CVPR 2026 Papers

Recent vision-language models (VLMs) achieve remarkable reasoning through reinforcement learning (RL), which provides a feasible solution for realizing continuous self-evolving large vision-language models (LVLMs) in the era of experience. However, RL for VLMs requires abundant high-quality multimodal data--especially challenging in specialized domains like chemistry, earth sciences, and multimodal mathematics. Existing strategies such as synthetic data and self-rewarding mechanisms suffer from limited distributions and alignment difficulties, ultimately causing reward hacking: models exploit high-reward patterns, collapsing policy entropy and destabilizing training. We propose DoGe (Decouple to Generalize), a dual-decoupling framework that guides models to first learn from context rather than problem solving by refocusing on the problem context scenarios overlooked by synthetic data methods. By decoupling learning process into dual components (Thinker and Solver), we reasonably quantify the reward signals of this process and propose a two-stage RL post-training approach from freely exploring context to practically solving tasks. Second, to increase the diversity of training data, DoGe constructs an evolving curriculum learning pipeline: an expanded native domain knowledge corpus and an iteratively evolving seed problems pool. Experiments show that our method consistently outperforms the baseline across various benchmarks, providing a scalable pathway for realizing self-evolving LVLMs.

We propose a multimodal-driven framework for high-fidelity long-term digital human animation termed Soul, which generates semantically coherent videos from a single-frame portrait image, text prompts, and audio, achieving precise lip synchronization, vivid facial expressions, and robust identity preservation. We construct Soul-1M, containing 1 million finely annotated samples with a precise automated annotation pipeline (covering portrait, upper-body, full-body, and multi-person scenes) to mitigate data scarcity, and we carefully curate Soul-Bench for comprehensive and fair evaluation of audio-/text-guided animation methods. The model is built on the Wan2.2-5B backbone, integrating audio-injection layers and multiple training strategies together with threshold-aware codebook replacement to ensure long-term generation consistency. Meanwhile, step/CFG distillation and a lightweight VAE are used to optimize inference efficiency, achieving an 11.4xspeedup with negligible quality loss. Extensive experiments show that Soul significantly outperforms current leading open-source and commercial models on video quality, video-text alignment, identity preservation, and lip-synchronization accuracy, demonstrating broad applicability in real-world scenarios such as virtual anchors and film production.

Developing generalizable deepfake detectors has become increasingly important with the rapid advancement of generative models. Adapting visual foundation models (VFMs), e.g., CLIP, through parameter-efficient finetuning (PEFT), with only a small subset of parameters updated, has been proven highly effective for generalizable detection. However, the success of "fewer-parameters" training raises an important question: although only a few parameters are tuned, have existing PEFT-based detectors truly exploited the most informative ones while eliminating redundant parameters for better generalization? In this work, we move beyond standard PEFT by proposing a joint optimization strategy that operates at both the layer and token levels. Since latent features across layers capture different semantic abstractions and tokens within the same layer convey varied forgery cues, we propose integrating both layer-level and token-level routing to maximize representational synergy. Specifically, at the layer level, we introduce "Early Layer Pruning", an adaptive truncation mechanism that enables the model to adaptively learn distinct forward depths for different types of instances. At the token level, "Token Selection" is guided by the Spearman rank loss to filter tokens irrelevant to forgery learning, enabling the model to focus on the most discriminative cues. Furthermore, a unified MoE architecture is applied that encourages diversity and thus reduces the potential model's overfitting to specific forgery types. Extensive benchmarking results demonstrate the effectiveness of our designs and show the superior performance of our method over existing state-of-the-arts.

Parameter-efficient fine-tuning (PEFT) methods have recently gained popularity for applying deep neural networks on small datasets as they reduce overfitting, simplify deployment, and enable fast training. We demonstrate that for dense image prediction tasks, a well-designed and lightweight dense readout on top of a frozen large backbone can surpass state-of-the-art PEFT methods in both efficiency and accuracy. Our parameter-efficient readout module combines interpolation and attention for fine-grained dense prediction. It integrates seamlessly with a wide range of pretrained vision backbones such as DINOv3 and can be combined with low-rank adaptation (LoRA). We achieve competitive or superior performance in semantic segmentation, object detection, pose estimation and semantic contour prediction, offering an efficient alternative to current PEFT techniques.

