CVPR 2026 Papers

Vision-Language-Action (VLA) models convert high-level language instructions into concrete, executable actions, a task that is especially challenging in open-world environments. We present Visual Foresight Planning (ForeAct), a general and efficient planner that guides a VLA step-by-step using imagined future observations and subtask descriptions. With an imagined future observation, the VLA can focus on visuo-motor inference rather than high-level semantic reasoning, leading to improved accuracy and generalization. Our planner comprises a highly efficient foresight image generation module that predicts a high-quality 640x480 future observation from the current visual input and language instruction within only 0.33s on an H100 GPU, together with a vision-language model that reasons over the task and produces subtask descriptions for both the generator and the VLA. Importantly, state-of-the-art VLAs can integrate our planner seamlessly by simply augmenting their visual inputs, without any architectural modification. The foresight generator is pretrained on over 1 million multi-task, cross-embodiment episodes, enabling it to learn robust embodied dynamics. We evaluate our framework on a benchmark that consists of 11 diverse, multi-step real-world tasks. It achieves an average success rate of 87.4%, demonstrating a +40.9% absolute improvement over the \pi_0 baseline (46.5%) and a +30.3% absolute improvement over \pi_0 augmented with textual subtask guidance (57.1%).

Vision-language models (VLMs) excel at multimodal understanding, yet their text-only decoding forces them to verbalize visual reasoning, limiting performance on tasks that demand visual imagination. Recent attempts train VLMs to render explicit images, but the heavy image-generation pre-training often hinders the reasoning ability. Inspired by the way humans reason with mental imagery--the internal construction and manipulation of visual cues--we investigate whether VLMs can reason through interleaved multimodal trajectories without producing explicit images. To this end, we present a Machine Mental Imagery framework, dubbed as "Mirage", which augments VLM decoding with latent visual tokens alongside ordinary text. Concretely, whenever the model chooses to "think visually", it recasts its hidden states as next tokens, thereby continuing a multimodal trajectory without generating pixel-level images. Begin by supervising the latent tokens through distillation from ground-truth image embeddings, we then switch to text-only supervision to make the latent trajectory align tightly with the task objective. A subsequent reinforcement learning stage further enhances the multimodal reasoning capability. Experiments on diverse benchmarks demonstrate that \Model unlocks stronger multimodal reasoning without explicit image generation.

Recent advances in 3D shape generation have achieved impressive results, but most existing methods rely on clean, unoccluded, and well-segmented inputs. Such conditions are rarely met in real-world scenarios. We present ShapeR, a novel approach for conditional 3D object shape generation from casually captured sequences. Given a image sequence, we leverage off-the-shelf visual-inertial SLAM,3D detection algorithms and VLMs to extract for each object, a set of sparse SLAM points, posed multi-view images, and machine-generated captions. A rectified flow transformer trained to effectively condition on these modalities then generates high-fidelity metric 3D shapes. To ensure robustness to the challenges of casually captured data, we employ a range of techniques including on-the-fly compositional augmentations, a curriculum training scheme spanning object- and scene-level datasets, and strategies to handle background clutter. Additionally, we introduce a new evaluation benchmark comprising 178 in the wild objects across 7 real-world scenes with geometry annotations. Experiments show that ShapeR significantly outperforms existing approaches in this challenging setting, achieving an improvement of 2.7x in Chamfer distance compared to SoTA.

Despite rapid progress in video generation, existing models are incapable of producing vector animation, a dominant and highly expressive form of multimedia on the Internet. Vector animations offer resolution-independence, compactness, semantic structure, and editable parametric motion representations, yet current generative models operate exclusively in raster space and thus cannot synthesize them. Meanwhile, recent advances in large multimodal models demonstrate strong capabilities in generating structured data such as slides , 3D meshes , LEGO sequences , and indoor layouts , suggesting that native vector animation generation may be achievable. In this work, we present the first framework for tokenizing and autoregressively generating vector animations. We adopt Lottie, a widely deployed JSON-based animation standard, and design a tailored Lottie Tokenizer that encodes layered geometric primitives, transforms, and keyframe-based motion into a compact and semantically aligned token sequence. To support large-scale training, we also construct LottieAnimation-660K, the largest and most diverse vector animation dataset to date, consisting of 660k real-world Lottie animation and 15M static Lottie image files curated from broad Internet sources. Building upon these components, we finetune Qwen-VL to create LottieGPT, a native multimodal model capable of generating coherent, editable vector animations directly from natural language or visual prompts. Experiments show that our tokenizer dramatically reduces sequence length while preserving structural fidelity, enabling effective autoregressive learning of dynamic vector content. LottieGPT exhibits strong generalization across diverse animation styles and outperforms previous state-of-the-art models on SVG generation (a special case of single-frame vector animation).

