ECCV 2026 Papers

Subject-driven video generation (SDV-Gen) aims to produce videos of a specific subject by adapting a pretrained video model, enabling personalized and application-driven content creation. To achieve this goal, per-subject tuning methods require approximately 200 A100 GPU hours to generate a customized video, whereas zero-shot methods avoid per-subject tuning but typically rely on millions of subjectvideo pairs for the supervision, incurring massive network fine-tuning costs (10K–200K A100 GPU hours). We propose a dataand computeefficient zero-shot SDV-Gen framework that avoids test-time per-subject tuning and the use of large-scale subject-video pairs. Our key idea decomposes SDV-Gen into (i) identity injection learned from subject-image pairs and (ii) motion-awareness preservation maintained by a small set of arbitrary videos. We optimize the two tasks with stochastic switching, using random reference-frame sampling and image-token dropout to prevent trivial first-frame copying. Our gradient analysis shows that the two objectives rapidly evolve toward nearly orthogonal update subspaces, explaining the stable optimization. Using CogVideoX-5B, we adapt a single model with 200K subject-image pairs and 4,000 arbitrary videos in 288 A100 GPU hours. This yields about 1% of compute compared to prior zero-shot baselines (i.e., 0.4% of VACE and 2.8% of Phantom) while using no subject-video pairs, yet remaining competitive in subject fidelity and motion quality. We show that the same recipe transfers to Wan 2.1-1.3B and Wan 2.2-5B.

Recent diffusion models enable high-quality video generation, but suffer from slow runtimes. The large transformer-based backbones used in these models are bottlenecked by spatiotemporal attention. In this paper, we identify that a significant fraction of token-totoken connections consistently yield negligible scores across various inputs, and their patterns often repeat across queries. Thus, the attention computation in these cases can be skipped with little to no effect on the result. This observation continues to hold for connections among local token blocks. Motivated by this, we introduce CalibAtt, a training-free method that accelerates video generation via calibrated sparse attention. CalibAtt performs an offline calibration pass that identifies block-level sparsity and repetition patterns that are stable across inputs, and compiles these patterns into optimized attention operations for each layer, head, and diffusion timestep. At inference time, we compute the selected input-dependent connections densely, and skip the unselected ones in a hardware-efficient manner. Extensive experiments on Wan 2.1 14B, Mochi 1, and few-step distilled models at various resolutions show that CalibAtt achieves up to 1.58× end-to-end speedup, outperforming existing training-free methods while maintaining video generation quality and text-video alignment.

End-to-end Vision-Language Models (VLMs) show immense potential in autonomous driving. However, standard Supervised FineTuning (SFT) often suffers from reasoning hallucinations and conservative biases. While traditional tool-augmented frameworks and Chain-ofThought (CoT) approaches mitigate these issues, they incur exorbitant token consumption and unacceptable latency, rendering real-time deployment impractical. To resolve this reliability-efficiency trade-off, we propose CritiqueDriveVLM, a novel unified three-stage framework internalizing reasoning directly into the VLM. First, we introduce Critique-Driven Multi-Turn Reinforcement Learning (RL) guided by a multi-dimensional verifier. By providing granular scalar feedback and a multi-turn penalty, we force the policy to internalize logical deduction, cultivating a robust System-2 Teacher that achieves high accuracy without fragile external tools. Subsequently, we propose Latent Thought Distillation to overcome the latency bottleneck. By aligning the Student’s latent representations with the Teacher’s fully converged reasoning states, we compress deep logical capabilities into a fast, CoT-free System-1 Student. Extensive experiments on the widely-used DriveLMM-o1 benchmark demonstrate remarkable improvements. Compared to the base model, our tool-free Teacher significantly boosts Multiple Choice Quality (MCQ) from 55.54% to a state-of-the-art 76.54%. Crucially, our distilled Student preserves competitive reasoning depth while drastically minimizing generation length to an average of merely 28 tokens. This slashes inference latency by 88% (from 3482 ms to 416 ms), paving a highly robust pathway for low-latency autonomous driving. Our source code is available at https://github.com/MICLAB-BUPT/CritiqueDriveVLM.

