ECCV 2026 Papers

Fine-tuning foundation models for Earth Observation is com-putationally expensive, with high training time and memory demands forboth training and deployment. Parameter-efficient methods reduce train-ing cost but retain full inference complexity, while post-hoc compressionoptimizes inference only after costly full fine-tuning. We introduce SIM-PLER, a pre–fine-tuning architecture selection method that reduces in-ference and deployment costs by identifying an effective model depthbefore adaptation. SIMPLER exploits stabilization of representations indeeper layers of pre-trained vision transformers: it computes layer-wiserepresentation similarity on unlabeled task data and applies an auto-mated scoring function to select redundant layers, with no gradients,magnitude heuristics, or hyperparameter tuning required. On Prithvi-EO-2, SIMPLER prunes up to 79% of parameters while retaining 94% ofbaseline performance, yielding a 2.1× training speedup and 2.6× infer-ence speedup. The method generalizes to TerraMind (a multimodal EOfoundation model) and ImageNet-pretrained ViT-MAE, demonstratingapplicability across tasks, architectures, and spectral modalities. Code isavailable at https://gitlab.citius.gal/hpc4rs/simpler.

Most existing knowledge distillation methods focus on ho-mogeneous models (e.g., CNN→CNN), thereby overlooking the flexi-bility and potential of knowledge transfer across heterogeneous mod-els. Due to intrinsic inductive bias discrepancies between heterogeneousmodels that cause spatial distribution inconsistencies, prior heteroge-neous distillation methods often weaken or discard spatial informationin heterogeneous representations. However, the spatial information inrepresentations often encodes transferable global structural semantics aswell as architecture-specific local details, and therefore should not bedirectly ignored. To better leverage the spatial information encoded inheterogeneous representations, we propose a Spatial–Frequency Joint-Aware Heterogeneous Knowledge Distillation framework (SFKD). Byleveraging the complementary properties of wavelet transform spatiallocality and Fourier representations in characterizing global energy dis-tributions, we first apply multi-level discrete wavelet transform to ex-plicitly decouple spatial information. The resulting wavelet sub-bandsare further refined by a dual-stream dual-stage refinement module, andfinally combined with a Gaussian-filtered frequency loss to selectivelycapture informative global information. Extensive experiments on mul-tiple benchmark datasets under both homogeneous and heterogeneousmodels demonstrate the superiority of our method. Code is available athttps://github.com/cpcpWang/SFKD.

Recent adapters for vision-language models (VLMs) still suf-fer from dense cross-modal interactions and limited control over align-ment capacity. To address this issue, we propose a two-stage Shallow-Deep Squeezing Adapter (SDSA) that explicitly regulates cross-modalinteraction density and alignment capacity. In the shallow squeezingstage, SDSA leverages token-level masking to impose structured spar-sity within a sparse bottleneck space. Then, in the deep squeezing stage,SDSA applies a shared low-rank transformation and a cross-modal atten-tion module: the low-rank module consolidates alignment into a compactset of dominant shared directions, while cross-modal attention refinesrepresentations through selective interaction. Extensive experiments on11 datasets show that SDSA delivers superior base-to-novel generaliza-tion and cross-dataset evaluation, and significantly improves the gener-alizability of VLMs under few-shot conditions. The code can be found athttps://github.com/haowang-ac/SDSA.

We present Scale3D, an autoregressive model for generatinghigh-quality outdoor 3D scenes at scale. Scaling 3D generation to vastenvironments like city blocks usually degrades local geometric detailsas the layout expands. Scale3D decouples this problem into a two-stageprocess. A causal autoregressive model _x001C_rst plans the global layout bypredicting scene chunks over a 2D spatial grid. A masked autoregressivemodel then synthesizes the actual 3D geometry within each chunk. Wereconstruct these local structures using a 3D chunk-based VAE designedto preserve sharp edges and high-_x001C_delity surfaces. We also introduce apipeline that extracts spatial relationships directly from geometry, al-lowing users to control complex scene layouts via simple text prompts.Unlike patch-based di_x001B_usion models that struggle with boundary arti-facts, Scale3D maintains long-range spatial coherence across expansiveoutdoor datasets while rendering detailed 3D structures.