Real-world documents combine text with tables, charts, photographs, and diagrams arranged in diverse layouts, yet existing research on multimodal large language models (MLLMs) for document QA predominantly produces text-only responses, underutilizing these visual elements. We introduce VinQA, a dataset designed for long-form answer generation where cited visual elements are explicitly interleaved with their supporting text and grounded in relevant document pages. To support this task, we study two encoding methods for feeding raw document page images into an MLLM, along with their visual-element citation mechanisms: (1) Page Encoding, which directly encodes full-page images with bounding boxes of visual elements and treats these boxed regions as citable units; and (2) Modality Encoding, which parses each page to extract text and crop visual elements, encodes them separately, and uses these cropped elements as citable units. In our experiments, we propose M-GroSE, a multimodal evaluation framework extending GroUSE to assess such answers along four dimensions: completeness, answer relevancy, faithfulness, and unanswerability. We additionally report Visual Source F1 to directly measure visual citation accuracy. Although proprietary frontier models still achieve the best overall scores on the VinQA test split, fine-tuning open Qwen2.5-VL models on the VinQA training split substantially improves their performance and markedly narrows this gap. Modality Encoding is initially more robust than Page Encoding for complex documents with long text, many visual elements, and diverse visual citation requirements. After training on VinQA, however, Page Encoding reaches a comparable performance level, showing that it can compete effectively even without the explicit parsing used in Modality Encoding. Finally, Visual G-Eval, an MLLM-based judge, confirms that fine-tuned models insert visual elements at semantically appropriate positions with faithful supporting text.

Video generation using recurrent architectures offers compelling efficiency advantages over attention-based transformers, particularly for long-sequence generation. However, chunked processing in recurrent models creates temporal discontinuities that harm long-range consistency. We introduce two complementary memory mechanisms to address this challenge at different granularities: (1) Context Memory maintains persistent global context within attention chunks through learnable sink columns and boundary buffers, adding only 150K parameters (\textless 0.1% overhead); (2) Latent Context-as-Memory (LCaM) extends memory across video segments by storing and retrieving historical latent embeddings, enabling cross-segment consistency without requiring camera annotations or frame reconstruction. Applied to Generalized Spatial-temporal Propagation Networks (GSTPN), our dual-memory approach achieves 1.54xfaster inference than attention-based transformers, while excelling in visual quality metrics. Our approach is particularly effective for knowledge distillation scenarios where only pre-extracted latent embeddings are available. This work demonstrates compelling efficiency-quality trade-offs for practical long video generation.

Recent advances in visual generation have increasingly explored the integration of reasoning capabilities. They incorporate textual reasoning, i.e., think, either before (as pre-planning) or after (as post-refinement) the generation process, yet they lack on-the-fly multimodal interaction during the generation itself. In this preliminary study, we introduce Thinking-while-Generating (TwiG), the first interleaved framework that enables co-evolving textual reasoning throughout the visual generation process. As visual content is progressively generating, textual reasoning is interleaved to both guide upcoming local regions and reflect on previously synthesized ones. This dynamic interplay produces more context-aware and semantically rich visual outputs. To unveil the potential of this framework, we investigate three candidate strategies, zero-shot prompting, supervised fine-tuning (SFT) on our curated TwiG-50K dataset, and reinforcement learning (RL) via a customized TwiG-GRPO strategy, each offering unique insights into the dynamics of interleaved reasoning. We hope this work inspires further research into interleaving textual reasoning for enhanced visual generation. Code is released at: https://github.com/ZiyuGuo99/Thinking-while-Generating.