Ensuring prompt faithfulness remains a central challenge for text-to-image diffusion models. Classifier-Free Guidance (CFG) improves prompt adherence but exhibits an inherent quality–alignment tension: increasing the guidance scale to strengthen conditioning on the prompt often degrades visual quality and introduces artifacts. To probe the limit of greedy alignment maximization, we first introduce SAP (Semantic Alignment Projection), a greedy update rule that projects each sampling update onto the guidance direction to maximize per-step alignment progress. We then show that SAP can fail due to the loss of orthogonal corrective freedom, discarding high-dimensional components that are crucial for rectifying accumulated trajectory drift. Based on this diagnosis, we propose SCALE (Semantic-CALibrated Guidance Enhancement), a drop-in, training-free guidance mechanism. SCALE selectively amplifies the semantic component along the guidance direction while preserving the orthogonal corrective component, improving semantic alignment without compromising structural fidelity. Across multiple text-to-image benchmarks, SCALE delivers substantial and consistent gains in prompt adherence and compositional alignment with negligible overhead over standard sampling. Project page: https://github.com/SudongCAI/SCALE.

Rolling shutter (RS) cameras equip virtually all consumerdevices, yet RS-aware relative pose estimation has remained impracti-cal: the state-of-the-art solver requires a minimum of 20 point corre-spondences, making RANSAC-based robust estimation prohibitively ex-pensive due to the exponential dependence of the iteration count onthe sample size. We make RS relative pose estimation practical by in-troducing affine correspondences (ACs) into the RS two-view geome-try. We derive novel RS-corrected affine constraints that account forthe coupling between point perturbations and the row-dependent es-sential matrix, providing two equations per correspondence beyond thestandard epipolar constraint. Building on these constraints, we developa linearized algebraic solver that estimates pose and RS motion fromonly 7 ACs. The solver exploits the physical smallness of RS parametersto linearize the constraints, eliminates the 12 RS unknowns via null-space projection, and solves the remaining degree-20 system via actionmatrices in 1.2 ms. On the TUM RS benchmark, our method achievesthe best pose and RS parameter accuracy among all tested methodsand, uniquely among RS solvers, provides accurate translational veloc-ity estimates – which are poorly conditioned from point correspondencesalone due to a v-t coupling. On the global-shutter EuRoC MAV dataset,the solver achieves comparable accuracy to the standard 5-point algo-rithm, demonstrating that it generalizes well to the GS setting. Code isat https://github.com/danini/rolling_shutter_made_practical.

Video Variational Auto-Encoders (Video VAEs) compress video data from the highly redundant pixel space into a compact latent representation, playing an important role in state-of-the-art video generation models. However, existing methods typically learn inter-frame correlations implicitly, overlooking the potential of breaking down video compression into two separate parts: keyframe encoding and inter-frame dynamic encoding, which is a fundamental design of traditional video codecs. To address this, we incorporate traditional video codec standard design into the Video VAE and introduce VC-VAE, a model that explicitly separates keyframe and inter-frame dynamic compression. We start by establishing a high-fidelity static keyframe anchor through initialization from a powerful pre-trained image VAE. Then, to explicitly model dynamic relative to this anchor, we introduce the Temporal Dynamic Difference Convolution (TDC), an operator designed to learn sparse motion residuals from inter-frame differences while maintaining a separate pathway for static content. Qualitative and quantitative experiments show that our proposed VC-VAE significantly outperforms baseline models in reconstruction quality, dynamic modelling, and training efficiency.

The performance of deep spiking neural networks (SNNs) often relies on batch normalization (BN). However, the advanced dynamic BN variants used in state-of-the-art models introduce runtime multiplications, which weaken the hardware-efficiency motivation of SNNs. To address this tension, we identify catastrophic firing-rate decay as a primary cause of severe performance degradation in normalization-free SNNs. Guided by this insight, this work proposes the Intrinsically Stable SNN (IS-SNN) architecture, which removes activation-normalization layers by enforcing signal homeostasis through topology-aware weight standardization and modified residual connections. By folding the standardization operations into static weights offline, IS-SNN removes the runtime statistics tracking and multiplications introduced by activation normalization, restoring an accumulation-oriented inference datapath. Comprehensive experiments show that IS-SNN achieves performance competitive with or superior to computationally expensive dynamic BN techniques across VGG, ResNet, and Transformer-based models. Notably, it achieves a competitive accuracy of 68.05% on ImageNet and overcomes the severe depth limitations of prior BN-free attempts. Together with a 96.4% reduction in FPGA lookup table resource consumption for neuron implementations, these results support IS-SNN as a practical framework for building accurate and hardware-friendly deep neuromorphic systems.