Multimodal Large Language Models (MLLMs) have shown impressive capabilities in jointly understanding text, images, and videos, often evaluated via Visual Question Answering (VQA). However, even state-of-the-art MLLMs struggle with domain-specific or knowledge-intensive queries, where relevant information is underrepresented in pre-training data. Knowledge-based VQA (KB-VQA) addresses this by retrieving external documents to condition answer generation, but current retrieval-augmented approaches suffer from low precision, noisy passages, and limited reasoning. To address this, we propose ReAG, a novel Reasoning-Augmented Multimodal RAG approach that combines coarse- and fine-grained retrieval with a critic model that filters irrelevant passages, ensuring high-quality additional context. The model follows a multi-stage training strategy leveraging reinforcement learning to enhance reasoning over retrieved content, while supervised fine-tuning serves only as a cold start. Extensive experiments on Encyclopedic-VQA and InfoSeek demonstrate that ReAG significantly outperforms prior methods, improving answer accuracy and providing interpretable reasoning grounded in retrieved evidence.

Synthesizing realistic human motion from natural language holds transformative potential for animation, robotics, and virtual reality. Recent methods handle single-action sequences and simple textual instructions, yet multi-action compositions and precise editing remain elusive due to limited data diversity, inadequate representations, and fragmented pipelines. Critically, most existing methods train motion generation models from scratch, failing to exploithe rich action semantics and long-horizon reasoning already encoded in pretrained Multimodal Large Language Models (MLLMs). Here we show that finetuning a pretrained MLLM with large-scale motion data yields strong zero-shot generalization across diverse text-guided motion generation and editing tasks. We present MotionMaster, a unified framework built on three components: MotionGB, a 10,000-hour dataset expanded from 400 hours of verified motion capture via spatial-temporal augmentation; an FSQ-based tokenizer that preserves both local joint accuracy and global trajectory coherence; and a finetuned MLLM with motion and language tokens in a shared embedding space. MotionMaster outperforms prior methods by 41.6% in multi-action semantic consistency and 20.8% in body-part composition. These results demonstrate that pretraining knowledge from MLLMs transfers effectively to motion understanding, opening a viable path toward general-purpose motion intelligence.

Novel view synthesis (NVS) under large view deviations remains an underexplored challenge for 3D Gaussian Splatting (3DGS). In urban scenes with limited training coverage, models often fail to maintain geometric consistency when extrapolating to unseen viewpoints, resulting in severe distortions and degraded rendering quality. We introduce Context-Aware Gaussian Splatting (CoRoGS), a Context-aware framework for Robust large-deviation novel view synthesis (LD-NVS) that embeds contextual reasoning into 3DGS. Instead of treating Gaussians as independent primitives, CoRoGS adopts a contextual formulation that explicitly models inter-Gaussian dependencies. This representation is implemented by constructing a 3D Gaussian graph, which propagates relational geometry and semantics via message passing, resulting in context-aware Gaussian updates. To further maintain structural consistency under substantial view deviation, we incorporate a progressive graph expansion strategy that adaptively grows and prunes Gaussians, leading to more coherent and complete scene reconstructions. Extensive experiments demonstrate that CoRoGS outperforms state-of-the-art 3DGS-based methods, producing higher-quality results. We highlight that CoRoGS robustly handles a wide range of view shifts, including lateral deviations (e.g., lane-level offsets) and cross-level transitions such as from ground-level driving views to elevated perspectives.

The ability to extrapolate dynamic 3D scenes beyond the observed timeframe is fundamental to advancing physical world understanding and predictive modeling. Existing dynamic 3D reconstruction methods have achieved high-fidelity rendering of temporal interpolation, but typically lack physical consistency in predicting the future. To overcome this issue, we propose ParticleGS, a physics-based framework that reformulates dynamic 3D scenes as physically grounded systems. ParticleGS comprises three key components: 1) an encoder that decomposes the scene into static properties and initial dynamic physical fields; 2) an evolver based on Neural Ordinary Differential Equations (Neural ODEs) that learns continuous-time dynamics for motion extrapolation; and 3) a decoder that reconstructs 3D Gaussians from evolved particle states for rendering. Through this design, ParticleGS integrates physical reasoning into dynamic 3D representations, enabling accurate and consistent prediction of the future. Experiments show that ParticleGS achieves state-of-the-art performance in extrapolation while maintaining rendering quality.