Aerial-ground visual localization is a challenging task due to the significant differences in scene scale and view point captured between two views. In this work, we explore the practical benefit of jointly learning camera calibration and bird's-eye-view (BEV) projection for estimating full 6 Degrees-of-freedom relative camera pose between uncalibrated aerial and ground views. We present Visual Geometry Alignment (VGA), a unified framework that jointly learns a global gravity-alignment prior inferred from dense monocular perspective fields, and a planar alignment prior complementing the unobserved azimuth angle through Procrustes alignment in a shared BEV plane. At inference, we jointly refine the relative camera pose by integrating the predicted per-camera gravity alignment and relative planar azimuth angle, yielding improved orientation and translation alignment from visual input with extreme wide base-lines and limited overlap. We evaluate our method on challenging MatrixCity, ACC-NVS1 and ULTRRA ground-aerial pairs, demonstrating that optimizing with learned geometric priors can further improve the camera pose estimation across diverse altitudes and environment.

Visual environments are inherently hierarchical, as a panoramic view naturally encompasses and organizes multiple perspective views within its field. Capturing this hierarchy is crucial for effective perspective-to-equirectangular (P2E) visual place recognition. In this work, we introduce HypeVPR, a hierarchical embedding framework in hyperbolic space specifically designed to address the challenges of P2E matching. HypeVPR leverages the intrinsic ability of hyperbolic space to represent hierarchical structures, allowing panoramic descriptors to encode both broad contextual information and fine-grained local details. To this end, we propose a hierarchical feature aggregation mechanism that organizes local-to-global feature representations within hyperbolic space. Furthermore, HypeVPR's hierarchical organization naturally enables flexible control over the accuracy-efficiency trade-off without additional training, while maintaining robust matching across different image types. This approach enables HypeVPR to achieve competitive performance while significantly accelerating retrieval and reducing database storage requirements. Project page: https://suhan-woo.github.io/HypeVPR/

Audio-visual large language models (AVLLMs) have made significant strides in understanding visual and auditory content. However, their ability to capture fine-grained temporal relationships between audio and visual streams remains insufficiently evaluated. To address this, we introduce FAVE (Fine-grained Audio-Visual Temporal Evaluation), a comprehensive benchmark targeting three core dimensions of temporal perception: cross-modal temporal alignment (FAVE-align), event temporal relationship (FAVE-low), and detailed moment captioning (FAVE-high). To construct FAVE, we propose a scalable annotation pipeline that integrates shot boundary detection, automated captioning, and GPT-assisted refinement to produce temporally grounded, high-quality data. Extensive experiments on twelve state-of-the-art multimodal LLMs, both open-source and closed-source, reveal key limitations in multimodal integration, temporal relationship and timestamp localization, especially for joint audio-visual tasks. These findings highlight the need for better temporal modeling to improve AVLLMs' understanding of real-world video content. FAVE serves as a rigorous testbed for advancing temporally aware multimodal systems, and will be publicly released upon acceptance.

Current visual text generation models struggle with the trade-off between text accuracy and overall image coherence. We find that achieving high text accuracy can reduce aesthetic quality and instruction-following capability. Although reinforcement learning approaches can alleviate the problem through aligning with multiple rewards, they are often unstable for text generation, as existing approaches normally optimize multiple rewards in a weighted-sum way. In addition, it is difficult to balance the weight of each reward. Moreover, reinforcement learning requires a set of training instructions. A large number of prompts require more training time and computing resources, while a small set leads to poor performance. Hence, how to select the prompts for efficient training is an unsolved problem. In this study, we propose Pareto-Optimal Curriculum Alignment (POCA), a framework that addresses this issue as a multi-objective problem by: 1) identifying the Pareto-optimal set to avoid simple scalarization and 2) designing an adaptive curriculum alignment strategy to manage a learning sequence of a multi-reward dataset using automatic difficulty assessment, which is crucial for optimal convergence as RL methods explore in a limited data environment. In synergy, POCA finds the Pareto-optimal set in a unified reward space, which eliminates inconsistent signals to find the best trade-off solution from different rewards under an easy-to-hard optimization landscape. The experimental results show that POCA significantly improves all metrics such as CLIP, HPS scores and sentence accuracy.

