ECCV 2026 Papers

Training-free motion customization imposes motion patterns from reference videos onto video generators through test-time computation. Most existing methods target full diffusion models, requiring many denoising steps and high computational cost. With the rise of efficient distilled models, a natural question arises: can test-time motion customization be applied directly to distilled generators with their accelerated sampling and efficiency gains? However, our analysis reveals that existing training-free techniques fail on distilled models. Distillation fundamentally alters the denoising dynamics that prior test-time guidance relies on, and the large denoising steps of distilled generators discard the dense intermediate states that score guidance requires, rendering existing motion control strategies incompatible with fast generation. To address this limitation, we propose MotionEcho, a novel training-free test-time distillation framework that enables motion customization for distilled video generators. The key idea is to correct the student model’s sampling trajectory with restricted usage of a high-quality diffusion teacher at inference time. Teacher supervises the student’s denoising by re-noising the student’s endpoint onto its dense trajectory to form a motion-aligned clean endpoint, then interpolating it with the student’s, while an adaptive scheduling mechanism determines when and how much teacher guidance is needed. As a result, MotionEcho restores generative trajectories for distilled video generators via lightweight, adaptive test-time teacher guidance, enabling accurate motion control without compromising generation efficiency. Extensive experiments on multiple distilled video generation models demonstrate that our method significantly improves motion fidelity and visual quality while retaining the efficiency advantages of distilled generation. Project page: MotionEcho.

Existing multimodal visual recognition methods addressing missing modalities have achieved great progress, mainly based on the feature reconstruction framework. However, feature reconstruction often introduces distribution inconsistency between synthesized and real features. Considering each modality as a different view of the same instance, each view is noisy and incomplete, but important representations, such as semantic details, tend to be shared between all modalities. Thus, we propose that AlignPrompt alleviates the distribution inconsistency problem by completing and aligning the semantic representation between synthesized and non-missing prompts. AlignPrompt is an instance-level prompt, implemented via a modal-shared-specific structure including a modalshared factor and a modal-specific factor. Specifically, for each modality pair, AlignPrompt aligns modality-specific factors of all modalities in a shared space by the modal-shared factor as an anchor, and completes the semantic representation by maximizing the mutual information between the synthesized factor and the non-missing factor. Besides, we also propose HyperPrompt to globally regulate the association of prompts of each layer based on different modalities missing scenarios: it transfers global prompts across layers by the inter-layer modules generated from an independent network. This collaborative prompts framework achieves state-of-the-art performance on multiple multimodal classification datasets while maintaining parameter efficiency.

Camera pose estimation is a key step in 3D reconstructionand view-synthesis pipelines. We present a deep, global Structure-from-Motion framework based on learned view-graph aggregation. Our methodemploys a permutation-equivariant, edge-conditioned graph neural net-work that takes noisy pairwise relative poses as input and outputs glob-ally consistent camera extrinsics. The network is trained without ground-truth supervision, relying solely on a relative-pose consistency objective.This is followed by 3D point triangulation and robust bundle adjust-ment. Our approach is efficient, scalable to more than a thousand images,and robust to graph density. We evaluate our method on MegaDepth,1DSfM, Strecha, and BlendedMVS. These experiments demonstrate thatour method achieves superior rotation and translation accuracy com-pared to deep track-centric methods while registering more images acrossmany scenes, and competitive results compared to state-of-the-art clas-sical pipelines, while being much faster.

Perception-centric systems are typically implemented with a modular encoderdecoder pipeline: a vision backbone for feature extraction and a separate decoder (or late-fusion module) for task prediction. We ask whether this separation is necessary for tasks such as openvocabulary segmentation and OCR. We present Falcon Perception, a single, early-fusion dense Transformer that processes image patches, text, and task tokens in a single shared parameter space. Falcon Perception uses a hybrid attention mask: bidirectional for image tokens and causal for text/task tokens, and keeps dense outputs practical by emitting only a small number of task tokens per instance and decoding continuous spatial outputs with lightweight heads, enabling parallel high-resolution mask prediction. On SA-Co, Falcon Perception achieves 68.0 F , compared to 62.3 for a strong state-of-the-art promptable segmentation baseline. 1 To measure compositional grounding beyond saturated referring benchmarks, we introduce PBench and show large gains on spatial and dense long-context regimes (e.g., 53.5 vs. 31.6 on spatial and 72.6 vs. 58.4 on Dense). Finally, we extend the same architecture to document OCR: a compact Falcon-OCR model with 300M parameters achieves 80.3% on olmOCR and 88.64% on OmniDocBench. Code: GitHub, PBench: HF Dataset, Blog: Project Blog.