Recent advances in vision-language models (VLMs) have significantly enhanced the visual grounding task, which involves locating objects in an image based on natural language queries. Despite these advancements, the security of VLM-based grounding systems has not been thoroughly investigated. This paper reveals a novel and realistic vulnerability: the first multi-target backdoor attack on VLM-based visual grounding. Unlike prior attacks that rely on static triggers or fixed targets, we propose IAG, a method that dynamically generates input-aware, text-guided triggers conditioned on any specified target object description to execute the attack. This is achieved through a text-conditioned UNet that embeds imperceptible target semantic cues into visual inputs while preserving normal grounding performance on benign samples. We further develop a joint training objective that balances language capability with perceptual reconstruction to ensure imperceptibility, effectiveness, and stealth. Extensive experiments on multiple VLMs (e.g., LLaVA, InternVL, Ferret) and benchmarks (RefCOCO, RefCOCO+, RefCOCOg, Flickr30k Entities, and ShowUI) demonstrate that IAG achieves the best ASRs compared with other baselines on almost all settings without compromising clean accuracy, maintaining robustness against existing defenses, and exhibiting transferability across datasets and models. These findings underscore critical security risks in grounding-capable VLMs and highlight the need for further research on trustworthy multimodal understanding. Code is in the supplementary material.

Lane detection is a crucial task in autonomous driving, which is conducive to ensuring the safe operation of vehicles. However, current datasets like CULane and TuSimple have relatively limited data under extreme weather conditions, such as rain, snow and fog, which makes detection models unreliable in extreme conditions, potentially leading to serious safety-critical failures on the road. In this direction, we propose HG-Lane, a High-fidelity Generation framework for Lane Scenes under adverse weather and lighting conditions, without the need for re-annotation. Based on our framework, we further propose a benchmark that includes adverse weather and lighting conditions, with 30,000 images. Experiment results demonstrate that our method constantly and significantly improves the performance of all the related lane detection networks. Taking the state-of-the-art CLRNet as an example, the overall mF1 on our benchmark increases by 20.87%. The F1@50 for the overall, normal, snow, rain, fog, night, and dusk categories increases by 19.75%, 8.63%, 38.8%, 14.96%, 26.84%, 21.5%, and 12.04%, respectively. The Code and dataset are available at https://github.com/zdc233/HG-Lane.

Video amodal completion (VAC) aims to mimic the human brain's ability to implicitly perceive the complete appearance of partially occluded objects, thereby facilitating recognition and understanding. Existing VAC methods finetune video generation models on custom datasets, yet these datasets often have unrealistic distributions and small scales due to the challenges of collecting real amodal data, and thus limit their performance and generalization. To address this, we propose a training-free framework by utilizing pre-trained image inpainting models for VAC and introducing in-context (IC) learning to enhance inter-frame consistency. However, despite the satisfactory performance of DiT-based IC Learning in generation tasks, task-agnostic global attention tends to utilize irrelevant contextual information, resulting in completion failures when applied to amodal completion. Additionally, IC Learning faces a cold-start problem with the exemplar construction. To this end, we propose a consistent video amodal completion with rectified in-context exemplar guidance. Specifically, we introduce rectified exemplar-guided completion by adjusting the attention weights of exemplar relative to the target images for consistent completion, and adopt a dual-frame calibrated exemplar rectification to tackle the cold-start issue. Quantitative and qualitative experiments demonstrate that our method outperforms SOTAs, especially in terms of generalization and robustness on uncommon data and under severe occlusion. Codes are available at https://github.com/xxyCSCV/IC-Amodal.