A sketch is a distilled form of visual abstraction that conveys core concepts through simplified yet purposeful strokes while omitting extraneous detail. Despite its expressive power, quantifying the efficiency of semantic abstraction in sketches remains challenging. Existing evaluation methods that rely on reference images, low-level visual features, or recognition accuracy do not capture abstraction, the defining property of sketches.To address these limitations, we introduce SEA (Sketch Evaluation metric for Abstraction efficiency), a reference-free metric that assesses how economically a sketch represents class-defining visual elements while preserving semantic recognizability. These elements are derived per class from commonsense knowledge about features typically depicted in sketches. SEA leverages a visual question answering model to determine the presence of each element and returns a quantitative score that reflects semantic retention under simple visual representations.To support this metric, we present CommonSketch, the first semantically annotated sketch dataset, comprising 23,100 human-drawn sketches across 300 classes, each paired with a caption and element-level annotations. Experiments show that SEA aligns closely with human judgments and reliably discriminates levels of abstraction efficiency, while CommonSketch serves as a benchmark providing systematic evaluation of element-level sketch understanding across various vision-language models.

Large Vision-Language Models (LVLMs) often suffer from object hallucination, generating objects that are absent from the image. Prior work largely attributes this to insufficient visual attention. However, in this work, we are surprised to find that both real and hallucinated objects receive equally strong visual attention in the model's mid-to-late layers. This indicates that the key issue may not be how much the model attends, but **what it attends to and why**. To this end, we decode the visual features of high-attention regions using Logit Lens, and observe that high-attention regions corresponding to real objects can be correctly decoded to the target object token, whereas those for hallucinated objects cannot. Building on this, we identify two distinct hallucination mechanisms: **(i) visual uncertainty**, triggered by semantically similar or confusable regions, masking these regions eliminates the hallucination. **(ii) contextual prior**, triggered by strong co-occurrence priors, even when the initially attended region is masked, the hallucination persists and attention drifts to other regions. Based on these findings, we propose a simple yet effective training-free **Detect-Mitigate framework** comprising a Logit-Lens Consistency Check to detect hallucination and targeted remedies: High-Attention Regions Masking (HARM) for visual uncertainty hallucination, and Visual Evidence Enhanced Decoding (VEED) for contextual prior hallucination, which leverages genuine visual evidence to suppress erroneous priors. Our approach achieves state-of-the-art results on multiple hallucination benchmarks.

Understanding egocentric videos plays a vital role for embodied intelligence. Recent multi-modal large language models (MLLMs) can accept both visual and audio inputs. However, due to the challenge of obtaining text labels with coherent joint-modality information, whether MLLMs can jointly understand both modalities in egocentric videos remains under-explored. To address this problem, we introduce EgoAVU, a scalable data engine to automatically generate egocentric audio-visual narrations, questions and answers. EgoAVU enriches human narrations with multimodal context and generates audio-visual narrations through cross-modal correlation modeling. Token-based video filtering and modular, graph based curation ensure both the data diversity and quality. Leveraging EgoAVU, we construct EgoAVU-Instruct - a large scale training dataset of 3M samples, and EgoAVU-Bench - a manually verified evaluation split of 3K samples. EgoAVU-Bench clearly reveals the limitation of existing MLLMs: they bias heavily towards visual signals, often neglecting audio cues or failing to correspond audio with the visual source. Finetuning MLLMs on EgoAVU-Instruct effectively solves this issue, enabling up to 113% performance improvement on EgoAVU-Bench. Such benefit can also transfer to other benchmarks such as EgoTempo and EgoIllusion, achieving up to 28% relative performance gain. Project page: https://cs20s030.github.io/EgoAVU/

Unified Multimodal Models (UMMs) are often constrained by the pre-training of their visual generation components, which typically relies on inefficient paradigms and scarce, high-quality text-image paired data. In this paper, we systematically analyze pre-training recipes for UMM visual generation and identify these two issues as the major bottlenecks. To address them, we propose Image-Only Training for UMMs (IOMM), a data-efficient two-stage training framework. The first stage pre-trains the visual generative component exclusively using abundant unlabeled image-only data, thereby removing the dependency on paired data for this costly phase. The second stage fine-tunes the model using a mixture of unlabeled images and a small curated set of text-image pairs, leading to improved instruction alignment and generative quality. Extensive experiments show that IOMM not only improves training efficiency but also achieves state-of-the-art (SOTA) performance. For example, our IOMM-B (3.6B) model was trained from scratch using only about 1050 H800 GPU hours (with the vast majority, 1000 hours, dedicated to the efficient image-only pre-training stage). It achieves 0.89 on GenEval and 0.55 on WISE--surpassing strong baselines such as BAGEL-7B (0.82 and 0.55) and BLIP3-o-4B (0.84 and 0.50). Code is available at: https://github.com/LINs-lab/IOMM

