CVPR 2026 Papers

Multimodal large language models (MLLMs) have expanded from vision-language systems to include audio, unlocking new capabilities in cross-modal reasoning and interaction. To address the limitation that existing benchmarks focus mainly on perception tasks and lack a unified cognitive evaluation framework, we propose Hierarchical Audio-Visual Evaluation Benchmark (HAVE-Bench). It systematically evaluates the audio-related capabilities of MLLMs along a three-level cognitive hierarchy: Perception, Reasoning, and Interaction, utilizing 2,451 curated samples and manually annotated multi-turn interaction-level tasks. Experiments using this unified framework reveal significant gaps in existing models at the reasoning and interaction levels, with speech-driven visual question answering (VQA) performance significantly lagging behind the text-image setting. These findings underscore the urgency of enhancing models' handling of long and complex audio and facilitating the transfer of reasoning capabilities from the vision-text to the audio-visual domain.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in multimodal tasks.Despite their impressive performance, MLLMs suffer from the modality imbalance issue, where visual information is often underutilized compared to textual representations in deeper layers, leading to degraded visual performance or hallucinations.This issue stems from the predominant reliance on next-text-token-prediction during training, which fails to provide direct visual supervisory signals, resulting in progressive homogenization of visual representations throughout the layers.To this end, we propose Latent Visual Reconstruction (LaVer), a novel training framework that facilitates MLLMs in learning more discriminative visual representations via masked image modeling in the joint latent semantic space of LLM.Our method offers direct visual activation to MLLMs, which exhibit increased visual attention allocation, indicating enhanced utilization of visual information.Extensive experiments across diverse benchmarks prove the superiority of our approach in various scenarios, especially those requiring dense visual capabilities.Code will be publicly available upon publication.

Reinforcement learning (RL), particularly GRPO, improves image generation quality significantly by comparing the relative performance of images generated within the same group. However, in the later stages of training, the model tends to produce homogenized outputs, lacking creativity and visual diversity, restricting the application scenarios of the model. This issue can be analyzed from both reward modeling and generation dynamics perspectives. First, traditional GRPO relies on single sample quality as the reward signal, driving the model to converge toward a few high reward generation modes while neglecting distribution-level diversity. Second, conventional GRPO regularization neglects the dominant role of early stage denoising in preserving diversity, causing a misaligned regularization budget that limits the achievable quality-diversity trade-off.Motivated by these insights, we revisit the diversity degradation problem from both reward modeling and generation dynamics. At the reward level, we propose a distributional creativity bonus based on semantic grouping. Specifically, we construct a distribution-level representation via spectral clustering over samples generated from the same caption, and adaptively allocate exploratory rewards according to group sizes to encourage the discovery of novel visual modes. At the generation level, we introduce a structure-aware regularization, which enforces stronger early-stage constraints to preserve diversity without compromising reward optimization efficiency. Experiments demonstrate that our method achieves an 13%~18% improvement in semantic diversity under matched quality scores, establishing a new Pareto frontier between image quality and diversity for GRPO-based image generation.

Digital characters are central to modern media, yet generating character videos with long-duration, consistent multi-view appearance and expressive identity remains challenging. Existing approaches either provide insufficient context to preserve identity or leverage non-character-centric information as the "memory", leading to suboptimal consistency.Recognizing that character video generation inherently resembles an "outside-looking-in" scenario. In this work, we propose represent the character's visual attributes through a compact set of anchor frames.This design provides stable references for consistency, while reference-based video generation inherently faces challenges of copy-pasting and multi-reference conflicts. To address these, we introduce two mechanisms: Superset Content Anchoring, providing intra- and extra-training clip cues to prevent duplication, and RoPE as Weak Condition, encoding positional offsets to distinguish multiple anchors.Furthermore, we construct a scalable pipeline to extract these anchors from massive videos. Experiments show our method generates high-quality character videos exceeding 10 minutes, and achieves expressive identity and appearance consistency across views, surpassing existing methods.

