fulfill goals specified via text prompts or spatial pokes. To achieve this, we first learn a highly compressed motion embedding with a temporal compression factor of 64x. In this space, we train a conditional flow-matching model to generate motion latents conditioned on task descriptions. The resulting motion distributions outperform those of both state-of-the-art video models and specialized task-specific approaches.

We introduce COT-FM, a general framework that reshapes the probability path in Flow Matching (FM) to achieve faster and more reliable generation. FM models often produce curved trajectories due to random or batch-wise couplings, which increase discretization error and reduce sample quality. COT-FM fixes this by clustering target samples and assigning each cluster a dedicated source distribution obtained by reversing pretrained FM models. This divide-and-conquer strategy yields more accurate local transport and significantly straighter vector fields, all without changing the model architecture. As a plug-and-play approach, COT-FM consistently accelerates sampling and improves generation quality across 2D datasets, image benchmarks, and robotic manipulation tasks.

Jigsaw puzzle solving has been an increasingly popular task in the computer vision research community. Recent works have utilized cutting-edge architectures and computational approaches to reassemble groups of pieces into a coherent image, while achieving increasingly good results on well established datasets. However, most of these approaches share a common, restricting setting: operating solely on strictly square puzzle pieces. In this work, we introduce GAP, a set of novel jigsaw puzzles datasets containing synthetic, heavily eroded pieces of unrestricted shapes, generated by a learned distribution of real-world archaeological fragments. We also introduce PuzzleFlow, a novel ViT and Flow-Matching based framework for jigsaw puzzle solving, capable of handling complex puzzle pieces and demonstrating superior performance on GAP when compared to both classic and recent prominent works in this domain.

Egocentric perception enables humans to experience and understand the world directly from their own point of view. Translating exocentric (third-person) videos into egocentric (first-person) videos opens up new possibilities for immersive understanding but remains highly challenging due to extreme camera pose variations and minimal view overlap. This task requires faithfully preserving visible content while synthesizing unseen regions in a geometrically consistent manner. To achieve this, we present EgoX, a novel framework for generating egocentric videos from a single exocentric input.EgoX leverages the pretrained spatio-temporal knowledge of large-scale video diffusion models through lightweight LoRA adaptation and introduces a unified conditioning strategy that combines exocentric and egocentric priors via width- and channel-wise concatenation.Additionally, a geometry-guided self-attention mechanism selectively attends to spatially relevant regions, ensuring geometric coherence and high visual fidelity.Our approach achieves coherent and realistic egocentric video generation while demonstrating strong scalability and robustness across unseen and in-the-wild videos.

Diffusion models have achieved outstanding success in image generation, yet their objectives are often limited to reconstruction, making it difficult to align with human preferences directly. Reinforcement learning (RL) offers a promising approach to address this by optimizing models using explicit reward signals. However, most studies apply RL across the entire denoising process, which is both computationally expensive and tends to weaken preference alignment, i.e., doing more but achieving less. We observe that the impact of RL fine-tuning varies significantly across denoising stages. In the early stage, image structures are unstable and distant from the final reward signal. Applying RL at this stage leads to delayed rewards and action-reward mismatching, resulting in high variance and inefficient updates. Conversely, in the later stage, reward gains saturate, and continued training tends to overfit local details, intensifying reward hacking. To tackle these challenges, we propose \ourmethod , an RL-enhanced plug-in that improves generation quality while reducing computational cost. Specifically, \ourmethod adaptively identifies the optimal timing for RL by perceiving the structural evolution and semantic consistency during denoising, and dynamically terminates training once the denoising converges and reward gains saturate. As a result, it achieves a rare `dual benefit': a reduction in computational costs alongside a significant performance improvement. Theoretical analysis from an entropy perspective and extensive experiments verify our claims: compared with state-of-the-art methods, \ourmethod improves performance by xx% while cutting computational cost by xx%.

Recent multimodal face generation models address the spatial control limitations of text-to-image diffusion models by augmenting text-based conditioning with spatial priors such as segmentation masks, sketches, or edge maps. This multimodal fusion enables controllable synthesis aligned with both high-level semantic intent and low-level structural layout. However, most existing approaches typically extend pre-trained text-to-image pipelines by appending auxiliary control modules or stitching together separate uni-modal networks. These ad hoc designs inherit architectural constraints, duplicate parameters, and often fail under conflicting modalities or mismatched latent spaces, limiting their ability to perform synergistic fusion across semantic and spatial domains. We introduce MMFace-DiT, a unified dual-stream diffusion transformer engineered for synergistic multimodal face synthesis. Its core novelty lies in a dual-stream transformer block that processes spatial (mask/sketch) and semantic (text) tokens in parallel, deeply fusing them through a shared Rotary Position-Embedded (RoPE) Attention mechanism. This design prevents modal dominance and ensures strong adherence to both text and structural priors to achieve unprecedented spatial-semantic consistency for controllable face generation. Furthermore, a novel Modality Embedder enables a single cohesive model to dynamically adapt to varying spatial conditions without retraining. MMFace-DiT achieves a 40% improvement in visual fidelity and prompt alignment over six state-of-the-art multimodal face generation models, establishing a flexible new paradigm for end-to-end controllable generative modeling.

