CVPR 2026 论文列表

Latent diffusion has become the default paradigm for visual generation, yet we observe a persistent reconstruction-generation trade-off as latent dimensionality increases: higher-capacity autoencoders improve reconstruction fidelity but generation quality eventually declines. We trace this gap to the different behaviors in high-frequency tokenization and detokenization. Through controlled perturbations in both RGB and latent domains, we analyze encoder/decoder behaviors and find that decoders depend strongly on high-frequency latent components to recover details, whereas encoders under-represent high-frequency contents, yielding insufficient exposure and underfitting in high-frequency bands for diffusion model training. To address this issue, we introduce FreqWarm, a plug-and-play frequency warm-up curriculum that increases early-stage exposure to high-frequency latent signals during diffusion or flow-matching training -- without modifying or retraining the autoencoder. Applied across several high-dimensional tokenizers, FreqWarm consistently improves generation quality: decreasing gFID by 14.11 on Wan2.2-VAE, 6.14 on LTX-VAE, and 4.42 on DC-AE-f32, while remaining architecture-agnostic and compatible with diverse backbones. Our study shows that explicitly managing frequency exposure can successfully turn high-dimensional latent spaces into more diffusible targets.

Image tokenization has significantly advanced visual generation and multimodal modeling, particularly when paired with autoregressive models. However, current methods face challenges in balancing efficiency and quality: high-resolution image generation either requires an excessive number of tokens or compromises critical details through token reduction. To resolve this, we propose Latent Consistency Tokenizer (LacTok) that bridges discrete visual tokens with the compact latent space of pretrained latent diffusion models (LDMs), enabling efficient representation of 1024x1024 images using only 256 tokens--a 16xcompression over VQGAN. LacTok integrates a transformer encoder, a quantized codebook, and a latent consistency decoder. Direct application of LDM as the decoder results in color and brightness discrepancies; thus, we convert it to latent consistency decoder, reducing multi-step sampling to 1-2 steps for direct pixel-level supervision. To endow LacTok with text-to-image generation capabilities, we seamlessly integrate it with an autoregressive transformer, forming LacTokGen. This transformer is trained by predicting compact token sequences conditioned on text instructions. Experiments demonstrate LacTok's superiority in high-fidelity reconstruction, with 10.8 reconstruction Frechet Inception Distance on MSCOCO-2017 5K benchmark for 1024x1024 image reconstruction. LacTokGen achieves 0.73 score on GenEval benchmark and 0.304 HPSv2 on MSCOCO-2017 dataset.

Optical imaging systems are inherently imperfect due to diffraction limits, lens manufacturing tolerances, assembly misalignment, and other physical constraints. In addition, unavoidable camera shake and object motion further introduce non-ideal degradations during acquisition. These aberrations and motion-induced variations are typically unknown, difficult to measure, and costly to model or calibrate in practice. Blind inverse problems offer a promising direction by jointly estimating both the latent image and the unknown degradation kernel. However, existing approaches often suffer from convergence instability, limited prior expressiveness, and sensitivity to hyperparameters. Inspired by recent advances in self-diffusion, we propose DeblurSDI, a zero-shot, self-supervised blind imaging framework that requires no pre-training. DeblurSDI formulates blind image recovery as an iterative reverse self-diffusion process that begins from pure noise and progressively refines both the sharp image and the blur kernel. Extensive experiments on combined optical aberrations and motion blur demonstrate that DeblurSDI consistently outperforms other methods by a substantial margin.

Existing humanoid control systems often rely on teleoperation or modular generation pipelines that separate language understanding from physical execution. However, the former is entirely human-driven, and the latter lacks tight alignment between language commands and physical behaviors. In this paper, we present SENTINEL, a fully end-to-end language-action framework for humanoid whole-body control. We construct a large-scale dataset by tracking human motions in simulation using a pretrained whole body controller, combined with their text annotations. The model directly maps language commands and proprioceptive inputs to low-level actions without any intermediate representation. The model generates action chunks using flow matching, which can be subsequently refined by a residual action head for real-world deployment. Our method exhibits strong semantic understanding and stable execution on humanoid robots in both simulation and real-world deployment, and also supports multi-modal extensions by converting inputs into texts.