Existing Multimodal Large Language Models (MLLMs) suffer from significant performance degradation on the long document understanding task as document length increases. This stems from two fundamental challenges: 1) a low Signal-to-Noise Ratio (SNR), with crucial evidence buried in irrelevant pages; and 2) supervision scarcity, as datasets offering only final short answers provide a weak learning signal. In this paper, we address these challenges by proposing a paradigm that requires the model to execute a structured "Analysis, Localization and Reasoning" workflow. To instill this capability, we design a two-stage training framework: we first perform Supervised Fine-Tuning on high-quality data generated via an efficient knowledge distillation strategy. Subsequently, we employ an Evidence-aware Group Relative Policy Optimization which jointly optimizes for both evidence localization and answer accuracy. Additionally, we introduce a Evidence-Guided Resolution Allocation strategy to mitigate memory constraints of training on multi-pages documents. Extensive experiments demonstrate that DocSeeker achieves superior performance on both in-domain and out-of-domain tasks. We show it robustly generalizes from short-page training to ultra-long documents and is naturally synergistic with visual Retrieval-Augmented Generation systems, serving as an ideal foundation for their implementation.

Recent progress in Multimodal Large Language Models (MLLMs) demonstrates that Chain-of-Thought (CoT) reasoning enables systematic solutions to complex understanding tasks. However, its extension to generation tasks remains nascent and limited by scenario-specific mechanisms that hinder generalization and adaptation. In this work, we present ThinkGen, the first think-driven visual generation framework that explicitly leverages MLLM's CoT reasoning in various generation scenarios. ThinkGen employs a decoupled architecture comprising a pretrained MLLM and a Diffusion Transformer (DiT), wherein the MLLM generates tailored instructions based on user intent, and DiT produces high-quality images guided by these instructions. We further propose a separable GRPO-based training paradigm (SepGRPO), alternating reinforcement learning between the MLLM and DiT modules. This flexible design enables joint training across diverse datasets, facilitating effective CoT reasoning for a wide range of generative scenarios. Extensive experiments demonstrate that ThinkGen achieves robust, state-of-the-art performance across multiple generation benchmarks.

We propose LiveGesture, the first fully streamable, speech-driven full-body gesture generation framework that operates with zero look-ahead and supports arbitrary sequence length. Unlike existing co-speech gesture methods--which are designed for offline generation and either treat body regions independently or entangle all joints within a single model--LiveGesture is built from the ground up for causal, region-coordinated motion generation. LiveGesture consists of two main modules: the Streamable Vector-Quantized Motion Tokenizer (SVQ) and the Hierarchical Autoregressive Transformer (HAR). The SVQ tokenizer converts the motion sequence of each body region into causal, discrete motion tokens, enabling real-time, streamable token decoding. On top of SVQ, HAR employs region-eXpert autoregressive (xAR) transformers to model expressive, fine-grained motion dynamics for each body region. A causal spatio-temporal fusion module (xAR-Fusion) then captures and integrates correlated motion dynamics across regions. Both xAR and xAR-Fusion are conditioned on live, continuously arriving audio signals encoded by a streamable causal audio encoder. To enhance robustness under streaming noise and prediction errors, we introduce autoregressive masking training, which leverages uncertainty-guided token masking and random region masking to expose the model to imperfect, partially erroneous histories during training. Experiments on the BEAT2 dataset demonstrate that LiveGesture produces coherent, diverse, and beat-synchronous full-body gestures in real time, matching or surpassing state-of-the-art offline methods under true zero-look-ahead conditions.

Pre-trained diffusion models excel at generating high-quality images but remain inherently limited by their native training resolution. Recent training-free approaches have attempted to overcome this constraint by introducing interventions during the denoising process; however, these methods incur substantial computational overhead, often requiring more than five minutes to produce a single 4K image.In this paper, we present PixelRush, the first tuning-free framework for practical high-resolution text-to-image generation. Our method builds upon the established patch-based inference paradigm but eliminates the need for multiple inversion and regeneration cycles. Instead, PixelRush enables efficient patch-based denoising within a low-step regime. To address artifacts introduced by patch blending in few-step generation, we propose a seamless blending strategy. Furthermore, we mitigate over-smoothing effects through a noise injection mechanism. PixelRush delivers exceptional efficiency, generating 4K images in approximately 20 seconds representing a 10xto 35xspeedup over state-of-the-art methods while maintaining superior visual fidelity. Extensive experiments validate both the performance gains and the quality of outputs achieved by our approach.