Recent advances in audio-visual representation learning have shown the value of combining contrastive alignment with masked reconstruction. However, jointly optimizing these objectives in a single forward pass forces the contrastive branch to rely on randomly visible patches designed for reconstruction rather than cross-modal alignment, introducing semantic noise and optimization interference. We propose TG-DP, a Teacher-Guided Dual-Path framework that decouples reconstruction and alignment into separate optimization paths. By disentangling the masking regimes of the two branches, TG-DP enables the contrastive pathway to use a visibility pattern better suited to cross-modal alignment. A teacher model further provides auxiliary guidance for organizing visible tokens in this branch, helping reduce interference and stabilize cross-modal representation learning. TG-DP achieves state-of-the-art performance in zero-shot retrieval. On AudioSet, it improves R@1 from 35.2% to 37.4% for video-to-audio retrieval and from 27.9% to 37.1% for audio-to-video retrieval. The learned representations also remain semantically robust, achieving state-of-the-art linear-probe performance on AS20K and VGGSound. Taken together, our results suggest that decoupling multimodal objectives and introducing teacher-guided structure into the contrastive pathway provide an effective framework for improving large-scale audio-visual pretraining.

Generative video-to-audio (V2A) models produce highly plausible soundtracks, but it remains unclear whether they capture the underlying physical processes. Existing evaluations emphasize perceptual realism and overlook physical correctness under controlled interventions. In this paper, we introduce FlatSounds, a benchmark that audits the physical reasoning of V2A models through: 1) controlled counterfactual pairs in which a single physical factor is varied, and 2) single-video pattern tests that probe internal consistency and directional trends. These settings test whether the generated audio correctly reflects specific physical properties and timings. Our evaluation of state-of-the-art models reveals a consistent trade-off: models rely more on text captions than the visual stream to infer physics and semantics. Captions generally improve physical and semantic accuracy, but paradoxically degrade temporal alignment. Our results highlight the need to move beyond audio quality toward learning physical processes directly from pixels. Finally, we find that our physics-based metrics correlate strongly with human preference tests on our own data.

Visual grounding, localizing objects from natural language descriptions, represents a critical bridge between language and vision understanding. While multimodal large language models (MLLMs) achieve impressive scores on existing benchmarks, a fundamental question remains: can MLLMs truly visually ground with human-like sophistication, or are they merely pattern-matching on simplified datasets? Current benchmarks fail to capture real-world complexity where humans effortlessly navigate intricate references and recognize when grounding is impossible. To rigorously assess MLLMs' true capabilities, we introduce GroundingME, a benchmark that systematically challenges models across four critical dimensions: (1) Discriminative: distinguishing highly similar objects, (2) Spatial: understanding complex relational descriptions, (3) Limited: handling occlusions or tiny objects, and (4) Rejection: recognizing ungroundable queries. Through careful curation combining automated generation with human verification, we create 1,005 challenging examples mirroring real-world complexity. Evaluating 25 state-of-the-art MLLMs reveals a profound capability gap: the best model achieves only 45.1% accuracy, while most score 0% on rejection tasks. We explore two strategies for improvements: (1) test-time scaling selects optimal response by thinking trajectory to improve overall performance by up to 4.5%, and (2) data-mixture training boosts rejection accuracy from 0% to 27.9%. GroundingME thus serves as both a diagnostic tool revealing current limitations in MLLMs and a roadmap toward human-level visual grounding. Project page: https://groundingme.github.io

Multilingual text-to-image (T2I) models have advanced rapidly in terms of visual realism and semantic alignment, and are now widely utilised. Yet outputs vary across cultural contexts: because language carries cultural connotations, images synthesized from multilingual prompts should preserve cross-lingual cultural consistency. We conduct a comprehensive analysis showing that current T2I models often produce culturally neutral or English-biased results under multilingual prompts.Analyses of two representative models indicate that the issue stems not from missing cultural knowledge but from insufficient activation of culture-related representations. We propose a probing method that localizes culture-sensitive signals to a small set of neurons in a few fixed layers. Guided by this finding, we introduce two complementary alignment strategies: (1) inference-time cultural activation that amplifies the identified neurons without backbone fine-tuned; and (2) layer-targeted cultural enhancement that updates only culturally relevant layers. Experiments on our CultureBench demonstrate consistent improvements over strong baselines in cultural consistency while preserving fidelity and diversity.