CVPR 2026 Papers

Face swapping aims to transfer the identity of a source face onto a target face while preserving target-specific attributes such as pose, expression, lighting, skin tone, and makeup. However, since real ground truth for face swapping is unavailable, achieving both accurate identity transfer and high-quality attribute preservation remains challenging. Recent diffusion-based approaches attempt to improve visual fidelity through conditional inpainting on masked target images, but the masked condition removes crucial appearance cues, resulting in plausible yet misaligned attributes. To address this limitation, we propose APPLE (Attribute-Preserving Pseudo-Labeling), a fully diffusion-based teacher-student framework for attribute-preserving face swapping. Our approach introduces a teacher design to produce pseudo-labels aligned with the target attributes through (1) a conditional deblurring formulation that improves the preservation of global attributes such as skin tone and illumination, and (2) an attribute-aware inversion scheme that further enhances fine-grained attribute preservation such as makeup. APPLE conditions the student on clean pseudo-labels rather than degraded masked inputs, enabling more faithful attribute preservation. As a result, APPLE achieves state-of-the-art performance in attribute preservation while maintaining competitive identity transferability.

Recent advances in generative AI have dramatically improved photorealistic image synthesis, yet they fall short for studio-level multi-object compositing. This task demands simultaneous (i) near-perfect preservation of each item's identity, (ii) precise background and color fidelity, (iii) layout and design elements control, and (iv) complete, appealing displays showcasing all objects. However, current state-of-the-art models often alter object details, omit or duplicate objects, and produce layouts with incorrect relative sizing or inconsistent item presentations. To bridge this gap, we introduce PLACID, a framework that transforms a collection of object images into an appealing multi-object composite. Our approach makes two main contributions. First, we leverage a pretrained image-to-video (I2V) diffusion model with text control to preserve objects consistency, identities and background details by exploiting temporal priors from videos. Second, we propose a novel data curation strategy that generates synthetic sequences where randomly placed objects smoothly move to their target positions. This synthetic data aligns with the video model's temporal priors during training. At inference, objects initialized at random positions consistently converge into coherent layouts guided by text, with the final frame serving as the composite image. Extensive quantitative evaluations and user studies demonstrate that PLACID surpasses state-of-the-art methods in multi-object compositing, achieving superior identity, background, and color preservation, with less omitted objects and visually appealing results.

Diffusion Transformers (DiTs) set the state of the art in visual generation, yet their quadratic self-attention cost fundamentally limits scaling to long token sequences. Recent Top-K sparse attention approaches reduce the computation of DiTs by compressing tokens into block-wise representation and selecting a small set of relevant key blocks, but still suffer from (i) quadratic selection cost on compressed tokens and (ii) increasing K required to maintain model quality as sequences grow. We identify that their inefficiency is due to the single-level design, as a single coarse level is insufficient to represent the global structure.In this paper, we introduce Log-linear Sparse Attention (LLSA), a trainable sparse attention mechanism for extremely long token sequences that reduces both selection and attention costs from quadratic to log-linear complexity by utilizing a hierarchical structure. LLSA performs hierarchical Top-K selection, progressively adopting sparse Top-K selection with the indices found at the previous level,and introduces a Hierarchical KV Enrichment mechanism that preserves global context while using fewer tokens of different granularity during attention computation. To support efficient training, we develop a high-performance GPU implementation that uses only sparse indices for both the forward and backward passes, eliminating the need for dense attention masks.We evaluate LLSA on high-resolution pixel-space image generation without using patchification and VAE encoding. LLSA accelerates attention inference by 28.27 xand DiT training by 6.09 xon 256 x256 pixel token sequences, while maintaining generation quality. The results demonstrate that LLSA offers a promising direction for training long-sequence DiTs efficiently.

Despite the remarkable success, recent reconstruction-based anomaly detection (AD) methods via diffusion modeling still involve fine-grained noise-strength tuning and computationally expensive multi-step denoising, leading to a fundamental tension between fidelity and efficiency. In this paper, we propose InvAD, a novel inversion-based anomaly detection approach for detection via noising in latent space, which circumvents explicit reconstruction. Importantly, we contend that the limitations of prior reconstruction-based methods originate from the prevailing "detection via denoising in RGB space" paradigm. To address this, we model AD under a reconstruction-free formulation that directly infers the final latent variable corresponding to the input image via DDIM inversion and then measures the deviation based on the known prior distribution for anomaly scoring. Specifically, when approximating the original probability flow ODE using the Euler method, we use only a few inversion steps to noise the clean image and improve inference efficiency. As the added noise is adaptively derived from the learned diffusion model, the original features of the clean test image can still be leveraged to yield high detection accuracy. We perform extensive experiments and detailed analysis across four widely used AD benchmarks under the unsupervised unified setting to demonstrate the effectiveness of our model, achieving state-of-the-art AD performance and about a twofold inference-time speedup without diffusion distillation.