Reinforcement learning (RL) has become a powerful tool for post-training visual generative models, with Group Relative Policy Optimization (GRPO) increasingly used to align generators with human preferences. However, existing GRPO pipelines rely on a single scalar reward per sample, treating each image or video as a holistic entity and ignoring the rich spatial and temporal structure of visual content. This coarse supervision hinders the correction of localized artifacts and the modeling of fine-grained perceptual cues. We introduce Visual Preference Policy Optimization (ViPO), a GRPO variant that lifts scalar feedback into structured, pixel-level advantages. ViPO employs a Perceptual Structuring Module that uses pretrained vision backbones to construct spatially and temporally aware advantage maps, redistributing optimization pressure toward perceptually important regions while preserving the stability of standard GRPO. Across both image and video benchmarks, ViPO consistently outperforms vanilla GRPO, improving in-domain alignment with human-preference rewards and enhancing generalization on out-of-domain evaluations. The method is architecture-agnostic, lightweight, and fully compatible with existing GRPO training pipelines, providing a more expressive and informative learning signal for visual generation.

While Multimodal Large Language Models (MLLMs) excel on benchmarks, their processing paradigm differs from the human ability to integrate visual information. Unlike humans who naturally bridge details and high-level concepts, models tend to treat these elements in isolation. Prevailing evaluation protocols often decouple low-level perception from high-level reasoning, overlooking their semantic and causal dependencies, which yields non-diagnostic results and obscures performance bottlenecks. We present VCU-Bridge, a framework that operationalizes a human-like hierarchy of visual connotation understanding: multi-level reasoning that advances from foundational perception through semantic bridging to abstract connotation, with an explicit evidence-to-inference trace from concrete cues to abstract conclusions. Building on this framework, we construct HVCU-Bench, a benchmark for hierarchical visual connotation understanding with explicit, level-wise diagnostics. Comprehensive experiments demonstrate a consistent decline in performance as reasoning progresses to higher levels. We further develop a data generation pipeline for instruction tuning guided by Monte Carlo Tree Search (MCTS) and show that strengthening low-level capabilities yields measurable gains at higher levels. Interestingly, it not only improves on HVCU-Bench but also brings benefits on general benchmarks (average +2.53%), especially with substantial gains on MMStar (+7.26%), demonstrating the significance of the hierarchical thinking pattern and its effectiveness in enhancing MLLM capabilities.

Omnimodal large language models (OmniLLMs) have attracted increasing research attention of late towards unified audio-video understanding. However, the high computational cost of processing longer joint audio-video token sequences has become a key bottleneck. Existing token compression methods have not addressed the emerging need to jointly compress multimodal tokens. To bridge this gap, we present OmniZip, a training-free, audio-guided audio-visual token-compression framework that optimizes multimodal token representation and accelerates model inference. Specifically, OmniZip first identifies salient audio tokens, then computes an audio retention score for each time group to capture information density, thereby dynamically guiding video token pruning and preserving cues from audio anchors enhanced by cross-modal similarity. For each time window, OmniZip compresses the video tokens using an interleaved spatio-temporal scheme. Extensive results demonstrate the merits of OmniZip: it achieves a 3.42x inference speedup and a 1.4x memory reduction over other top-performing counterparts, while maintaining the performance of OmniLLMs without training.

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.

Continuous image tokenizers enable efficient visual generation, and those based on variational frameworks can learn smooth, structured latent representations through KL regularization. Yet this often leads to posterior collapse when using fewer tokens, where the encoder fails to encode informative features into the compressed latent space. To address this, we introduce **MacTok**, a **M**asked **A**ugmenting 1D **C**ontinuous **Tok**enizer that leverages image masking and representation alignment to prevent collapse while learning compact and robust representations. MacTok applies both random masking to regularize latent learning and DINO-guided semantic masking to emphasize informative regions in images, forcing the model to encode robust semantics from incomplete visual evidence. Combined with global and local representation alignment, MacTok preserves rich discriminative information in a highly compressed 1D latent space, requiring only 64 or 128 tokens. On ImageNet, MacTok achieves a competitive gFID of 1.44 at 256x256 and a state-of-the-art 1.52 at 512x512 with SiT-XL, while reducing token usage by up to 64x. These results confirm that masking and semantic guidance together prevent posterior collapse and achieve efficient, high-fidelity tokenization.