Fully Few-shot Class-incremental Audio Classification (FFCAC) is challenging since the training samples are limited both in the incremental sessions and in the base session. Existing few-shot learning methods suffer from catastrophic forgetting and overfitting when applied to FFCAC.Pre-trained Audio-Language Models (ALMs) have achieved success in many audio learning tasks. However, we find that it is impractical to directly use ALM on FFCAC, since misalignment between text and audio causes even severe catastrophic forgetting and overfitting. We propose a Task-Adaptive Prototype Evolution (TAPE) framework to facilitate ALMs to tackle the challenges of FFCAC, which consists of two key components:(1) A Task-Adapter that isolates audio features in a metric space to mitigate catastrophic forgetting while preserving knowledge across sessions, and (2) A Prototype Evolution mechanism that dynamically refines class prototypes using query samples during inference, thereby enabling adaptive learning and reducing overfitting.To the best of our knowledge, we are the first to use ALMs on the FFCAC task. We conduct experiments on three audio datasets: NSynth-100 (instrument recognition), FSC-89 (event detection), and LBS-100 (voice recognition). The experimental results show that our proposed approach TAPE significantly surpasses the baselines. Specifically, it averagely improves upon the second best from 54.93% to 82.76% in terms of Average Accuracy (AA \uparrow), and from 28.74% to 12.56% in terms of Performance Dropping rate (PD \downarrow). The code for our work is available at https://github.com/YvoGao/TAPE.

We present Vanast, a unified framework that generates garment-transferred human animation videos directly from a single human image, garment images, and a pose guidance video. Conventional two-stage pipelines treat image-based virtual try-on and pose-driven animation as separate processes, which often results in identity drift, garment distortion, and front-back inconsistency. Our model addresses these issues by performing the entire process in a single unified step to achieve coherent synthesis. To enable this setting, we construct large-scale triplet supervision. Our data generation pipeline includes generating identity-preserving human images in alternative outfits that differ from garment catalog images, capturing full upper and lower garment triplets to overcome the single-garment-posed video pair limitation, and assembling diverse in-the-wild triplets without requiring garment catalog images. We further introduce a Dual Module architecture for video diffusion transformers to stabilize training, preserve pretrained generative quality, and improve garment accuracy, pose adherence, and identity preservation while supporting zero-shot garment interpolation. Together, these contributions allow Vanast to produce high-fidelity, identity-consistent animation across a wide range of garment types.

Recent advancements in video generation technologies have been significant, resulting in their widespread application across multiple domains. However, concerns have been mounting over the potential misuse of generated content. Tracing the origin of generated videos has become crucial to mitigate potential misuse and identify responsible parties. Existing video attribution methods require additional operations or the training of source attribution models, which may degrade video quality or necessitate large amounts of training samples. To address these challenges, we define for the first time the "few-shot training-free generated video attribution" task and propose SWIFT, which is tightly integrated with the temporal characteristics of the video. By leveraging the "Pixel Frames(many) to Latent Frame(one)" temporal mapping within each video chunk, SWIFT applies a fixed-length sliding window to perform two distinct reconstructions: normal and corrupted. The variation in the losses between two reconstructions is then used as an attribution signal. We conducted an extensive evaluation of five state-of-the-art (SOTA) video generation models. Experimental results show that SWIFT achieves over 90% average attribution accuracy with merely 20 video samples across all models and even enables zero-shot attribution for HunyuanVideo, EasyAnimate, and Wan2.2. Our source code is available at https://github.com/wangchao0708/SWIFT.

Multimodal Large Language Models (MLLMs) like LLaVA have demonstrated remarkable capabilities in multi-modal understanding and generation. This success motivates us to investigate whether the inherent prior knowledge embedded within such MLLMs contains sufficient spatial awareness for dense prediction tasks, without requiring any task-specific fine-tuning. Thus, in this paper, we explore the utilization of LLaVA for training-free open-vocabulary semantic segmentation. We discover that certain layers within the LLM part of LLaVA can generate localized features corresponding to given object classes. Building on this intrinsic capability, we design three modules: A question-answer pipeline to identify target classes in the image, a text-visual response module to extract initial reliable pixel-level activations for the target class, and a visual generation module to produce reliable refined prompts, which further serve as guidance for SAM to generate the predictions. Our LLaVA-based approach achieves new state-of-the-art performance on "Thing" category datasets, e.g., PASCAL VOC 2012 and COCO-object. Moreover, our method does not require explicit background class names, demonstrating its exceptional potential for handling open-world scenarios. Code is available at https://github.com/zbf1991/FSeg-LLaVA.

