CVPR 2026 Papers

Multimodal large language models suffer from substantial inference overhead since multimodal KV Cache grows proportionally with the visual input length. Existing multimodal KV cache compression methods mostly rely on attention score to reduce cache size, which makes them are incompatible with established efficient attention kernels (e.g., FlashAttention) and ignores the contribution of value vectors to the attention output. In this work, we revisit multimodal KV cache compression from the perspective of the KV matrices' distribution. First, we observe that frequency-domain energy of multimodal KV matrices is predominantly concentrated in low-frequency and extract this principal energy via a low-pass filter. Further, we find that removing KV pairs that deviate substantially from this principal energy leads to a pronounced performance drop, which we define as Outlier KVs. Considering Outlier KVs are more likely to encode features critical for inference, we propose FlashCache, a frequency-domain-guided, Outlier-KV-aware KV Cache compression framework. First, we introduce an Outlier KV Recognition Module that models the principal component of multimodal KV matrices in the frequency domain and preferentially retains KV pairs that significantly deviate from it. Furthermore, Dynamic Budget Allocation Module is designed to adaptively determine the per-layer KV cache size to retain more Outlier KVs. Experiments on multiple MLLMs and benchmarks demonstrate that FlashCache outperforms state-of-the-art multimoal KV compression methods, achieving up to 1.69x faster decoding with 80% lower KV memory usage while maintaining task performance.

Text-to-Image (T2I) generation has long been an open problem, with compositional synthesis remaining particularly challenging. This task requires accurate rendering of complex scenes containing multiple objects that exhibit diverse attributes as well as intricate spatial and semantic relationships, demanding both precise object placement and coherent inter-object interactions. In this paper, we propose a novel compositional curriculum reinforcement learning framework named CompGen that addresses compositional weakness in existing T2I models. Specifically, we leverage scene graphs to establish a novel difficulty criterion for compositional ability and develop a corresponding adaptive Markov Chain Monte Carlo graph sampling algorithm. This difficulty-aware approach enables the synthesis of training curriculum data that progressively optimize T2I models through reinforcement learning. We integrate our curriculum learning approach into Group Relative Policy Optimization (GRPO) and investigate different curriculum scheduling strategies. Our experiments reveal that CompGen exhibits distinct scaling curves under different curriculum scheduling strategies, with easy-to-hard and Gaussian sampling strategies yielding superior scaling performance compared to random sampling. Extensive experiments demonstrate that CompGen significantly enhances compositional generation capabilities for both diffusion-based and auto-regressive T2I models, highlighting its effectiveness in improving the compositional T2I generation systems.

Few-shot Font Generation (FFG) aims to create a complete font from a limited number of references, offering significant practical value. However, existing methods neglect glyph spatial information, which leads to two critical limitations. At the pipeline level, distorted rendering introduces spatial bias, impairing vectorization and dataset quality, and this problem is compounded by the lack of unified standards, which undermines a unified benchmark. At the model level, the implicit coupling of shape and position hinders fine-grained optimization and generalization. We address these challenges in the context of Chinese font generation, where glyph complexity demands superior model capability. Consequently, we first propose a Spatial-Preserving Rendering (SPR) scheme, which eliminates spatial bias and enables accurate vectorization. Alongside, we release an OFL-licensed Chinese font dataset to establish a unified benchmark. Then, technically, we propose GlyphSpatialNet, a two-stage framework to explicitly model glyph spatial information in pixel space. In first stage, we design a Shape-Position Decoupling (SPD) architecture and a Gradient Broadcasting Module (GBM) to achieve font style transfer in low resolution. In second stage, we design Style Detail Enhancement (SDE), which refines the style details for high resolution outputs. Extensive experiments demonstrate the effectiveness of our approach. Code and dataset are available at https://github.com/sp777g/GlyphSpatialNet.

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

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

Prompt tuning of large-scale vision-language models such as CLIP enables efficienttask adaptation without updating model weights. However, it often leads to poorconfidence calibration and unreliable predictive uncertainty. We address thisproblem by proposing a calibration framework that enhances predictive reliabilitywhile preserving the geometry of the pretrained CLIP embedding space, which isrequired for robust generalization. Our approach extends the standard cross-entropyloss with two complementary regularizers: (1) a mean-variance margin penalty thatstabilizes inter-class logit margins by maximizing their average while minimizingdispersion, mitigating underconfidence and overconfidence spikes; and (2) a textmoment-matching loss that aligns the first and second moments of tuned textembeddings with their frozen CLIP counterparts, preserving semantic dispersioncrucial for generalization. Through extensive experiments across 7 prompt-tuningmethods and 11 diverse datasets, we demonstrate that our approach significantlyreduces the Expected Calibration Error (ECE) compared to competitive calibrationtechniques on both base and novel classes.