Despite the impressive performance of diffusion models such as Stable Diffusion (SD) in image generation, their slow inference limits practical deployment. Recent works accelerate inference by distilling multi-step diffusion into one-step generators. To better understand the distillation mechanism, we analyze U-Net/DiT weight changes between one-step students and their multi-step teacher counterparts. Our analysis reveals that changes in weight direction significantly exceed those in weight norm, highlighting it as the key factor during distillation. Motivated by this insight, we propose the Low-rank Rotation of weight Direction (LoRaD), a parameter-efficient adapter tailored to one-step diffusion distillation. LoRaD is designed to model these structured directional changes using learnable low-rank rotation matrices. We further integrate LoRaD into Variational Score Distillation (VSD), resulting in Weight Direction-aware Distillation (WaDi)--a novel one-step distillation framework. WaDi achieves state-of-the-art FID scores on COCO 2014 and COCO 2017 while using only approximately 10% of the trainable parameters of the U-Net/DiT. Furthermore, the distilled one-step model demonstrates strong versatility and scalability, generalizing well to various downstream tasks such as controllable generation, relation inversion, and high-resolution synthesis.

Diffusion-based methods have shown great promise in single image super-resolution (SISR); however, existing approaches often produce blurred fine details due to insufficient guidance in the high-frequency domain. To address this issue, we propose a High-Frequency Guided Diffusion Network based on Wavelet Decomposition (HDW-SR), which replaces the conventional U-Net backbone in diffusion frameworks. Specifically, we perform diffusion only on the residual map, allowing the network to focus more effectively on high-frequency information restoration. We then introduce wavelet-based downsampling in place of standard CNN downsampling to achieve multi-scale frequency decomposition, enabling sparse cross-attention between the high-frequency subbands of the pre-super-resolved image and the low-frequency subbands of the diffused image for explicit high-frequency guidance. Moreover, a Dynamic Thresholding Block (DTB) is designed to refine high-frequency selection during the sparse attention process. During upsampling, the invertibility of the wavelet transform ensures low-loss feature reconstruction. Experiments on both synthetic and real-world datasets demonstrate that HDW-SR achieves competitive super-resolution performance, excelling particularly in recovering fine-grained image details. Code is available at https://github.com/Baty2023/HDW-SR.

Diffusion models achieve state-of-the-art video generation quality, but their inference remains expensive due to the large number of sequential denoising steps. This has motivated a growing line of research on accelerating diffusion inference. Among training-free acceleration methods, caching reduces computation by reusing previously computed model outputs across timesteps. Existing caching methods rely on heuristic criteria to choose cache/reuse timesteps and require extensive tuning. We address this limitation with a principled sensitivity-aware caching framework. Specifically, we formalize the caching error through an analysis of the model output sensitivity to perturbations in the denoising inputs, i.e., the noisy latent and the timestep, and show that this sensitivity is a key predictor of caching error. Based on this analysis, we propose Sensitivity-Aware Caching (SenCache), a dynamic caching policy that adaptively selects caching timesteps on a per-sample basis. Our framework provides a theoretical basis for adaptive caching, explains why prior empirical heuristics can be partially effective, and extends them to a dynamic, sample-specific approach. Experiments on Wan 2.1, CogVideoX, and LTX-Video show that SenCache achieves better visual quality than existing caching methods under similar computational budgets.

Diffusion transformers have demonstrated remarkable generation quality, albeit requiring longer training iterations and numerous inference steps. In each denoising step, diffusion transformers encode the noisy inputs to extract the lower-frequency semantic component and then decode the higher frequency with identical modules. This scheme creates an inherent optimization dilemma: encoding low-frequency semantics necessitates reducing high-frequency components, creating tension between semantic encoding and high-frequency decoding. To resolve this challenge, we propose a new \color ddtD ecoupled \color ddtD iffusion \color ddtT ransformer(\color ddtDDT ), with a decoupled design of a dedicated condition encoder for semantic extraction alongside a specialized velocity decoder. Our experiments reveal that a more substantial encoder yields performance improvements as model size increases. For ImageNet 256x256, Our DDT-XL/2 achieves a new state-of-the-art performance of 1.31 FID (nearly 4xfaster training convergence compared to previous diffusion transformers). For ImageNet 512x512, Our DDT-XL/2 achieves a new state-of-the-art FID of 1.28. Additionally, as a beneficial by-product, our decoupled architecture enhances inference speed by enabling the sharing self-condition between adjacent denoising steps. To minimize performance degradation, we propose a novel statistical dynamic programming approach to identify optimal sharing strategies.