Recent advances in diffusion models have demonstrated impressive capability in generating high-quality images for simple prompts. However, when confronted with complex prompts involving multiple objects and hierarchical structures, existing models struggle to accurately follow instructions, leading to issues such as concept omission, confusion, and poor compositionality. To address these limitations, we propose a Hierarchical Compositional Generative framework (HiCoGen) built upon a novel Chain of Synthesis (CoS) paradigm. Instead of monolithic generation, HiCoGen first leverages a Large Language Model (LLM) to decompose complex prompts into minimal semantic units. It then synthesizes these units iteratively, where the image generated in each step provides crucial visual context for the next, ensuring all textual concepts are faithfully constructed into the final scene. To further optimize this process, we introduce a reinforcement learning (RL) framework. Crucially, we identify that the limited exploration of standard diffusion samplers hinders effective RL. We theoretically prove that sample diversity is maximized by concentrating stochasticity in the early generation stages and, based on this insight, propose a novel Decaying Stochasticity Schedule to enhance exploration. Our RL algorithm is then guided by a hierarchical reward mechanism that jointly evaluates the image at the global, subject, and relationship levels. We also construct HiCoPrompt, a new text-to-image benchmark with hierarchical prompts for rigorous evaluation. Experiments show our approach significantly outperforms existing methods in both concept coverage and compositional accuracy.

Synthesizing realistic human-object interaction (HOI) is essential for 3D computer vision and robotics, underpinning animation and embodied control. Existing approaches often require manually specified intermediate waypoints and place all optimization objectives on a single network, which increases complexity, reduces flexibility, and leads to errors such as unsynchronized human and object motion or penetration. To address these issues, we propose Decoupled Generative Modeling for Human-Object Interaction Synthesis (DecHOI), which separates path planning and action synthesis. A trajectory generator first produces human and object trajectories without prescribed waypoints, and an action generator conditions on these paths to synthesize detailed motions. To further improve contact realism, we employ adversarial training with a discriminator that focuses on the dynamics of distal joints. The framework also models a moving counterpart and supports responsive, long-sequence planning in dynamic scenes, while preserving plan consistency. Across two benchmarks, FullBodyManipulation and 3D-FUTURE, DecHOI surpasses prior methods on most quantitative metrics and qualitative evaluations, and perceptual studies likewise prefer our results.

Though rectified flow models have achieved remarkable performance in image, video, and 3D generation, their practical deployments are challenged by slow inference speeds. Prior acceleration methods reuse cached features from previous steps, which neglects the growing mismatch between static caches and the evolving input, leading to reduced output fidelity. This work proposes Velocity Decomposition and Estimation (VDE), a training-free acceleration method that shifts the paradigm from caching-and-reusing to decomposing-and-estimating. Specifically, VDE decomposes the model's velocity into components parallel and orthogonal to the input, exploiting their temporal predictability and directional consistency for precise, input-adaptive estimation. To prevent error accumulation, it periodically anchors the model's state via full forward passes. Extensive experiments on image and video generation tasks demonstrate that VDE achieves substantial acceleration with minimal loss in visual quality. Notably, VDE accelerates Flux by 3.22x and achieves an LPIPS of 0.069 on Qwen-Image, outperforming the best baseline with a 52.2% reduction. Code: https://github.com/Tan-Junwen/VDE

Diffusion models (DMs) have recently achieved remarkable success across diverse modalities, including high-fidelity image and video synthesis.However, their inherent step sequential denoising process introduces substantial cumulative latency, which significantly degrades user experience. While existing multi-GPU parallelization motheds can alleviate latency, they often incur prohibitive GPU-GPU communication overhead, offsetting much of the performance gain.We present Otil (Only Transmit Informative Latents), a communication-efficient parallel framework for accelerating diffusion inference.% Otil can minimizes redundant data exchange across GPUs while preserving generation quality.Our key insight is that latent activations change only marginally between consecutive denoising process. Leveraging this property, Otil identifies and synchronizes only the most informative latent sub-blocks and introduces a dynamic polling mechanism that periodically revisits all spatial regions, ensuring complete coverage without unnecessary communication. The framework is fully plug-and-play and remains compatible with fast sample and architectural acceleration algorithms, without requiring any retraining or architectural modification.Otil reduces GPU-GPU communication up to 87.5% compared with SOTA parallelism methods, achieving 1.8xspeedup on two GPUs with Stable Diffusion v1.5 and 2.6xon four GPUs with Stable Diffusion XL. When combined with few-step samplers (30 steps) and LoRA models, the acceleration further increases to 2.46x-2.84xon 2 GPUs. These demonstrate the strong potential of Otil for scalable and efficient multi-GPU diffusion inference while preserving generation fidelity.