Parallel test-time scaling typically trains separate generation and verification models, incurring high training and inference costs. We propose Advantage Decoupled Preference Optimization (ADPO), a unified reinforcement learning framework that jointly learns answer generation and self-verification within a single policy. ADPO introduces two innovations: a preference verification reward improving verification capability and a decoupled optimization mechanism enabling synergistic optimization of generation and verification. Specifically, the preference verification reward computes mean verification scores from positive and negative samples as decision thresholds, providing positive feedback when prediction correctness aligns with answer correctness. Meanwhile, the advantage decoupled optimization computes separate advantages for generation and verification, applies token masks to isolate gradients, and combines masked GRPO objectives, preserving generation quality while calibrating verification scores. ADPO achieves up to +34.1% higher verification AUC and -53.5% lower inference time, with significant gains of +2.8%/+1.4% accuracy on MathVista/MMMU, +1.9 cIoU on ReasonSeg, and +1.7%/+1.0% step success rate on AndroidControl/GUI Odyssey. Code is available at https://github.com/ZJUSCL/ADPO.

Existing robot policies based on learned visual embeddings lack explicit structure and are sensitive to visual distractions.Thus, the representations that drive their behaviour are often opaque, making their decision-making process difficult to interpret.To address this, we introduce Structured Image Representation, a method that leverages Scene Graphs as an intermediate representation for robot policy learning.Our approach first constructs a fully connected graph, using 2D or 3D image-derived features as initial node representations. Then, a module learns to sparsify this graph end-to-end, creating a minimal, task-relevant sub-graph that is passed to the action generation model.This process makes our model intrinsically explainable.Evaluations on RoboCasa show that our sparse graph policies outperform image-based baselines on average with 19.5% vs 14.81% success rate.We also demonstrate that our graph-based representations are significantly more robust to distractor objects, showing almost no performance degradation, as opposed to image representations.Most importantly, we show that the learned sparse graphs are a powerful tool for introspection.By analysing when the model's sub-graph deviates from human expectation, such as by including distractor nodes or omitting key objects, we successfully uncover dataset biases, including spurious correlations and positional biases.

The "think-with-image" paradigm has recently gained traction for complex visual reasoning tasks. However, existing approaches often struggle with inference inefficiency due to a fixed number of redundant reasoning steps, as well as training instability. This challenge primarily arises from the direct use of standard reinforcement learning policies, which do not incorporate improvements for the think-with-image multi-turn conversational scenario. To address this challenge, we propose VisionLeaf, an entropy-guided, tree-based reasoning framework. Unlike conventional GRPO, where all nodes expand from the root and each leaf has only a single branch, our method grows the reasoning tree from the leaf nodes and selects the most valuable nodes based on entropy for thorough rollout exploration. This leaf-first expansion naturally aligns with the hierarchical nature of multi-step image analysis. Without modifying any model or training data, our VisionLeaf achieves a 4.2% performance improvement on benchmarks such as VSTAR and HRBench, while reducing the number of inference rounds by nearly half--demonstrating significant gains in both accuracy and speed. All our code will be released.

Modern computer vision models from different architecture families, e.g., CNNs, Vision Transformers, and MLP-Mixers, achieve remarkably similar aggregate performance on standard benchmarks, masking potential systematic differences in how they process visual information. We introduce a simple yet revealing framework to identify where pretrained model families diverge: by systematically mapping the high-disagreement tail versus the low-disagreement tail of the image distribution. Analyzing 12 pretrained models spanning three architecture families on ImageNet validation set, we discover that controversial images exhibit approximately 4.5xhigher disagreement than consensus images (Controversy Score: 4.46). Despite mean accuracy around 80%, models show structured disagreement patterns: within-family agreement exceeds cross-family agreement, with CNNs and ViTs forming distinct clusters while MLPs show lower overall alignment. Crucially, only the top 10% most controversial images drive the majority of architectural divergence, constituting a small but informationally dense subset that reveals fundamental differences masked by aggregate metrics. Our analysis demonstrates that architectural choice matters most on this concentrated controversy space, providing researchers with actionable guidance for model selection and ensemble construction.

