CVPR 2026 论文列表

Diffusion models face a fundamental trade-off between generation quality and computational efficiency. Latent Diffusion Models (LDMs) offer an efficient solution but suffer from potential information loss and non-end-to-end training. In contrast, existing pixel space models bypass VAEs but are computationally prohibitive for high-resolution synthesis. To resolve this dilemma, we propose DiP, an efficient pixel space diffusion framework. DiP decouples generation into a global and a local stage: a Diffusion Transformer (DiT) backbone operates on large patches for efficient global structure construction, while a co-trained lightweight Patch Detailer Head leverages contextual features to restore fine-grained local details. This synergistic design achieves computational efficiency comparable to LDMs without relying on a VAE. DiP is accomplished with up to 10xfaster inference speeds than previous method while increasing the total number of parameters by only 0.3%, and achieves an 1.79 FID score on ImageNet 256x256.

Condensing the large-scale, high-resolution ImageNet-1K dataset remains a challenge for dataset distillation (DD). Existing methods typically match batch normalization (BN) statistics, i.e., mean and variance, between real and synthetic datasets. Although effective with soft labels, their performance degrades substantially under hard labels. In this paper, we theoretically identify that BN matching mainly aligns the scales of real and synthetic gradients but overlooks their directions. However, experimental evidence demonstrates that gradient direction, rather than scale, is pivotal to model training, clarifying the limitations of prior methods. Building on this insight, we introduce Orthogonal Gradient Matching (OGM), which explicitly aligns the intrinsic direction of gradients, i.e., singular vectors. Specifically, OGM first orthogonalizes real and synthetic gradients by setting all singular values to one, eliminating their scales, and then minimizes the distance between these orthogonal gradients so that their singular vectors coincide. To further reduce computation, OGM employs a least-squares loss whose gradients can be obtained in the forward pass, avoiding back-propagation. Extensive experiments on ImageNet-1K validate the effectiveness of OGM. With only ten images per class (IPC = 10), OGM achieves 47.0% accuracy with soft labels and 16.7% with hard labels, outperforming training-based DD methods and RDED.

Multi-view diffusion models have recently emerged as a powerful paradigm for novel view synthesis, yet the underlying mechanism that enables their view consistency remains unclear. In this work, we first verify that the attention maps of these models acquire geometric correspondence throughout training, attending to the geometrically corresponding regions across reference and target views for view-consistent generation. However, this correspondence signal remains incomplete, with its accuracy degrading under large viewpoint changes. Building on these findings, we introduce CAMEO, a simple yet effective training technique that directly supervises attention maps using geometric correspondence to enhance both the training efficiency and generation quality of multi-view diffusion models. Notably, supervising a single attention layer is sufficient to guide the model toward learning precise correspondences, thereby preserving the geometry and structure of reference images, accelerating convergence, and improving novel view synthesis performance. CAMEO reduces the number of training iterations required for convergence by half while achieving superior performance at the same iteration counts. We further demonstrate that CAMEO is model-agnostic and can be applied to any multi-view diffusion model. Project page is available at https://cvlab-kaist.github.io/CAMEO/.

Denoising-based diffusion transformers, despite their strong generation performance, suffer from inefficient training convergence. Existing methods addressing this issue, such as REPA (relying on external representation encoders) or SRA (requiring dual-model setups), inevitably incur heavy computational overhead during training due to external dependencies. To tackle these challenges, this paper proposes SRA 2, a lightweight intrinsic self-representation alignment framework for efficient diffusion training. SRA 2 leverages off-the-shelf pre-trained Variational Autoencoder (VAE) features: their reconstruction property ensures inherent encoding of visual priors like rich texture details, structural patterns, and basic semantic information. Specifically, SRA 2 aligns the intermediate latent features of diffusion transformers with VAE features via a lightweight projection layer, supervised by a feature alignment loss. This design accelerates training without extra representation encoders or dual-model maintenance, resulting in a simple yet effective pipeline. Extensive experiments demonstrate that SRA 2 improves both generation quality and training convergence speed compared to vanilla diffusion transformers, matches or outperforms state-of-the-art acceleration methods, and incurs merely 4% extra GFLOPs with zero additional cost for external guidance models.