Video diffusion models have achieved remarkable generative performance, but their substantial computational and memory costs pose significant challenges for deployment, especially on consumer GPUs. As recent advances in attention optimization mitigate previous computational bottlenecks, linear layers now dominate both computational cost and inference memory. In this work, we focus on quantizing both weights and activations to 4 bits to accelerate these layers. Previous methods, such as SVDQuant, overlook the highly dynamic nature of activations across denoising timesteps, where outlier channels and magnitudes vary dramatically. However, video data inherently exhibits strong activation similarity among neighboring tokens in space and time, which we term spatiotemporal activation similarity, analogous to how video codecs exploit intra- and inter-frame redundancy. Leveraging this property, we introduce DeltaQuant, which partitions activations into local 3D spatiotemporal cubes and uses each cube's mean token as a \coretoken, quantizing only the small differences (delta tokens) to 4 bits while keeping core tokens in FP8. This decomposition substantially reduces quantization error with minimal overhead.For weight quantization, DeltaQuant incorporates SVDQuant's low-rank decomposition to further reduce quantization error.We also implement an efficient kernel that translates DeltaQuant's computational benefits into real-world speedups.Extensive experiments on Wan 2.2 I2V, Wan 2.2 T2V, and LTX-Video T2V demonstrate that DeltaQuant maintains high generation fidelity.On Wan 2.2, it compresses model size by 2.9x and reduces memory footprint by 2.3x. DeltaQuant is compatible with efficient attention mechanisms and few-step distillation. When integrated with these techniques, it achieves an additional 3.0x acceleration, for a total 111.8x end-to-end speedup. Code and models will be released upon publication.

Latent diffusion models such as Stable Diffusion 1.5 offer strong generative priors that are highly valuable for image restoration, yet their full pipelines remain too computationally heavy for deployment on edge devices. Existing lightweight variants predominantly compress the denoising U-Net or reduce the diffusion trajectory, which disrupts the underlying latent manifold and limits generalization beyond a single task. We introduce NanoSD, a family of Pareto-optimal diffusion foundation models distilled from Stable Diffusion 1.5 through network surgery, feature-wise generative distillation, and structured architectural scaling jointly applied to the U-Net and the VAE encoder-decoder. This full-pipeline co-design preserves the generative prior while producing models that occupy distinct operating points along the accuracy-latency-size frontier (e.g., 130M-315M parameters, achieving real-time inference down to 20ms on mobile-class NPUs). We show that parameter reduction alone does not correlate with hardware efficiency, and we provide an analysis revealing how architectural balance, feature routing, and latent-space preservation jointly shape true on-device latency. When used as a drop-in backbone, NanoSD enables state-of-the-art performance across image super-resolution, image deblurring, face restoration, and monocular depth estimation, outperforming prior lightweight diffusion models in both perceptual quality and practical deployability. NanoSD establishes a general-purpose diffusion foundation model family suitable for real-time visual generation and restoration on edge devices.

Generative AI models pose a significant challenge to intellectual property (IP), as they can replicate unique artistic styles and concepts without attribution. While watermarking offers a potential solution, existing methods often fail in complex scenarios where multiple concepts (e.g., an object and an artistic style) are composed within a single image. These methods struggle to disentangle and attribute each concept individually. In this work, we introduce TokenTrace, a novel proactive watermarking framework for robust, multi-concept attribution. Our method embeds secret signatures into the semantic domain by simultaneously perturbing the text prompt embedding and the initial latent noise that guide the diffusion model's generation process. For retrieval, we propose a query-based TokenTrace module that takes the generated image and a textual query specifying which concepts need to be retrieved (e.g., a specific object or style) as inputs. This query-based mechanism allows the module to disentangle and independently verify the presence of multiple concepts from a single generated image. Extensive experiments show that our method achieves state-of-the-art performance on both single-concept (object and style) and multi-concept attribution tasks, significantly outperforming existing baselines while maintaining high visual quality and robustness to common transformations.