We propose a federated framework for training diffusion models on decentralized and private datasets. The method learns a shared generative model alongside personalized client models, allowing clients to benefit from cross-client structure while ensuring that the shared model cannot reproduce any client's data on its own. We provide formal differential privacy guarantees for each client and establish utility bounds for conditional generation under a Gaussian mixture model, showing that collaboration improves sample quality relative to private non-collaborative training. Experiments on CIFAR-10, Colorized MNIST, and CelebA support these results: the method generates high-fidelity samples, improves performance on minority and underrepresented classes, and maintains strong protection against membership inference, memorization, and reconstruction attacks.

Caching-based acceleration methods have recently driven significant progress in efficient video generation with diffusion models. However, we identify a critical limitation when directly applying these acceleration techniques to auto-regressive video diffusion models, which generate long videos by sequentially synthesizing segments conditioned on historical context. In such settings, any approximation errors introduced by acceleration tend to propagate and accumulate over time, resulting in severe error accumulation and progressive degradation of video quality. To address this challenge, we propose ARCache, the first training-free caching-based acceleration framework specifically designed for auto-regressive video diffusion models. ARCache improves both the timing and quality of caching through two key components. First, History-Guided Cache (HGC) leverages historical information to adaptively schedule caching for each segment, enabling more accurate and efficient cache utilization. Second, Enhanced Residual Correction (ERC) adaptively refines the residual trajectory for subsequent segments, effectively mitigating error accumulation while introducing minimal computational overhead. Extensive experiments on FramePack-F1, SkyReels-V2, and auto-regressive world model Matrix-Game demonstrate that ARCache achieves state-of-the-art acceleration and visual fidelity.

Even when told to "do nothing," modern diffusion models subtly alter their output relative to the input they are supposed to preserve. We call this effect No-Op Drift. We introduce the Drift Kernel K_M(sigma), defined as the expected perceptual deviation induced when running a diffusion model at noise strength sigma under a null instruction. Using 120,000 baseline samples (30,000 per model across SD15, SD21, SDXL, and InstructPix2Pix) and 9,600 ablation samples (four strengths, null versus strict copy prompts), we show that variance-driven diffusion models follow a quadratic form K_M(sigma) approximately equal to k_M sigma^2 plus c_M, with aggregate R^2 equal to 0.97. We derive this scaling from first principles via a Taylor expansion of the decoder, yielding k_M equal to Tr(J_D J_D^T), which depends only on the decoder Jacobian and not on prompts. To validate mechanistic structure, we construct synthetic decoders that reproduce the two regimes seen in practice: quadratic variance-driven drift and flat, high-variance edit-driven drift. We show that prompt wording has negligible effect (less than 17 percent coefficient difference), proving that drift is structural and not prompt-induced. We release NoOp-Bench, a benchmark with 10,000 inputs and full code for reproducible kernel estimation. Additional proofs, ablations, LPIPS and CLIP metrics, and extended visualizations appear in the Supplementary Material.

Existing methods following the integrated hard-regression or decoupling optimization paradigms exhibit limited fusion performance under complex degradations. To address these paradigm-level shortcomings, we propose ReCoFuse, an ultra-robust image fusion framework based on restorative multi-modal diffusion reciprocal coupling. ReCoFuse redefines the relationship between information restoration and integration, deriving a novel reciprocal coupling optimization paradigm through their mutual reinforcement. It first constructs two restoration branches using diffusion modules (DiM) to capture modality-specific restoration priors. Then, time-aware cross-modal integration modules (TIM) are introduced as a bridge to couple restoration and integration, embedded at each DiM sampling timestep to aggregate multi-modal information. The aggregated variable not only feeds back to each restoration branch to enhance degradation removal via cross-modal complementarity, but also generates high-quality fused images that comprehensively represent the scene. Moreover, an alternating regularization mechanism is designed to iteratively optimize DiM and TIM along the gradient path, ensuring effective collaboration between restoration and integration. Extensive experiments show that ReCoFuse achieves state-of-the-art performance under challenging degradations such as low light, haze, noise, low contrast, and stripes. The code is publicly available at https://github.com/HaoZhang1018/ReCoFuse.