With the rapid progress of controllable generation, training data synthesis has become a promising way to expand labeled datasets and alleviate manual annotation in remote sensing (RS). However, the complexity of semantic mask control and the uncertainty of sampling quality often limit the utility of synthetic data in downstream semantic segmentation tasks. To address these challenges, we propose a task-oriented data synthesis framework (TODSynth), including a Multimodal Diffusion Transformer (MM-DiT) with unified triple attention and a plug-and-play sampling strategy guided by task feedback. Built upon the powerful DiT-based generative foundation model, we systematically evaluate different control schemes, showing that a text-image-mask joint attention scheme combined with full fine-tuning of the image and mask branches significantly enhances the effectiveness of RS semantic segmentation data synthesis, particularly in few-shot and complex-scene scenarios. Furthermore, we propose a control-rectify flow matching (CRFM) method, which dynamically adjusts sampling directions guided by semantic loss during the early high-plasticity stage, mitigating the instability of generated images and bridging the gap between synthetic data and downstream segmentation tasks. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art controllable generation methods, producing more stable and task-oriented synthetic data for RS semantic segmentation.

Generating complete digital twins from videos requires precise camera control, global scene coverage, and strict spatial-temporal consistency--constraints that remain challenging for perspective video generators due to their limited field of view (FoV). Their narrow FoV forces long or multi-view trajectories, amplifying cross-view inconsistency and temporal drift.We argue that 360deg video generation offers a natural solution: panoramic coverage simplifies trajectory design and provides strong global context for maintaining coherence. We introduce Pantheon360: Taming Digital Twin Generation via 3D-Aware 360deg Video Diffusion, a controllable 360deg video generation framework that synthesizes high-fidelity videos from sparse 360deg inputs. The key idea is an explicit 3D Cache, reconstructed from the input, which serves as a geometric scaffold for any user-defined camera path. This allows the diffusion model to focus on photorealistic texture refinement while the 3D Cache enforces global geometric consistency.Experiments show that Pantheon360 achieves superior visual quality and unmatched geometric coherence, enabling reliable and flexible 360deg scene generation for downstream simulation and digital-twin applications.

Causality--referring to temporal, uni-directional cause-effect relationships between components--underlies many complex generative processes, including videos, language, and robot trajectories.Current causal diffusion models entangle temporal reasoning with iterative denoising, applying causal attention across all layers, at every denoising step, and over the entire context.In this paper, we show that the causal computation in these models is separable from the multi-step denoising process.Through systematic probing of autoregressive video diffusers, we uncover two key regularities:(1) early blocks produce highly similar features across denoising steps, indicating redundant computation along the diffusion trajectory; and(2) deeper blocks exhibit sparse cross-frame attention and primarily perform intra-frame rendering.Motivated by these findings, we introduce Separable Causal Diffusion (SCD), a new architecture that explicitly decouples once-per-frame temporal reasoning, via a causal transformer encoder, from multi-step frame-wise rendering, via a lightweight diffusion decoder.Extensive experiments on both pretraining and post-training tasks across synthetic and real benchmarks show that CSD significantly improves throughput and latency while matching or surpassing the generation quality of strong causal diffusion baselines.

Text-to-image diffusion models excel at generating high-quality, diverse images from natural language prompts. However, they often fail to produce semantically accurate results when the prompt contains concept combinations that contradict their learned priors. We define this failure mode as contextual contradiction, where one concept implicitly negates another due to entangled associations learned during training. To address this, we propose a stage-aware prompt decomposition framework that guides the denoising process using a sequence of proxy prompts. Each proxy prompt is constructed to match the semantic content expected to emerge at a specific stage of denoising, while ensuring contextual coherence. To construct these proxy prompts, we leverage a large language model (LLM) to analyze the target prompt, identify contradictions, and generate alternative expressions that preserve the original intent while resolving contextual conflicts. By aligning prompt information with the denoising progression, our method enables fine-grained semantic control and accurate image generation in the presence of contextual contradictions. Experiments across a variety of challenging prompts show substantial improvements in alignment to the textual prompt.

Pruning-based unlearning has recently emerged as a fast, training-free, and data-independent approach to remove undesired concepts from diffusion models. It promises high efficiency and robustness, offering an attractive alternative to traditional fine-tuning or editing-based unlearning. However, in this paper we uncover a hidden danger behind this promising paradigm. We find that the locations of pruned weights, typically set to zero during unlearning, can act as side-channel signals that leak critical information about the erased concepts.To verify this vulnerability, we design a novel attack framework capable of reviving erased concepts from pruned diffusion models in a fully data-free and training-free manner. Our experiments confirm that pruning-based unlearning is not inherently secure, as erased concepts can be effectively revived without any additional data or retraining.Finally, we explore potential defense strategies and advocate safer pruning mechanisms that conceal pruning locations while preserving unlearning effectiveness, providing practical insights for designing more secure pruning-based unlearning frameworks.