Text-to-multiview (T2MV) diffusion models have shown great promise in generating multiple views of a scene from a single text prompt. While few-step backbones enable real-time T2MV generation, they often compromise key aspects of generation quality, such as per-view fidelity and cross-view consistency. Reinforcement learning (RL) finetuning offers a potential solution, yet existing approaches designed for single-image diffusion do not readily extend to the few-step T2MV setting, as they neglect cross-view coordination and suffer from weak learning signals in few-step regimes. To address this, we propose MVC-ZigAL, a tailored RL finetuning framework for few-step T2MV diffusion models. Specifically, its core insights are: (1) a new MDP formulation that jointly models all generated views and assesses their collective quality via a joint-view reward; (2) a novel advantage learning strategy that exploits the performance gains of a self-refinement sampling scheme over standard sampling, yielding stronger learning signals for effective RL finetuning; and (3) a unified RL framework that extends advantage learning with a Lagrangian dual formulation for multiview-constrained optimization, balancing single-view and joint-view objectives through adaptive primal-dual updates under a self-paced threshold curriculum that harmonizes exploration and constraint enforcement. Collectively, these designs enable robust and balanced RL finetuning for few-step T2MV diffusion models, yielding substantial gains in both per-view fidelity and cross-view consistency. Code is available at https://github.com/ZiyiZhang27/MVC-ZigAL.

Diffusion transformers (DiTs) offer excellent scalability for high-fidelity generation, but their computational overhead poses a great challenge for practical deployment. Existing acceleration methods primarily exploit the temporal dimension, whereas spatial acceleration remains underexplored. In this work, we investigate spatial acceleration for DiTs via latent upsampling. We found that naive latent upsampling for spatial acceleration introduces artifacts, primarily due to aliasing in high-frequency edge regions and mismatching from noise-timestep discrepancies. Then, based on these findings and analyses, we propose a training-free spatial acceleration framework, dubbed Region-Adaptive Latent Upsampling (RALU), to mitigate those artifacts while achieving spatial acceleration of DiTs by our mixed-resolution latent upsampling. RALU achieves artifact-free, efficient acceleration with early upsampling only on artifact-prone edge regions and noise-timestep matching for different latent resolutions, leading to up to 7.0xspeedup on FLUX-1.dev and 3.0xon Stable Diffusion 3 with negligible quality degradation. Furthermore, our RALU is complementarily applicable to existing temporal acceleration methods and timestep-distilled models, leading to up to 15.9xspeedup.

Single-image reflection removal is a highly ill-posed problem, where existing methods struggle to reason about the composition of corrupted regions, causing them to fail at recovery and generalization in the wild. This work reframes an editing-purpose latent diffusion model to effectively perceive and process highly ambiguous, layered image inputs, yielding high-quality outputs. We argue that the challenge of this conversion stems from a critical yet overlooked issue, i.e., the latent space of semantic encoders lacks the inherent structure to interpret a composite image as a linear superposition of its constituent layers. Our approach is built on three synergistic components, including a reflection-equivariant VAE that aligns the latent space with the linear physics of reflection formation, a learnable task-specific text embedding for precise guidance that bypasses ambiguous language, and a depth-guided early-branching sampling strategy to harness generative stochasticity for promising results. Extensive experiments reveal that our model achieves new SOTA performance on multiple benchmarks and generalizes well to challenging real-world cases.

Acquiring physically plausible motor skills across diverse and unconventional embodiments, including humanoids and quadrupeds, is essential for advancing character simulation and robotics. Traditional methods, such as reinforcement learning (RL), require extensive reward function engineering. Imitation learning (IL) offers an alternative but relies heavily on curated 3D expert demonstrations, which are scarce and difficult to obtain for non-human morphologies. Video diffusion models, on the other hand, are capable of generating realistic-looking videos of various morphologies, from humans to ants. However, these videos are often not physically plausible, which limits their direct use for skill acquisition. We introduce "No-data Imitation Learning" (NIL): an imitation learning framework that replaces curated expert demonstrations with videos generated by a pretrained video diffusion model. Our key insight is that the physics simulator enforces physical constraints, while the video provides visual guidance. NIL learns 3D motor skills in a physics simulator from 2D-generated videos, with generalization capability to unconventional forms. Specifically, NIL computes a discriminator-free imitation reward that combines (i) a video-embedding similarity between generated and simulated videos using a pretrained video vision transformer, and (ii) an image-based similarity term derived from video segmentation masks. We evaluate NIL on locomotion and whole-body control tasks across unique body configurations. Our experiments show that in humanoid locomotion, NIL matches the performance of state-of-the-art IL baselines trained on motion capture data; and in whole-body manipulation, it exceeds the performance of RL baselines without requiring any curated data. Our project page, including videos and code, is available at https://nil.is.tue.mpg.de.