Large-scale vision-language models such as CLIP have achieved remarkable success in zero-shot image recognition, yet their predictions remain largely opaque to human understanding. In contrast, Concept Bottleneck Models provide interpretable intermediate representations by reasoning through human-defined concepts, but they rely on concept supervision and lack the ability to generalize to unseen classes. We introduce EZPC that bridges these two paradigms by explaining CLIP's zero-shot predictions through human-understandable concepts. Our method projects CLIP's joint image-text embeddings into a concept space learned from language descriptions, enabling faithful and transparent explanations without additional supervision. The model learns this projection via a combination of alignment and reconstruction objectives, ensuring that concept activations preserve CLIP's semantic structure while remaining interpretable. Extensive experiments on five benchmark datasets, CIFAR-100, CUB-200-2011, Places365, ImageNet-100, and ImageNet-1k, demonstrate that our approach maintains CLIP's strong zero-shot classification accuracy while providing meaningful concept-level explanations. By grounding open-vocabulary predictions in explicit semantic concepts, our method offers a principled step toward interpretable and trustworthy vision-language models. Code is available at https://github.com/oonat/ezpc.

While continual visual instruction tuning (CVIT) has shown promise in adapting multimodal large language models (MLLMs), existing studies predominantly focus on models without safety alignment. This critical oversight ignores the fact that real-world MLLMs inherently require such mechanisms to mitigate potential risks. In this work, we shift our focus to CVIT for safety-aligned MLLMs and observe that during continual adaptation, the model not only suffers from task forgetting but also exhibits degradation in its safety. Achieving a harmonious balance between safety and task performance remains a crucial challenge. To address this, we propose Harmonious Parameter Adaptation (HPA), a post-training framework composed of focusing-based parameter partition, harmoniously balanced parameter selection, and orthogonal parameter adjustment. Specifically, HPA partitions parameters into two types based on their focus on safety or task performance, and selects the focused ones to preserve from a balanced perspective. In addition, HPA imposes orthogonality constraints on parameter updates to further alleviate catastrophic forgetting. Extensive experiments on the CVIT benchmark and safety evaluation datasets demonstrate that HPA better maintains high safety and mitigates forgetting than existing baselines. Code is available at https://github.com/Minato-Zackie/HPA.

Video snapshot compressive imaging (SCI) offers a promising alternative to high-speed cameras by encoding multiple frames into a single 2D measurement. However, SCI requires algorithms to reconstruct the high-speed video, and as resolution increases, reconstruction becomes computationally expensive and memory-intensive. Much of the resource is wasted on recovering large background regions that contain little useful information, highlighting the need for selective, object-driven reconstruction. Existing object detectors struggle to perform accurately on SCI measurements due to the spatial-temporal aliasing introduced by coded exposure. To address this challenge, we propose DetectSCI, the first framework enabling object-guided region-of-interest (ROI) reconstruction for high-resolution SCI. The inside detector comprises two key components: an encoder built from weight-sharing Mamba-Implicit Modules (MIM) for progressive feature refinement, and a Frequency Mamba (FM) module dedicated to frequency-aware query selection. MIM enhances features via multi-scale dilated convolutions and implicit representations, while FM restores discriminative details by decomposing and reweighting frequency bands. Experiments on the SportsMOT dataset show that DetectSCI achieves 80.9 Average Precision (AP), surpassing the best CNN-based detector by at least 2.8 AP and the best Transformer-based detector by at least 4.1 AP, while maintaining comparable efficiency. Code will be released.

As multimodal LLM-driven agents advance in autonomy and generalization, traditional static datasets face inherent scalability limitations and are insufficient for fully assessing their capabilities in increasingly complex and diverse tasks. Existing studies have attempted to generate agent tasks using LLMs, but due to the inherent hallucinations of LLMs and the lack of internal data relationship modeling, these tasks often exhibit semantic inconsistencies and solvability issues. To address these challenges, we introduce Graph2Eval, a knowledge-graph-driven framework for automated, scalable, and semantically grounded agent task generation. At its core, Graph2Eval leverages a knowledge graph built from heterogeneous external data sources as a structured task space, generating multimodal agent tasks through subgraph sampling and task construction guided by task templates and meta-path strategies. To further ensure task reliability, a multi-stage filtering pipeline based on node reachability analysis, LLM scoring, and similarity analysis ensures the diversity and solvability of the generated tasks. By unifying both RAG Agent and Web Agent scenarios, Graph2Eval enables efficient generation of multimodal document understanding tasks and multi-step web interaction tasks. We instantiate the framework with Graph2Eval-Bench, a curated dataset of 1,319 tasks spanning document understanding and web interaction scenarios. Extensive experiments show that, on average, Graph2Eval improves task semantic consistency by 20% and solvability by 17% over baselines, while Graph2Eval-Bench effectively distinguishes agent performance, offering a new perspective on automated agent evaluation.