Referring Multi-Object Tracking has attracted increasing at-tention due to its human-friendly interactive characteristics, yet it ex-hibits limitations in low-visibility conditions, such as nighttime, smoke,and other challenging scenarios. To overcome this limitation, we proposea new RGB-Thermal RMOT task, named RT-RMOT, which aimsto fuse RGB appearance features with the illumination robustness ofthe thermal modality to enable all-day referring multi-object tracking.To promote research on RT-RMOT, we construct the first Ref erringMulti-Object Tracking dataset under RGB-Thermal modality, namedRefRT. It contains 388 language descriptions, 1,250 tracked targets,and 166,147 Language-RGB-Thermal (L-RGB-T) triplets. Furthermore,we propose RTrack, a framework built upon a multimodal large lan-guage model (MLLM) that integrates RGB, thermal, and textual fea-tures. Since the initial framework still leaves room for improvement, weintroduce a Group Sequence Policy Optimization (GSPO) strategy tofurther exploit the model’s potential. To alleviate training instabilityduring RL fine-tuning, we introduce a Clipped Advantage Scaling (CAS)strategy to suppress gradient explosion. In addition, we design StructuredOutput Reward and Comprehensive Detection Reward to balance explo-ration and exploitation, thereby improving the completeness and accu-racy of target perception. Extensive experiments on the RefRT datasetdemonstrate the effectiveness of the proposed RTrack framework.

In DiT-based video generation models equipped with3D Rotary Position Embeddings (3D RoPE), the attention mech-anism remains a primary computational bottleneck due to its quadraticcomplexity with respect to sequence length. While quantized FlashAt-tention o_x001B_ers a promising path toward hardware acceleration, exist-ing low-bit quantization methods overlook two critical challenges in thissetting: 1) applying online rotation matrices_x0016_a widely used techniquefor mitigating outliers in Queries (Q) and Keys (K )_x0016_is di_x001E_cult toreconcile with RoPE; and 2) the non-negative attention matrix P =exp(QK − max(QK)) makes symmetric quantization waste half of the 4-bit dynamic range. In this work, we observe that the outlier distributionsof Q and K are strongly a_x001B_ected by the dimensional partitioning of 3DRoPE. Based on this _x001C_nding, we propose RotateAttention, an e_x001E_-cient mixed-precision INT4 FlashAttention framework tailored forDiT-based video generation models with 3D RoPE, using selec-tive FP16 fallback for accuracy-sensitive attention blocks and denoisingsteps. RotateAttention introduces two core techniques: 1) RoPE-awareRotation, which employs either mergeable rotation matrices that canbe fused into RoPE or negligible-overhead matrices to mitigate RoPE-induced outliers in Q and K ; and 2) Range-optimized P Quantiza-tion, which uses _x001C_xed scales and zero-points to fully exploit the INT4numerical range with minimal computational overhead. Experimentsshow that RotateAttention preserves video generation quality nearlyidentical to full-precision baselines while achieving up to 1.68× end-to-end speedup and 2.2× kernel-level acceleration.

The latent space of diffusion models fundamentally deter-mines their learning efficiency and generation quality. While recent ad-vances in the latent space have driven substantial progress in single-imagegeneration, the optimal latent space for novel view synthesis (NVS) re-mains largely unexplored. In particular, NVS requires geometrically con-sistent generation across viewpoints, but existing approaches typicallyoperate in a view-independent latent space. In this paper, we proposeGeometric Latent Diffusion (GLD), a framework that repurposesthe feature space of a geometric foundation model as the latent space formulti-view diffusion. We show that the features of the geometric founda-tion model not only support high-fidelity RGB reconstruction but alsoencode strong cross-view geometric correspondences, providing a well-suited latent space for NVS. Through experiments, GLD outperformsboth VAE and RAE on 2D image quality and 3D consistency metrics,accelerating training by more than 4.4× compared to the VAE latentspace. Notably, GLD remains competitive with state-of-the-art methodsthat leverage large-scale text-to-image pretraining, despite training itsdiffusion model from scratch without such generative pretraining.

Recent advancements in reinforcement learning, particularlythrough Group Relative Policy Optimization (GRPO), have significantlyimproved multimodal large language models for complex reasoning tasks.However, two critical limitations persist: 1) they often produce unfo-cused, verbose reasoning chains that obscure salient spatiotemporal cues,and 2) binary rewarding fails to account for partially correct answers,resulting in high reward variance and ine!cient learning. In this pa-per, we propose TW-GRPO, a novel framework that enhances visualreasoning with focused thinking and dense reward granularity. Specifi-cally, we employ a token weighting mechanism that prioritizes tokenswith high informational density (estimated by intra-group informationentropy), suppressing redundant tokens like generic reasoning prefixes.Furthermore, we reformulate RL training by shifting from single-choice tomulti-choice QA tasks, where soft rewards enable finer-grained gradientestimation by distinguishing partial correctness. Additionally, we proposequestion-answer inversion, a data augmentation strategy to generatediverse multi-choice samples from existing benchmarks. Experimentsdemonstrate superior performance on several video reasoning and un-derstanding benchmarks. Notably, under identical training settings, ourmethod significantly outperforms the Video-R1, achieving an averageimprovement of 2.6% across four comparable benchmarks, e.g., +3.8%on MVBench, +2.4% on TempCompass, and +2.3% on VideoMME. Ourcodes are available at https://github.com/longmalongma/TW-GRPO.