Microscaling floating-point (MXFP) formats have emerged as a promising standard for deploying Multi-modal Large Language Models (MLLMs) and Large Language Models (LLMs) on modern accelerator architectures. However, existing Post-Training Quantization (PTQ) methods, particularly rotation-based techniques designed for integer formats, suffer from severe performance collapse when applied to MXFP4. Recent studies attribute this failure to a fundamental format mismatch: global orthogonal rotations inadvertently transfer outlier energy across quantization blocks, inducing new outliers that disrupt local block-wise scaling, while often creating bimodal activation distributions that underutilize the limited quantization range. To address these issues, we propose BATQuant (Block-wise Affine Transformation), which restricts transformations to align with MXFP granularity to prevent cross-block outlier propagation, while relaxing orthogonality constraints to optimize distribution shaping. To ensure parameter efficiency, we introduce Global and Private Kronecker (GPK) decomposition to effectively reduces storage and runtime overhead and incorporate Block-wise Learnable Clipping to suppress residual outliers. Extensive experiments on both MLLMs and LLMs demonstrate that BATQuant establishes new state-of-theart results under aggressive W4A4KV16 configurations, recovering up to 96.43% of full-precision performance on multimodal benchmarks and clearly outperforming existing methods across diverse tasks.

We propose A2-Edit, a unified inpainting framework for arbitrary object categories, which allows users to replace any target region with a reference object using only a coarse mask. To address the issues of severe homogenization and limited category coverage in existing datasets, we construct a large-scale, multi-category dataset UniEdit-500K, which includes 8 major categories, 209 fine-grained subcategories, and a total of 500,104 image pairs. Such rich category diversity poses new challenges for the model, requiring it to automatically learn semantic relationships and distinctions across categories. To this end, we introduce the Mixture of Transformer module, which performs differentiated modeling of various object categories through dynamic expert selection, and further enhances cross-category semantic transfer and generalization through collaboration among experts. In addition, we propose a Mask Annealing Training Strategy (MATS) that progressively relaxes mask precision during training, reducing the model’s reliance on accurate masks and improving robustness across diverse editing tasks. Extensive experiments on benchmarks such as VITON-HD and AnyInsertion demonstrate that A2-Edit consistently outperforms existing approaches across all metrics, providing a new and efficient solution for arbitrary object editing. The code is released at https://github.com/huayu-zheng/A2Edit.

Spatial reasoning from a single RGB image, including the inference of distances, sizes, and inter-object relationships without depth sensors or multiple views, remains a critical blind spot for current Vision Language Models (VLMs). Existing benchmarks either rely on multiview video sequences that expose explicit geometric cues, or are confined to indoor environments too small for model training. To close this gap, we introduce MonoSR, a large-scale dataset for open-world monocular spatial reasoning comprising over 1M QA pairs from 230K images spanning indoor, outdoor, and object-centric domains across 98 semantic categories. Every QA pair is guaranteed answerable from a single RGB image via a four-stage observability filter validated by human audit. Tasks are organized into three cognitive levels: foundational perception, perspective-aware imagination, and situational reasoning, supporting both large-scale training and fine-grained evaluation. Comprehensive benchmarking of state-of-the-art openand closed-source VLMs reveals consistent limitations across all three domains, with object-centric metric estimation emerging as the most challenging frontier. We further conduct a systematic auxiliary information study by injecting scene context, 2D visual prompts, and 3D bounding boxes, allowing us to quantify the geometric gap that future monocular perception modules must close and provide actionable design guidelines. Our project page is available at this link.

Multi-teacher distillation has emerged as a way to combine complementary teacher models into a single student model that exhibits the strengths of all its teachers. The student is trained to mimic the output of the teachers on a set of images, typically the union of the individual teacher’s training sets, assuming this data is available. In this paper, we question that assumption and explore alternative options. We first study how far one can go when distilling from teachers fed with different types of noise. Then, we show that information contained in the teachers can be leveraged to tailor the noise for multi-teacher distillation: we propose a method that, thanks to decorrelation losses at both patch and image levels, generates teacher-specific, improved samples optimized for data-free distillation. Experiments show that our most effective samples, IDeaL, lead to strong students that successfully capture complementary information from the teachers, yielding surprisingly competitive results that substantially narrow the gap with students distilled from real images. Moreover, given a limited budget of 1K images for distillation, students distilled using our IDeaL samples match or surpass the performance of those distilled using a 1K-image subset of ImageNet.