Driving planning is a critical component of end-to-end (E2E) autonomous driving. However, prevailing Imitative E2E Planners often suffer from multimodal trajectory mode collapse, failing to produce diverse trajectory proposals. Meanwhile, Generative E2E Planners struggle to incorporate crucial safety and physical constraints directly into the generative process, necessitating an additional optimization stage to refine their outputs. In this paper, we propose GuideFlow, a novel planning framework that leverages Constrained Flow Matching. Concretely, GuideFlow explicitly models the flow matching process, which inherently mitigates mode collapse and allows for flexible guidance from various conditioning signals. Our core contribution lies in directly enforcing explicit constraints within the flow matching generation process, rather than relying on implicit constraint encoding. Crucially, GuideFlow unifies the training of the flow matching with the Energy-Based Model (EBM) to enhance the model's autonomous optimization capability to robustly satisfy physical constraints. Secondly, GuideFlow parameterizes driving aggressiveness as a control signal during generation, enabling precise manipulation of trajectory style. Extensive evaluations on major driving benchmarks (Bench2Drive, NuScenes, NavSim and ADV-NuScenes) validate the effectiveness of GuideFlow. Notably, on the NavSim test hard split (Navhard), GuideFlow achieved SOTA with an EPDMS score of 43.0. The code will be released in https://github.com/adept-thu/GuideFlow.

Recently, great progress has been achieved in text-to-video (T2V) generation by scaling transformer-based diffusion models to billions of parameters, which can generate high-quality videos. However, existing models typically produce only short clips offline, restricting their use cases in interactive and real-time applications. This paper addresses these challenges by proposing StreamDiT, a streaming video generation model. StreamDiT training is based on flow matching by adding a moving buffer. We design mixed training with different partitioning schemes of buffered frames to boost both content consistency and visual quality. StreamDiT modeling is based on adaLN DiT with varying time embedding and window attention. To practice the proposed method, we train a StreamDiT model with 4B parameters. In addition, we propose a multistep distillation method tailored for StreamDiT. Sampling distillation is performed in each segment of a chosen partitioning scheme. After distillation, the total number of function evaluations (NFEs) is reduced to the number of chunks in a buffer. Finally, our distilled model reaches real-time performance at 16 FPS on one GPU, which can generate video streams at 512p resolution. We evaluate our method through both quantitative metrics and human evaluation. Our model enables real-time applications, e.g. streaming generation, interactive generation, and video-to-video.

In controllable driving-scene reconstruction and 3D scene generation, maintaining geometric fidelity while synthesizing visually plausible appearance under large viewpoint shifts is crucial. However, effective fusion of geometry-based 3DGS and appearance-driven diffusion models faces inherent challenges, as the absence of pixel-wise, 3D-consistent editing criteria often leads to over-restoration and geometric drift. To address these issues, we introduce **FaithFusion**, a 3DGS-diffusion fusion framework driven by pixel-wise Expected Information Gain (EIG). EIG acts as a unified policy for coherent spatio-temporal synthesis: it guides diffusion as a spatial prior to refine high-uncertainty regions, while its pixel-level weighting distills the edits back into 3DGS. The resulting plug-and-play system is free from extra prior conditions and structural modifications. Extensive experiments on the Waymo dataset demonstrate that our approach attains SOTA performance across NTA-IoU, NTL-IoU, and FID, maintaining an FID of 107.47 even at 6 meters lane shift.

Diffusion bridge models offer a powerful framework for connecting two data distributions, such as in image restoration and translation. Many existing methods learn this bridge by mimicking the score-matching formulation of standard diffusion models. In this work, we find that this way leads to an anomalous underfitting phenomenon near the target endpoint, as the process approaches the target distribution (t \to 0). This underfitting, characterized by significant drift in the predicted variance and direction, results from an excessively large discrepancy in noise levels between the network's input and its regression target. To resolve this issue, we propose the Noise-Aligned Diffusion Bridge (NADB). Our approach reformulates the diffusion bridge by first employing a mean network to provide a cleaner conditional target, and then introducing a novel, noise-aligned mapping relationship. This new formulation resolves the noise mismatch and corrects the underfitting near the target endpoint.Experimental validation across multiple image restoration and image translation tasks demonstrates the effectiveness of our approach. Code is available at https://github.com/gyr02/NADB.

Despite remarkable progress in text-to-image diffusion models, controlling the semantic and spatial relationships between interacting instances remains a fundamental challenge. Current methods that inject spatial constraints often fail to model the intrinsic functional dependencies between entities, leading to implausible interactions. In this paper, we introduce Semantic Derivative Flow (SDF), a novel graph-guided framework that structures the diffusion process within a directed acyclic interaction graph. Our core innovation is a theoretically-motivated derivative attention mechanism, which explicitly enforces the semantic representation of a predicate to be derived from its subject, and the object from the predicate, formalizing a differentiable semantic graph. This principled approach compels the generative process to adhere to the logical chain of interaction. We further integrate a global context node and a real-time regional refinement module to ground the graph in the visual domain holistically. Extensive experiments demonstrate that our model, an instantiation of SDF, establishes a new state-of-the-art in fidelity and controllability on the HICODet benchmark. We complement our empirical results with a theoretical analysis, framing our method as structured message passing on interaction graphs, which provides a rigorous justification for its efficacy and generalization benefits.