Diffusion models have seen widespread adoption for text-driven human motion generation and related tasks due to their impres-sive generative capabilities and flexibility. However, current motion dif-fusion models face two major limitations: a representational gap causedby pre-trained text encoders that lack motion-specific information, anderror propagation during the iterative denoising process. This paper in-troduces Reconstruction-Anchored Diffusion Model (RAM) to addressthese challenges. First, RAM leverages a motion latent space as interme-diate supervision for text-to-motion generation. To this end, RAM co-trains a motion reconstruction branch with two key objective functions:self-regularization to enhance the discrimination of the motion space andmotion-centric latent alignment to enable accurate mapping from textto the motion latent space. Second, we propose Reconstructive ErrorGuidance (REG), a testing-stage guidance mechanism that exploits themotion diffusion model’s inherent self-correction ability to mitigate errorpropagation. At each denoising step, REG uses the motion reconstruc-tion branch to reconstruct the previous estimate, reproducing the priorerror patterns. By amplifying the residual between the current predic-tion and the reconstructed estimate, REG highlights the improvementsin the current prediction. Extensive experiments demonstrate that RAMachieves significant improvements and state-of-the-art performance. Ourcode will be released on https://feifeifeiliu.github.io/RAM .

We present a real-time LiDAR-based framework for Gaus-sian Splatting SLAM that tightly couples fast G-ICP registration withspherical rasterization-based dense mapping for large-scale sequences.Leveraging LiDAR geometry rather than appearance, we reuse tracking-estimated local covariances to initialize Gaussians with range-aware scalesand to derive surface normals for geometry-aware map optimization. Wefurther introduce a covariance-derived geometry score that measures lo-cal complexity and drives pruning in planar regions and selective densifi-cation in structurally rich areas, while optimized Gaussians and LiDAR-specific confidence cues are fed back to improve tracking robustness. Onthe Newer College dataset, our method achieves an F-score of 86.78%using purely online trajectories at real-time speed (>20 FPS), and addi-tional experiments on other datasets confirm its stability and scalability.Code is available at github.com/Lab-of-AI-and-Robotics/LiDAR-GS-SLAM.Project Page: lab-of-ai-and-robotics.github.io/GS-SLAM-Family

Due to the absence of temporal annotations, recent weaklysupervised temporal action localization (WTAL) methods often adoptpseudo-label learning to boost localization. A segment-level WTAL base-line first generates proposals and constructs pseudo labels, which thensupervise a fully supervised localization head. However, the constructionof pseudo labels depends on the proposals’ original confidence scores.Accumulated noise in proposal generation induces a mismatch betweenthese confidence scores and proposal quality, thereby limiting the achiev-able quality of the pseudo labels. To address this, we propose SCLR, aproposal-level score realignment framework driven by Semantic Com-pleteness Learning (SCL). SCL proceeds in three steps. First, for eachaction class, we construct a core semantic center that focuses on localhigh-response cues and a global teacher center that aggregates class-levelcontextual commonalities. Second, guided by the two centers, we self-supervise the decomposition of the core semantic center into generic andspecific components, then reconstruct the class center by a weighted sum.Third, we derive soft labels from proposal–to–center similarity to super-vise a per-proposal semantic-completeness score, which is then used forscore realignment. Extensive experiments on two public datasets demon-strate that SCLR, when integrated as a plug-and-play framework, consis-tently reduces score misalignment across diverse baselines and improveslocalization accuracy. Further, leveraging re-aligned proposals as pseudolabels to train the localization head yields state-of-the-art performance.