Referring multi-object tracking (RMOT) aims to track every instance in a video that matches a given language expression. Despite the recent integration of multimodal large language models (MLLMs) to enhance generalization, existing methods predominantly relegate them to the role of caption generators, necessitating external modules for final decision-making. This paradigm not only introduces extra latency but also severely underutilizes the inherent vision–language alignment capabilities of MLLMs. To address these limitations, we propose YesTrack, a novel two-stage RMOT method that reformulates referring as a discriminative task, directly leveraging MLLMs for Yes/No verification without explicit text generation. To further enhance the reliability and efficiency of this MLLM-based verification, we introduce two lightweight temporal consistency constraints: Temporal Confidence Prior (TCP) and Temporal Reference Propagation (TRP). We further validate the generality of this discriminative paradigm by proposing YesTrack-MOT, a straightforward yet highly effective instantiation for generic multiobject tracking (MOT). Experiments on Refer-KITTI and Refer-KITTIV2 show that YesTrack significantly outperforms existing state-of-theart methods while maintaining high efficiency, even when implemented with the smallest variant of Qwen3-VL. Code is released at https: //github.com/ggbondrighthere24/YesTrack.

Despite rapid advances in video generative models, robust metrics for evaluating visual and temporal correctness of complex human actions remain elusive. Critically, existing pure-vision encoders and Multimodal Large Language Models (MLLMs) are strongly appearancebiased, lack temporal understanding, and thus struggle to discern intricate motion dynamics and anatomical implausibilities in generated videos. We tackle this gap by introducing a novel evaluation metric derived from a learned latent space of real-world human actions. Our method first captures the nuances, constraints, and temporal smoothness of real-world motion by fusing appearance-agnostic human skeletal geometry features with appearance-based features. We posit that this combined feature space provides a robust representation of action plausibility. Given a generated video, our metric quantifies its action quality by measuring the distance between its underlying representations and this learned real-world action distribution. For rigorous validation, we develop a new multi-faceted benchmark specifically designed to probe temporally challenging aspects of human action fidelity. Through extensive experiments, we show that our metric achieves substantial improvement of more than 68% compared to existing state-of-the-art methods on our benchmark, performs competitively on established external benchmarks, and has a stronger correlation with human perception. Our in-depth analysis reveals critical limitations in current video generative models and establishes a new standard for advanced research in video generation. Code is available at https://xthomasbu.github.io/video-gen-evals/

We address the problem of recovering high-speed videos from dynamic scenes under extreme photon sparsity. Existing methods rely on aggregating photon detections in local spatiotemporal windows to improve signal-to-noise ratio; however, this local grouping discards global structure and fails in low-light regimes where photon detections are sparse in space and time. In this work, we show that the information needed to recover both motion and illumination is encoded in correlations over the full space-time pattern of photon arrivals. Building on this insight, we develop a spatiotemporal flux probing theory and an algorithm that estimates the Fourier coefficients of the underlying intensity directly from the photon stream. We demonstrate that our approach (1) recovers fast motion and temporal illumination dynamics with substantially fewer photons than prior methods, (2) enables velocity-selective videography that automatically refocuses video onto specific detected motions, and (3) generalizes across sensing modalities including singlephoton, event, and spike cameras.

Table recognition (TR) aims to transcribe table images into semi-structured representations (e.g., HTML or Markdown). However, current methods still struggle with complex structures, particularly involving merged cells, which are essential for accurate table parsing and downstream understanding tasks. While some works improve TR through large-scale training and global supervision, these local fine-grained yet crucial merged-cell attributes are often overlooked, becoming a key bottleneck for further progress. To tackle this, we introduce Table-MCR2TR, a reinforced multimodal large language model framework that leverages the enhanced merged-cell recognition (MCR) ability as contextual guidance to improve table recognition quality. To enable this capability, we first develop a high-fidelity data pipeline to construct a large-scale table dataset covering many complex merged cells, alleviating their scarcity in existing datasets. Additionally, the merged-cell-aware table recognition (MCATR) optimization strategy is proposed to improve table parsing via collaborative reinforcement learning. The core of MCATR lies in integrating three complementary objectives: TR and MCR respectively learn the global table structure and fine-grained merged-cell information; based on these learned signals, the MCR2TR alignment task transfers merged-cell knowledge to guide TR generation. Experimental results show that Table-MCR2TR achieves SOTA across multiple benchmarks, with average TR/MCR accuracy of 92.3%/83.9%, outperforming top-tier models and further boosting downstream table QA performance.