We present MeshFlow, a new method for compressing and generating artist-like 3D meshes. Current mesh generators often adopt Auto-Regressive (AR) next-token prediction, a natural choice given the discrete nature of mesh connectivity, which, however, scales poorly due to the inference cost being quadratic in mesh size. AR methods also require discretizing the vertex coordinates, which introduces quantization errors and can cause vertex collapse. To address these challenges, we introduce a Variational Autoencoder (VAE) that, supervised with a contrastive loss, represents both continuous vertex positions and discrete connectivity in a continuous latent space.This latent space is significantly more compact than prior token-based mesh representations. We then build a 3D generator based on a Rectified-Flow transformer, which generates all mesh vertices and edges in parallel. This model samples meshes 26xfaster than the fastest AR generator while also achieving state-of-the-art accuracy across standard mesh-generation metrics.

This work studies how latent space structure impacts the performance of Latent Diffusion Models (LDMs). We show that effective generation requires a latent space that is simultaneously locally smooth, enabling stable and reliable reconstruction, and globally dispersive, allowing the model to draw diverse and meaningful samples without collapsing into narrow regions of the latent space. However, existing approaches often emphasize smoothness, which may lead to concentrated latent regions and limited exploration of the broader space. To address these limitations, we propose Self-Organized Representation Learning (SORL), a bottom-up training paradigm inspired by self-organization in complex systems, where global structure emerges naturally from simple local interactions. The critical latent properties of smoothness and maximal dispersity are not explicitly imposed. Instead, SORL promotes these properties through two complementary local mechanisms: local attraction, which encourages coherent reconstructions among nearby latent codes, and local repulsion, which prevents latent codes from collapsing into dense clusters. Through their interaction, SORL induces a latent manifold that maintains both local smoothness and global dispersity, leading to improved reconstruction and generation.

In this work, we introduce LLaDA-V, a purely diffusion-based Multimodal Large Language Model (MLLM) that integrates visual instruction tuning with masked diffusion models, representing a departure from the autoregressive paradigms dominant in current multimodal approaches. Built upon LLaDA, a representative large language diffusion model, LLaDA-V incorporates a vision encoder and MLP connector that projects visual features into the language embedding space, leveraging diffusion language models' bidirectional attention to capture spatial relationships in visual data more effectively than causal, sequential processing. Our empirical investigation reveals several intriguing results: First, LLaDA-V demonstrates promising multimodal performance despite its language model being weaker on purely textual tasks than counterparts like LLaMA3-8B and Qwen2-7B. When trained on the same instruction data, LLaDA-V is highly competitive with LLaMA3-V across multimodal tasks and shows promising data scaling behavior on several benchmarks. It also narrows the performance gap to Qwen2-VL, suggesting the effectiveness of its architecture for multimodal tasks. Second, LLaDA-V achieves state-of-the-art performance in multimodal understanding compared to existing purely diffusion-based MLLMs. Our findings suggest that large language diffusion models show promise in multimodal contexts and warrant further investigation. To facilitate future research, we open-source LLaDA-V together with its training and evaluation code at https://github.com/ML-GSAI/LLaDA-V.

Object-centric reconstruction seeks to recover the 3D structure of a scene through composition of independent objects. While this independence can simplify modeling, it discards strong signals that could improve reconstruction, notably repetition where the same object model is seen multiple times in a scene, or across scans. We propose the Joint Reconstruction Model (JRM) to leverage repetition by framing object reconstruction as one of personalized generation: multiple observations share a common subject that should be consistent for all observations, while still adhering to the specific pose and state from each. Prior methods in this direction rely on explicit matching and rigid alignment across observations, making them sensitive to errors and difficult to extend to non-rigid transformations. In contrast, JRM is a 3D flow-matching generative model that implicitly aggregates unaligned observations in its latent space, learning to produce consistent and faithful reconstructions in a data-driven manner without explicit constraints. Evaluations on synthetic and real-world data show that JRM's implicit aggregation removes the need for explicit alignment, improves robustness to incorrect associations, and naturally handles non-rigid changes such as articulation. Overall, JRM outperforms both independent and alignment-based baselines in reconstruction quality.