The rise of vision foundation models (VFMs) calls for systematic evaluation. A common approach pairs VFMs with large language models (LLMs) as general-purpose heads, followed by evaluation on broad Visual Question Answering (VQA) benchmarks. However, this protocol has two key blind spots: (i) Instruction tuning data may not align with VQA test distributions, meaning a wrong prediction can stem from such data mismatch rather than VFMs' visual shortcomings; (ii) VQA benchmarks often require multiple visual abilities in a single question, making it difficult to determine whether errors arise from the lack of all required abilities or just one key ability. To address these gaps, we introduce AVA-Bench, the first benchmark that explicitly disentangles 14 Atomic Visual Abilities (AVAs), foundational skills such as localization, depth estimation, and spatial understanding, which collectively support complex visual reasoning tasks. By decoupling AVAs and matching training and test distributions within each, AVA-Bench pinpoints exactly where a VFM excels or falters. Applying AVA-Bench to leading VFMs thus reveals distinctive "ability fingerprints," turning VFM selection from educated guesswork into principled engineering. Notably, we find that a 0.5B LLM yields similar VFM rankings as a 7B LLM while cutting GPU hours by 8x, enabling more efficient evaluation. By offering a comprehensive and transparent benchmark, we hope AVA-Bench lays the foundation for the next generation of VFMs.

Digital humans are fundamental to immersive interaction, yet creating a unified model for holistic modalities, including text, audio, motion, and visual content, remains an open challenge. In this paper, we present Archon, a fully pretrained, human-centric unified multimodal model for holistic avatar generation. Archon unifies seven modalities with modality-specific tokenizers, and a native autoregressive unified multimodal model pretrained on synchronized modalities and 72 diverse tasks to model holistic joint distributions. To address the token explosion challenge in high-fidelity talking videos, we introduce a memory-efficient semantic video reparameterization, achieving 4x token reduction while preserving fine-grained dynamics, coupled with a semantic-driven video diffusion decoder. We further propose a "Thinking in Modality" that decomposes ambiguous cross-modal tasks into stepwise thinking in an alternative chain of modality, progressively enhancing fidelity and controllability. Extensive experiments demonstrate that Archon achieves superior or comparable performance across diverse digital human generation tasks, validating the effectiveness of our unified framework.

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.

Autoregressive (AR) video generative models rely on video tokenizers that compress pixels into discrete token sequences. The length of these token sequences is crucial for balancing reconstruction quality against downstream generation computational cost. Traditional video tokenizers apply a uniform token assignment across temporal blocks of different videos, often wasting tokens on simple, static, or repetitive segments while underserving dynamic or complex ones. To address this inefficiency, we introduce **EVATok**, a framework to produce **E**fficient **V**ideo **A**daptive **Tok**enizers. Our framework estimates optimal token assignments for each video to achieve the best quality-cost trade-off, develops lightweight routers for fast prediction of these optimal assignments, and trains adaptive tokenizers that encode videos based on the assignments predicted by routers. We demonstrate that EVATok delivers substantial improvements in efficiency and overall quality for video reconstruction and downstream AR generation. Enhanced by our advanced training recipe that integrates video semantic encoders, EVATok achieves superior reconstruction and state-of-the-art class-to-video generation on UCF-101, with at least 24.4% savings in average token usage compared to the prior state-of-the-art LARP and our fixed-length baseline.

Self-supervised representations excel at many vision and speech tasks, but their potential for audio-visual deepfake detection remains underexplored. Unlike prior work that uses these features in isolation or buried within complex architectures, we systematically evaluate them across modalities (audio, video, multimodal) and domains (lip movements, generic visual content). We assess three key dimensions: detection effectiveness, interpretability of encoded information, and cross-modal complementarity. We find that most self-supervised features capture deepfake-relevant information, and that this information is complementary. Moreover, models primarily attend to semantically meaningful regions rather than spurious artifacts (such as the leading silence). Among the investigated features, audio-informed representations generalize best and achieve state-of-the-art results. However, generalization to realistic in-the-wild data remains challenging. Our analysis indicates this gap stems from intrinsic dataset difficulty rather than from features latching onto superficial patterns.