Range-view (RV) based LiDAR diffusion has recently made huge strides towards 2D photo-realism. However, it neglects 3D geometry realism and often generates various RV artifacts such as depth bleeding and wavy surfaces. We design L3DR, a 3D-aware LiDAR Diffusion and Rectification framework that can regress and cancel RV artifacts in 3D space and restore local geometry accurately. Our theoretical and empirical analysis reveals that 3D models are inherently superior to 2D models in generating sharp and authentic boundaries. Leveraging such analysis, we design a 3D residual regression network that rectifies RV artifacts and achieves superb geometry realism by predicting point-level offsets in 3D space. On top of that, we design a Welsch Loss that helps focus on local geometry and ignore anomalous regions effectively. Extensive experiments over multiple benchmarks including KITTI, KITTI360, nuScenes and Waymo show that the proposed L3DR achieves state-of-the-art generation and superior geometry-realism consistently. In addition, L3DR is generally applicable to different LiDAR diffusion models with little computational overhead.

We present ShapeAR, a novel autoregressive latent diffusion framework that decomposes raster images into editable, artist-like vector shape layers. Unlike conventional raster-to-SVG methods that rely on boundary tracing or joint path optimisation, ShapeAR generates non-overlapping RGBA shape layers directly in latent space via flow-matching diffusion. To scale generation to complex scenes with many shapes, we formulate the process auto-regressively, conditioning each step on both the input image (global context) and partial composition of previously generated layers (local context). In addition, we propose geometry-aware evaluation metrics that quantify aesthetic and structural quality of the generated shapes.

We present Gen3R, a method that bridges the strong priors of foundational reconstruction models and video diffusion models for scene-level 3D generation. We repurpose the VGGT reconstruction model to produce geometric latents by training an adapter on its tokens, which are regularized to align with the appearance latents of pre-trained video diffusion models. By jointly generating these disentangled yet aligned latents, Gen3R produces both RGB videos and corresponding 3D geometry, including camera poses, depth maps, and global point clouds. Experiments demonstrate that our approach achieves state-of-the-art results in single- and multi-image conditioned 3D scene generation. Additionally, our method can enhance the robustness of reconstruction by leveraging generative priors, demonstrating the mutual benefit of tightly coupling reconstruction and generative models.

We address text-based 3D human motion editing, where the goal is to preserve the style and structure of a source motion while applying edits described in natural language. The release of the MotionFix dataset has spurred active research into training-based diffusion models that directly generate an edited motion from a source motion and a text instruction. While previous works have focused primarily on learning when an edit should occur temporally, our goal is to create a model that understands not only this temporal aspect but also which specific joints are responsible for the change. Targeting this, we propose a novel architecture and a complementary auxiliary task to aid its training. Our architecture consists of two axis-anchored transformers, which extract distinct features along the joint and time dimensions respectively, and a cross-axis fusion block that integrates these representations. We further introduce an auxiliary task that trains the joint-anchored transformer to regress the Soft-DTW distance between source and target joint rotations. This objective teaches the module to understand which joints to modify and which to preserve. Through comprehensive experiments on the MotionFix dataset, we demonstrate that our method significantly improves semantic alignment with both the text instruction and the source motion, as well as the overall fidelity of the generated motion, achieving state-of-the-art results.

The new era has witnessed a remarkable capability to extend Vision-Language Models (VLMs) for tackling tasks of video understanding. While current VLMs excel at event- or story-level understanding, their ability to capture fine-grained motion details remains limited, primarily due to their focus on high-level static semantic structures and macro-event logic. In contrast, Video Diffusion Models (VDMs) are adept at modeling dynamic motion patterns, benefiting from large-scale video data and the intrinsic requirement of temporal generation. In this paper, we introduce MotionEnhancer, a novel approach that leverages motion priors distilled from a powerful video diffusion model as auxiliary supervision to enhance the motion understanding capability of a VLM via attention alignment. MotionEnhancer comprises two simple parameter-free modules, Motion-sensitive Head Selection (MHS) and Motion-salient Text Token Identification (MTTI), to directly extract and optimize motion-related attentions from the VDM in a computation-only manner. MotionEnhancer provides a scalable solution for motion understanding without additional training parameters, modifications to existing architectures, or tool calling. Extensive experiments demonstrate that MotionEnhancer can achieve consistent improvements over state-of-the-art VLMs on two motion-level video understanding benchmarks, especially on motion-related metrics.