Ocean dynamics drive global climate patterns and extreme weather events, making accurate spatiotemporal forecasting essential for climate monitoring and marine operations. Traditional Global Ocean Forecasting Systems (GOFSs) offer high accuracy predictions, yet remain computationally expensive and fail to fully leverage growing historical data. Recent deep learning models have achieved notable success, but still face three fundamental challenges: (1) they homogenize ocean variables despite strong physical coupling via equation-of-state relationships; (2) they neglect spherical geometry, resulting in severe distortions at high latitudes; and (3) they struggle to model multi-scale temporal dynamics. We introduce PhyOceanCast, a physics-informed diffusion model that overcomes these limitations through two key innovations. First, the Spherical Graph Attention Network for Multi-scale Ocean Coupling (SGAN-MOC) preserves spherical topology while enabling cross-variable interactions via heterogeneous encoding and k-hop-constrained attention. Second, the Physics-Informed Wavelet Temporal Coherence (PWTC) module that decomposes ocean dynamics across multiple scales with advection-diffusion constraints. PhyOceanCast forecasts 145 ocean variables, including temperature, salinity, and velocity fields, across 36 depth levels plus sea surface height. Extensive experiments demonstrate superior performance over diffusion, transformer, and hybrid baselines, promising a new paradigm for global ocean canonical variable forecasting.

As a key technique in multi-modal processing, infrared and visible image fusion (IVIF) plays a crucial role in integrating complementary spectral information for visual enhancement and downstream vision tasks. Despite remarkable progress, existing methods struggle to flexibly accommodate heterogeneous demands. Achieving adaptive fusion that aligns with various preferences from both human and machine vision remains an open and challenging problem. To address this challenge, we propose DPOFusion, a direct preference optimization (DPO) framework integrating the property-aligned latent diffusion model (PALDM) and the preference-controllable latent diffusion model (PCLDM), enabling task-guided, preference-adaptive IVIF for both human and machine vision. The PALDM leverages a latent fusion prior and a joint conditional loss to generate diverse candidate fusion results with various properties. PCLDM is subsequently fine-tuned via instance direct preference optimization (IDPO), enabling direct control of the final fusion results with heterogeneous preference signals.Experimental results demonstrate that our framework not only attains precise preference alignment among humans, vision-language models, and task-driven networks, but also sets a new benchmark for adaptive fusion quality and task-oriented transferability. Our code will be available at https://github.com/suweijian1996/DPOFusion

We present NaTex, a native texture generation framework that predicts texture color directly in 3D space. In contrast to previous approaches that rely on baking 2D multi-view images synthesized by geometry-conditioned Multi-View Diffusion models (MVDs), NaTex avoids several inherent limitations of the MVD pipeline. These include difficulties in handling occluded regions that require inpainting, achieving precise mesh-texture alignment along boundaries, and maintaining cross-view consistency and coherence in both content and color intensity. NaTex features a novel paradigm that addresses the aforementioned issues by viewing texture as a dense color point cloud. Driven by this idea, we propose latent color diffusion, which comprises a geometry-awared color point cloud VAE and a multi-control diffusion transformer (DiT), entirely trained from scratch using 3D data, for texture reconstruction and generation. To enable precise alignment, we introduce native geometry control that conditions the DiT on direct 3D spatial information via positional embeddings and geometry latents. We co-design the VAE-DiT architecture, where the geometry latents are extracted via a dedicated geometry branch tightly coupled with the color VAE, providing fine-grained surface guidance that maintains strong correspondence with the texture. With these designs, NaTex demonstrates strong performance, significantly outperforming previous methods in texture coherence and alignment. Moreover, NaTex also exhibits strong generalization capabilities, either training-free or with simple tuning, for various downstream applications, e.g., material generation, texture refinement, and part segmentation and texturing.