Existing methods for human motion control in video generation typically rely on either 2D poses or explicit 3D parametric models (e.g., SMPL) as control signals. However, 2D poses rigidly bind motion to the driving viewpoint, precluding novel-view synthesis. Explicit 3D models, though structurally informative, suffer from inherent inaccuracies (e.g., depth ambiguity and inaccurate dynamics) which, when used as a strong constraint, override the powerful intrinsic 3D awareness of large-scale video generators. In this work, we revisit motion control from a 3D-aware perspective, advocating for an implicit, view-agnostic motion representation that naturally aligns with the generator's spatial priors rather than depending on externally reconstructed constraints. We introduce 3DiMo, which jointly trains a motion encoder with a pretrained video generator to distill driving frames into compact, view-agnostic motion tokens, injected semantically via cross-attention. To foster 3D awareness, we train with view-rich supervision--single-view, multi-view, and moving-camera videos--forcing motion consistency across diverse viewpoints. Additionally, we use auxiliary geometric supervision that leverages SMPL only for early initialization and is annealed to zero, enabling the model to transition from external 3D guidance to learning genuine 3D spatial motion understanding from the data and the generator's priors. Experiments confirm that 3DiMo faithfully reproduces driving motions with flexible, text-driven camera control, significantly surpassing existing methods in both motion fidelity and visual quality.

Recent advances in Vision-Language-Action (VLA) models demonstrate that visual signals can effectively complement sparse action supervisions. However, letting VLA directly predict high-dimensional visual states can distribute model capacity and incur prohibitive training cost, while compressing visual states into more compact supervisory signals inevitably incurs information bottlenecks. Moreover, existing methods often suffer from poor comprehension and reasoning capabilities due to the neglect of language supervision. This paper introduces Mantis, a novel framework featuring a Disentangled Visual Foresight (DVF) to tackle these issues. Specifically, Mantis decouples visual foresight prediction from the backbone with the combination of meta queries and a diffusion Transformer (DiT) head. With the current visual state provided to the DiT via a residual connection, a simple next-state prediction objective enables the meta queries to automatically capture the latent actions that delineate the visual trajectory, and hence boost the learning of explicit actions. The disentanglement reduces the burden of the VLA backbone, enabling it to maintain comprehension and reasoning capabilities through language supervision. Empirically, pretrained on human manipulation videos, robot demonstrations, and image-text pairs, Mantis achieves a 96.5% success rate on LIBERO benchmark after fine-tuning, surpassing powerful baselines while exhibiting high convergence speed. Real-world evaluations show that Mantis outperforms \pi_ 0.5 , a leading open-source VLA model, particularly in instruction-following capability, generalization to unseen instructions, and reasoning ability. We also introduce the adaptive temporal ensemble (ATE) strategy to balance computational efficiency and motion stability during inference, yielding the Mantis-ATE variant, which reduces inference counts by 45% while maintaining performance. We have released the code and model weights to the open-source community, which are available at https://github.com/SJTU-DENG-Lab/Mantis.

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

Modern scene reconstruction methods, such as 3D Gaussian Splatting, deliver photo-realistic novel view synthesis at real-time speeds, yet their adoption in interactive graphics applications has been limited. A major bottleneck is the difficulty of interacting with these representations compared to traditional, human-authored 3D assets. While previous research has attempted to impose semantic decomposition on these models, significant challenges remain regarding segmentation quality and consistency. To address this, we introduce Semantic Foam, extending the recently proposed Radiant Foam representations to semantic decomposition tasks.Our approach integrates the natural spatial volumetric decomposition of Radiant Foam's Voronoi mesh with an explicit semantic feature field parameterized at the cell level. This explicit structure enables direct spatial regularization, which prevents artifacts caused by occlusion or inconsistent supervision across views - common pitfalls for other point-based representations. Experimental results show that our method achieves superior object-level segmentation performance compared to state-of-the-art methods like Gaussian Grouping and SAGA.

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