Monocular 3D object detection is crucial for scalable per-ception across fields like autonomous driving, robotics, and surveillance.However, progress is hindered by limited 3D annotations and the in-herent ambiguity of single-image geometry. Existing methods often relyon strong geometric assumptions or carefully curated datasets, whichlimit their applicability to real-world scenarios. In this paper, we presentPLOT (Pseudo-Labeling via Object Tracking), a framework that gen-erates 3D annotations from monocular videos without auxiliary sen-sors or model retraining. PLOT tracks object and background trajecto-ries to estimate camera motion and perform object association in pose-unknown settings. These trajectories provide point correspondences thatalign frame-wise pseudo-LiDARs, which are then fused via simple opti-mization into a unified object shape robust to occlusion and viewpointshifts. Recognizing temporal coherence as a fundamental requirementfor reliable shape fusion and video perception, we design a global ob-ject memory that preserves consistent object identities across frames.PLOT achieves robust annotation quality and strong generalization onboth M3OD video benchmarks and in-the-wild videos, proving its effec-tiveness across diverse and unconstrained domains. Project page: https://plot-eccv.github.io.

Personalizing Multimodal Large Language Models (MLLMs)aims to recognize users’ unique concepts from visual data and providepersonalized responses. Although prior work has shown the benefit ofconcept descriptions and reasoning for this task, MLLM descriptions ofteninclude information, such as state and context, that does not help andmay in fact hinder the unique identification of the target concept amongother visually similar items. Effective descriptions of personal conceptsshould instead be accurate, discriminative, and free of distracting details.To achieve such descriptions, we introduce Reinforced Reference Game(RRG), a learning framework that promotes discriminative descriptionsthrough a novel reinforced multimodal reference game. The MLLM playsboth the roles of speaker and listener in a contrastive game setting, whosegoal is to effectively communicate discriminative information about atarget concept. Our approach formulates a verifiable contrastive rewardover hard positives (dissimilar views of the same concept) and hardnegatives (visually similar but different concepts). Empirically, RRGachieves state-of-the-art across multiple tasks on three personalizationbenchmarks. RRG generalizes to unseen domains and outperforms existingmethods based on concept descriptions and personalization-specific RLframeworks. We will release code and models in the project page.

Traditional in-camera colorimetric mapping relies on corre-lated color temperature (CCT)–based interpolation between pre-calibratedtransforms optimized for Planckian illuminants such as CIE A and D65.However, modern lighting technologies such as LEDs can deviate sub-stantially from the Planckian locus, exposing the limitations of rely-ing on conventional one-dimensional CCT for illumination characteriza-tion. This paper demonstrates that transitioning from 1D CCT (on thePlanckian locus) to a 2D chromaticity space (off the Planckian locus) im-proves colorimetric accuracy across various mapping approaches. In addi-tion, we replace conventional CCT interpolation with a lightweight multi-layer perceptron (MLP) that leverages 2D chromaticity features for ro-bust colorimetric mapping under non-Planckian illuminants. A lightbox-based calibration procedure incorporating representative LED sourcesis used to train our MLP. Validated across diverse LED lighting, ourmethod reduces angular reproduction error by 22% on average in LED-lit scenes, maintains backward compatibility with traditional illuminants,accommodates multi-illuminant scenes, and supports real-time in-cameradeployment with negligible additional computational cost. Code and datacan be found on the project webpage: ccmmlp.github.io

Recent Video Super-Resolution (VSR) methods rely heavilyon transformers and explicit optical flow, creating computational over-head and custom operations that hinder deployment on hardware acceler-ators like TensorRT. To address this, we introduce NanoVSR, a scalable,fully convolutional architecture designed for resource-constrained edgedevices. Using structural reparameterization, NanoVSR collapses intostandard convolutions during inference, ensuring seamless hardware com-patibility and negligible runtime overhead. Furthermore, despite lackingexplicit motion compensation, it maintains competitive restoration qual-ity by implicitly learning spatio-temporal alignments through progres-sive training. Evaluated on the REDS4 benchmark, NanoVSR demon-strates an exceptional balance between accuracy and computational ef-ficiency, significantly improving the trade-off for compact architectures.Our NanoVSR-644k baseline yields 28.64 dB PSNR on REDS4 whiledelivering 27.20 FPS on the NVIDIA Jetson Orin NX 16GB (25W),offering massive speed gains over heavier models. The scaled NanoVSR-1.7M variant reaches 29.15 dB with a throughput of 19.58 FPS, providingsuperior, edge-optimized upscaling.