Recent text-to-image (T2I) diffusion and flow-matching models can produce highly realistic images from natural language prompts. In practical scenarios, T2I systems are often run in a "generate--then--select" mode: many seeds are sampled and only a few images are kept for use. However, this pipeline is highly resource-intensive since each candidate requires tens to hundreds of denoising steps, and evaluation metrics such as CLIPScore and ImageReward are post-hoc. In this work, we address this inefficiency by introducing Probe-Select, a plug-in module that enables efficient evaluation of image quality within the generation process. We observe that certain intermediate denoiser activations, even at early timesteps, encode a stable coarse structure, object layout and spatial arrangement--that strongly correlates with final image fidelity. Probe-Select exploits this property by predicting final quality scores directly from early activations, allowing unpromising seeds to be terminated early. Across diffusion and flow-matching backbones, our experiments show that early evaluation at only 20% of the trajectory accurately ranks candidate seeds and enables selective continuation. This strategy reduces sampling cost by over 60% while improving the quality of the retained images, demonstrating that early structural signals can effectively guide selective generation without altering the underlying generative model. Code is available at https://github.com/Guhuary/ProbeSelect.

Transformer architectures, particularly Diffusion Transformers (DiTs), have become widely used in diffusion and flow-matching models due to their strong performance compared to convolutional UNets. However, the isotropic design of DiTs processes the same number of patchified tokens in every block, leading to relatively heavy computation during training process. In this work, we introduce a multi-patch transformer design in which early blocks operate on larger patches to capture coarse global context, while later blocks use smaller patches to refine local details. This hierarchical design could reduces computational cost by up to 50% in GFLOPs while achieving good generative performance. In addition, we also propose improved designs for time and class embeddings that accelerate training convergence. Extensive experiments on the ImageNet dataset demonstrate the effectiveness of our architectural choices.

Recent advances in diffusion transformers have shown remarkable generalization in visual synthesis, yet most dense perception methods still rely on text-to-image (T2I) generators designed for stochastic generation. We revisit this paradigm and show that image editing diffusion models are inherently image-to-image consistent, providing a more suitable foundation for dense perception task. We introduce Edit2Perceive, a unified diffusion framework that adapts editing models for depth, normal, and matting. Built upon the FLUX.1 Kontext architecture, our approach employs full-parameter fine-tuning and a pixel-space consistency loss to enforce structure-preserving refinement across intermediate denoising states. Moreover, our single-step deterministic inference yields up to faster runtime while training on relatively small datasets.Extensive experiments demonstrate comprehensive state-of-the-art results across all three tasks, revealing the strong potential of editing-oriented diffusion transformers for geometry-aware perception.

Generating dynamic 4D objects from sparse inputs is difficult because it demands joint preservation of appearance and motion coherence across views and time while suppressing artifacts and temporal drift. We hypothesize that the view discrepancy arises from supervision limited to pixel- or latent-space video-diffusion losses, which lack explicitly temporally aware, feature-level tracking guidance.We present Track4DGen, a two-stage framework that couples a multi-view video diffusion model with a foundation point tracker and a hybrid 4D Gaussian Splatting (4D-GS) reconstructor. The central idea is to explicitly inject tracker-derived motion priors into intermediate feature representations for both multi-view video generation and 4D-GS. In Stage One, we enforce dense, feature-level point correspondences inside the diffusion generator, producing temporally consistent features that curb appearance drift and enhance cross-view coherence. In Stage Two, we reconstruct a dynamic 4D-GS using a hybrid motion encoding that concatenates co-located diffusion features (carrying Stage-One tracking priors) with Hex-plane features, and augment them with 4D Spherical Harmonics for higher-fidelity dynamics modeling.Track4DGen surpasses baselines on both multi-view video generation and 4D generation benchmarks, yielding temporally stable, text-editable 4D assets. Lastly, we curate Sketchfab28, a high-quality dataset for benchmarking object-centric 4D generation and fostering future research.

The rapid growth of the text-to-image (T2I) community has fostered a thriving online ecosystem of expert models, which are variants of pretrained diffusion models specialized for diverse generative capabilities. Yet, existing model merging methods remain limited in fully leveraging abundant online expert resources and still struggle to meet diverse in-the-wild user needs. We present DiffGraph, a novel automated agent-driven graph-based model merging framework, which automatically harnesses online experts and flexibly merges them for diverse user needs. Our DiffGraph constructs a scalable graph and organizes ever-expanding online experts within it through node registration and calibration. Then, DiffGraph dynamically activates specific subgraphs based on user needs, enabling flexible combinations of different experts to achieve user-desired generation. Extensive experiments show the efficacy of our method.

While existing generation and unified models excel at general image generation, they struggle with tasks requiring deep reasoning, planning, and precise data-to-visual mapping abilities beyond general scenarios. To push beyond the existing limitations, we introduce a new and challenging task: creative table visualization, requiring the model to generate an infographic that faithfully and aesthetically visualizes the data from a given table. To address this challenge, we propose ShowTable, a pipeline that synergizes MLLMs with diffusion models via a progressive self-correcting process. The MLLM acts as the central orchestrator for reasoning the visual plan and judging visual errors to provide refined instructions, the diffusion execute the commands from MLLM, achieving high-fidelity results. To support this task and our pipeline, we introduce three automated data construction pipelines for training different modules. Furthermore, we introduce TableVisBench, a new benchmark with 800 challenging instances across 5 evaluation dimensions, to assess performance on this task. Experiments demonstrate that our pipeline, instantiated with different models, significantly outperforms baselines, highlighting its effective multi-modal reasoning, generation, and error correction capabilities.