Large-scale video diffusion models exhibit strong world-simulation and temporal reasoning capabilities, yet their potential as zero-shot image editors remains underexplored. We present \ifedit IF-Edit (Image Edit by Generating Frames), a tuning-free framework that repurposes pre-trained image-to-video diffusion models for instruction-driven image editing. \ifedit IF-Edit addresses three core obstacles--prompt misalignment, redundant temporal latents, and blurry late-stage frames--via: (1) a Chain-of-Thought Prompt Enhancement module that reformulates static editing instructions into temporally grounded reasoning prompts; (2) a Temporal Latent Dropout strategy that compresses frame latents after the expert-switch point, accelerating denoising while preserving global semantics and temporal coherence; and (3) a Self-Consistent Post-Refinement step that refines the sharpest late-stage frame through a brief still-video trajectory, leveraging the video prior for sharper and more faithful results. Extensive experiments across four public benchmarks--covering non-rigid deformations, physical and temporal reasoning, and general instruction editing--show that \ifedit IF-Edit achieves strong performance on non-rigid and reasoning-centric tasks while remaining competitive on general-purpose edits. Our study offers a systematic view of video diffusion models as image editors, revealing their unique strengths, limitations, and a simple recipe for unified video-image generative reasoning.

Recent progress in multimodal large language models (MLLMs) has highlighted the challenge of efficiently bridging pre-trained Vision-Language Models (VLMs) with Diffusion Models. While methods using a fixed number of learnable query tokens offer computational efficiency, they suffer from task generalization collapse, failing to adapt to new tasks that are distant from their pre-training tasks. To overcome this, we propose Noisy Query Tokens, which learn a distributed representation space between the VLM and Diffusion Model via end-to-end optimization, enhancing continual learning. Additionally, we introduce a VAE branch with linear projection to recover fine-grained image details. Experimental results confirm our approach mitigates generalization collapse and enables stable continual learning across diverse tasks.

MeanFlow (MF) is a diffusion-motivated generative model that enables efficient few-step generation by learning long jumps directly from noise to data. In practice, it is often used as a latent MF by leveraging the pre-trained Stable Diffusion variational autoencoder (SD-VAE) for high-dimensional data modeling. However, MF training remains computationally demanding and is often unstable. During inference, the SD-VAE decoder dominates the generation cost, and MF depends on complex guidance hyperparameters for class-conditional generation. In this work, we develop an efficient training and sampling scheme for MF in the latent space of a Representation Autoencoder (RAE), where a pre-trained vision encoder (e.g., DINO) provides semantically rich latents paired with a lightweight decoder. We observe that naive MF training in the RAE latent space suffers from severe gradient explosion. To stabilize and accelerate training, we adopt Consistency Mid-Training for trajectory-aware initialization and use a two-stage scheme: distillation from a pre-trained flow matching teacher to speed convergence and reduce variance, followed by an optional bootstrapping stage with a one-point velocity estimator to further reduce deviation from the oracle mean flow. This design removes the need for guidance, simplifies training configurations, and reduces computation in both training and sampling. Empirically, our method achieves a 1-step FID of 2.03, outperforming vanilla MF's 3.43, while reducing sampling GFLOPS by 38% and total training cost by 83% on ImageNet 256. We further scale our approach to ImageNet 512, achieving a competitive one-step FID of 3.23 with the lowest GFLOPS among all baselines.

With the surge of pre-trained text-to-image flow matching models, text-based image editing performance has gained remarkable improvement, especially for **simple editing** that only contains a single editing target. However, to satisfy the exploding editing requirements, the **complex editing** that contains multiple editing targets is posed as a more challenging task. However, current complex editing solutions: single-round and multi-round editing are limited by long text following and cumulative inconsistency, respectively. Thus, they struggle to strike a balance between semantic alignment and source consistency.In this paper, we propose **FlowDC**, which decouples the complex editing into multiple sub-editing effects and superposes them in parallel during the editing process. Meanwhile, we observed that the velocity quantity that is orthogonal to the editing displacement harms the source structure preserving. Thus, we decompose the velocity and decay the orthogonal part for better source consistency.To evaluate the effectiveness of complex editing settings, we construct a complex editing benchmark: Complex-PIE-Bench. On two benchmarks, FlowDC shows superior results compared with existing methods. We also detail the ablations of our module designs.