End-to-end autonomous driving systems commonly follow a cascaded two-stage pipeline where a perception stage compresses multimodal sensor inputs into a compact context and a downstream planner predicts trajectories conditioned on this context. We argue that this oneway perception-to-planning interface forces sensor inputs into a compact representation, losing the fine-grained details critical for planning. Moreover, by constraining the planner to this compressed context, it is difficult to leverage the rich representations offered by modern vision foundation models. To address these issues, we propose MOJITO, a unified sensorto-action framework for end-to-end autonomous driving built on modal joint learning. MOJITO removes the cascaded interface and instead performs block-wise Modal Joint Attention that simultaneously updates action, image, and LiDAR features, allowing the planner to directly access multi-modal features during action generation. MOJITO achieves 88.9 PDMS on the NAVSIM v1 dataset and 88.4 EPDMS on the more challenging NAVSIM v2 dataset, setting a new state-of-the-art. Extensive experiments further demonstrate strong scalability, instruction following, and diverse trajectory generation. Code and models are available at https://github.com/mumucc01/MOJITO.

Video diffusion models achieve strong frame-level fidelity butstill struggle with motion coherence, dynamics and realism, often pro-ducing jitter, ghosting, or implausible dynamics. A key limitation is thatthe standard denoising MSE objective provides no direct supervision ontemporal consistency, allowing models to achieve low loss while still gen-erating poor motion. We propose MoGAN, a motion-centric post-trainingframework that improves motion realism without reward models or hu-man preference data. Built atop a 3-step distilled video diffusion model,we train a DiT-based optical-flow discriminator to differentiate real fromgenerated motion, combined with a distribution-matching regularizer topreserve visual fidelity. With experiments on Wan2.1-T2V-1.3B, MoGANsubstantially improves motion quality across benchmarks. On VBench,MoGAN boosts motion score by +7.3% over the 50-step teacher and+13.3% over the 3-step DMD model. On VideoJAM-Bench, MoGAN im-proves motion score by +7.4% over the teacher and +8.8% over DMD,while maintaining comparable or even better aesthetic and image-qualityscores. A human study further confirms that MoGAN is preferred formotion quality (52% vs. 38% for the teacher; 56% vs. 29% for DMD).Overall, MoGAN delivers significantly more realistic motion without sac-rificing visual fidelity or efficiency, offering a practical path toward fast,high-quality video generation.

Multimodal large language models (MLLMs) have advancedrapidly, yet heterogeneity in architecture, alignment strategies, and effi-ciency means that no single model is uniformly superior across tasks. Inpractical deployments, workloads span lightweight OCR to complex mul-timodal reasoning; using one MLLM for all queries either over-provisionscompute on easy instances or sacrifices accuracy on hard ones. Query-level model selection (routing) addresses this tension, but extending rout-ing from text-only LLMs to MLLMs is nontrivial due to modality fusion,wide variation in computational cost across models, and the absence ofa standardized, budget-aware evaluation.We present MMR-Bench, a unified benchmark that isolates the mul-timodal routing problem and enables comparison under fixed candidatesets and cost models. MMR-Bench provides (i) a controlled environmentwith modality-aware inputs and variable compute budgets, (ii) a broadsuite of vision–language tasks covering OCR, general VQA, and mul-timodal reasoning, and (iii) strong single-model baselines, oracle upperbounds, and representative routing policies. Using MMR-Bench, we showthat incorporating multimodal signals improves routing quality. Empiri-cally, these cues improve the cost–accuracy frontier and enable the routedsystem to exceed the strongest single model’s accuracy at roughly 33\%of its cost. Furthermore, policies trained on a subset of models and tasksgeneralize zero-shot to new datasets and text-only benchmarks withoutretuning, establishing MMR-Bench as a foundation for studying adaptivemultimodal model selection and efficient MLLM deployment.

Masked autoregressive (MAR) models unify the strengths ofmasked and autoregressive generation by predicting tokens in a fixed or-der using bidirectional attention for image generation. Although effective,MAR models incur substantial computational overhead because they re-compute attention and feed-forward representations for every token ateach decoding step, even though the majority of tokens remain seman-tically unchanged across steps. We propose a training-free generationframework MARché to address this inefficiency through two key com-ponents: cache-aware attention and selective KV refresh. Cache-awareattention partitions tokens into active and cached sets, enabling sepa-rate computation paths that allow efficient reuse of previously computedkey/value projections without compromising full-context modeling. How-ever, a cached token cannot be used indefinitely without recomputationdue to the changing contextual information over multiple steps. MARchérecognizes this challenge and applies a technique called selective KV re-fresh. Selective KV refresh identifies contextually relevant tokens basedon attention scores from newly generated tokens and updates only thosetokens that require recomputation, while preserving image generationquality. MARché significantly reduces redundant computation in MARwithout modifying the underlying architecture. Empirically, MARchéachieves up to 1.7× speedup with negligible impact on image quality,offering a scalable and broadly applicable solution for efficient maskedtransformer generation.