We push the frontier of large-scale spatial intelligence in Vision-Language Models (VLMs) and introduce the first benchmark that probes geographical layout understanding from real-world videos, spanning up to 1km distances. Inspired by the cognitive science literature, we evaluate models against the hierarchical stages of human spatial awareness: anchoring via landmarks, connecting them through routes, and integrating these into global mental maps. Extensive experiments reveal a fundamental divergence in how current AI models process spatial information. Instead of utilising true path integration or forming geometric survey knowledge, we find that VLMs rely almost entirely on 2D visual recognition and text-matching to bypass complex spatial reasoning. The benchmark is publicly available at https://perception-testchallenge.github.io/kilometervision.html.

Active learning (AL) aims to build high-quality labeled datasets by iteratively selecting the most informative samples from an unlabeled pool under limited annotation budgets. However, in fine-grained image classification, assessing this informativeness reliably is especially challenging due to subtle differences between classes. In this paper, we introduce a novel active learning method, combining DiscrEpancy-Confusion uncErtainty and calibRatioN diversity for active fine-grained image classification (DECERN), to effectively perceive the distinctiveness between fine-grained images and evaluate the sample value. DECERN introduces a multifaceted informativeness measure that combines discrepancyconfusion uncertainty and calibration diversity. The discrepancy-confusion uncertainty quantifies the structural stability and category directionality of fine-grained unlabeled data during local feature fusion. Subsequently, uncertainty-weighted clustering is performed to diversify the uncertainty samples. Then we calibrate the diversity to maximize the global diversity of the selected sample while maintaining its local representativeness. Extensive experiments conducted on 7 fine-grained image datasets across 39 distinct experimental settings demonstrate that our method achieves superior performance compared to state-of-the-art methods.

The surging demand for adapting long-form cinematic content into short videos has motivated the need for versatile automatic video compilation systems. However, existing compilation methods are limited to predefined tasks, and the community lacks a comprehensive benchmark to evaluate the cinematic compilation. To address this, we introduce CineBench, the first benchmark for instruction-driven cinematic video compilation, featuring diverse user instructions and highquality ground-truth compilations annotated by professional editors. To overcome contextual collapse and temporal fragmentation, we present CineAgents, a multi-agent system that reformulates cinematic video compilation into “design-and-compose” paradigm. CineAgents performs script reverse-engineering to construct a hierarchical narrative memory to provide multi-level context and employs an iterative narrative planning process that refines a creative blueprint into a final compiled script. Extensive experiments demonstrate that CineAgents significantly outperforms existing methods, generating compilations with superior narrative coherence and logical coherence.

Diffusion models typically reconstruct coarse, low-frequency structure before recovering high-frequency detail during the reverse denoising. However, the standard forward process injects Gaussian white noise that corrupts all frequencies uniformly, creating a mismatch between forward corruption and reverse denoising and hindering the recovery of fine detail. We address this mismatch with an adaptive noise covariance schedule that evolves along the geodesic on the Symmetric Positive Definite (SPD) manifold under the Bures-Wasserstein (BW) metric. Specifically, we evolve the covariance along the BW geodesic from a blue-noise endpoint to a white-noise endpoint, so the noise power spectrum transitions smoothly from high-frequency emphasis to uniform. This transition better aligns forward corruption with reverse denoising and enhances textural detail. We provide theoretical and quantitative analyses showing that this transition path yields smoother noise power spectrum evolution and lower noise power fluctuations. In addition, we adapt the transition speed to the characteristics of each image, yielding finer forward-reverse alignment and improved perceptual quality. Experiments on multiple datasets show consistent gains over deterministic diffusion baselines in FID and KID. Our code and pretrained models are available at https://github.com/BolinDeng/RMDM.