We introduce LAM, a system that explores collaboration between large language models and vision-language models to generate articulated objects from text prompts without a visual prior or prebuilt 3D assets. In contrast, we formulate articulated object generation as a unified code-generation task, in which geometry and articulations can be co-designed from scratch. Given an input text, LAM coordinates a team of specialized modules to generate code to represent the desired articulated object procedurally. LAM first reasons about the hierarchical structure of parts (links) with Link Designer, then writes code, compiles it, and debugs it with Geometry and Articulation Coders and self-corrects with Geometry and Articulation Checkers. The code serves as a structured, interpretable bridge between individual links, ensuring the correct relationships among them. Experiments demonstrate the power of leveraging code as a generative medium within a collaboration system, showcasing its effectiveness in automatically constructing complex articulated objects.

Video temporal grounding (VTG) is a critical task in video understanding and a key capability for extending video large language models (Vid-LLMs) to broader applications. However, existing Vid-LLMs rely on uniform frame sampling to extract video information, resulting in a sparse distribution of key frames and the loss of crucial temporal cues. To address this limitation, we propose Grounded Visual Token Sampling (GroundVTS), a Vid-LLM architecture that focuses on the most informative temporal segments. GroundVTS employs a fine-grained, query-guided mechanism to filter visual tokens before feeding them into the LLM, thereby preserving essential spatio-temporal information and maintaining temporal coherence. Futhermore, we introduce a progressive optimization strategy that enables the LLM to effectively adapt to the non-uniform distribution of visual features, enhancing its ability to model temporal dependencies and achieve precise video localization. We comprehensively evaluate GroundVTS on three standard VTG benchmarks, where it outperforms existing methods, achieving a 7.7-point improvement in mIoU for moment retrieval and 12.0-point improvement in mAP for highlight detection. Code is available at the \href https://github.com/Florence365/GroundVTS GroundVTS code repository .

Despite the rapid progress of Large Vision-Language Models (LVLMs), the integration of visual modalities introduces new safety vulnerabilities that adversaries can exploit to elicit biased or malicious outputs. In this paper, we demonstrate an underexplored vulnerability via semantic slot filling, where LVLMs complete missing slot values with unsafe content even when the slot types are deliberately crafted to appear benign. Building on this finding, we propose StructAttack, a simple yet effective single-query jailbreak framework under black-box settings. StructAttack decomposes a harmful query into a central topic and a set of benign-looking slot types, then embeds them as structured visual prompts (e.g., mind maps, tables, or sunburst diagrams) with small random perturbations. Paired with a completion-guided instruction, LVLMs automatically recompose the concealed semantics and generate unsafe outputs without triggering safety mechanisms. Although each slot appears benign in isolation (local benignness), StructAttack exploits LVLMs' reasoning to assemble these slots into coherent harmful semantics. Extensive experiments on multiple models and benchmarks show the efficacy of our proposed StructAttack.

Unsupervised deraining has attracted attention for its ability to learn the real-world distribution of rain without paired supervision. However, the lack of strong constraints makes it difficult for the network to converge, especially with the complex diversity of rain degradation. A key motivation is that high-quality deraining results occasionally emerge during training, which can be leveraged to guide the optimization process. To overcome these challenges, we introduce RGSUD (Reward-Guided Self-Reinforcement Unsupervised Image Deraining), comprising two key stages: reward recycling and self-reinforcement (SR) training. In the reward recycling stage, we propose an Image Quality Assessment (IQA)-based dynamic reward recycling mechanism that selects optimal derained outputs during training and continuously collects high-quality deraining images. In the SR training stage, we incorporate these rewards into the model's optimization process, constraining the optimization space and improving alignment between derained outputs and clean images. By leveraging IQA-based self reinforced loss and dynamically updated rewards, we enhance the quality of synthesized pseudo-paired data and stabilize the optimization. Extensive experiments demonstrate that our method achieves SOTA performance across multiple datasets, including paired synthetic, paired real, and unpaired real images, outperforming existing unsupervised deraining approaches in both subjective and objective IQA metrics. Additionally, we show that the Self Reinforcement Strategy is adaptable to other unsupervised deraining methods.