Long-range 3D object detection is critical for safe autonomous driving at highway speeds, yet existing radar-camera fusion methods face notable limitations at extended ranges. BEV-based approaches effectively encode scene-level context but incur rapidly growing computational cost and struggle to preserve fine-grained object-level detail, while querybased methods provide efficient object-centric encoding but lack sufficient scene-level context. Temporal fusion introduces additional challenges: distant objects produce only a few radar returns and occupy only a few image pixels, requiring scene-level accumulation, while high-speed motion causes large inter-frame displacements that require object-level motion modeling. BEV-based aggregation alleviates sparsity through multi-frame accumulation but is less suited to individual object motion, whereas query-based modeling captures object-level motion but provides limited scene-level temporal context. In this paper, we propose Horizon3D, a sparse radar-camera fusion framework for long-range 3D object detection that jointly captures object-level detail and scene-level context in both spatial and temporal dimensions through a hybrid representation that combines Gaussian primitives with sparse BEV features. Horizon3D first employs Keypoint-Guided Gaussian Initialization (KGGI) to initialize Gaussian primitives at object keypoints estimated from radar and camera features. Object-Centric Sparse Fusion (OCSF) aggregates cross-modal features around these primitives and splats the refined Gaussians onto the BEV plane, where they are fused with sparse radar BEV features to combine object-level detail with scene-level context. Finally, Dual-Path Temporal Fusion (DPTF) aggregates temporal cues through a BEV path for multi-frame feature accumulation and a Gaussian path for propagating primitives across frames to encode per-object motion. Extensive evaluations on TruckScenes demonstrate that Horizon3D achieves stateof-the-art performance for radar-camera 3D object detection. On the validation set, our approach outperforms the previous best method by +3.0 NDS and +1.6 mAP while maintaining a sparse representation with competitive inference speed.

Single-image-to-3D generative models can now produce highquality geometry, yet conditioning on a single view inevitably introduces ambiguity about unseen regions. Multi-view conditioning can reduce this ambiguity, but existing methods either require fixed canonical viewpoints or rely on external reconstruction modules that impose heavy training costs and limit generation quality. We observe that pretrained singleview models already possess strong 2D-to-3D grounding that can be reused for multi-view conditioning. However, a closer analysis reveals that their conditioning mechanism entangles orientation control with geometry transfer, two functions that conflict when images from different viewpoints are naively combined. Based on this analysis, we propose ROAR-3D, a lightweight method that upgrades a pretrained singleview model to accept an arbitrary number of unposed images. A tokenwise view router assigns each 3D latent token to its most relevant view, implicitly establishing 2D-to-3D correspondences without explicit pose input. A dual-stream attention design preserves the pretrained primaryview behavior while routing auxiliary views through a separate path dedicated to geometric enrichment. An orientation perturbation strategy ensures the auxiliary path learns orientation-independent geometry transfer. These components introduce lightweight architectural additions and incur only low additional inference overhead relative to the singleview baseline. ROAR-3D achieves state-of-the-art multi-view 3D generation quality and supports test-time view scaling from 1 to 12+ views with consistent improvements.

Recent indoor occupancy prediction methods adopt Gaussian primitives as a sparse 3D representation for computational efficiency. However, their training relies on voxel classification, which imposes only local constraints and lacks global supervision on the distribution of the primitives. Therefore, they inevitably predict spurious primitives in empty regions, undermining both representational and computational efficiency. To address this, we propose Feed-forward Likelihood Maximization (FLM), a novel framework that reformulates occupancy prediction as voxel distribution estimation. In FLM, a network is trained to predict a mixture model that maximizes the likelihood over groundtruth occupied voxels in a feed-forward manner. To enable end-to-end training of networks and voxelization of a standard mixture model, we define mixture weights as normalized primitive volumes to implicitly enforce simplex constraints and derive novel voxelization formulas. Based on FLM, our FLM-Occ, a novel method that is capable of relocating randomly initialized primitives over long distances to model a scene. On Occ-ScanNet, FLM-Occ achieves superior accuracy using only 32 superquadrics, 2.7% of the prior SoTA, while running 3.7× faster.

Diffusion models excel in Real-World Image Super-Resolution (Real-ISR) due to their powerful generative priors but suffer from slow iterative sampling. Although existing one-step distillation methods accelerate inference, they typically require auxiliary teacher models that inflate training memory and restrict scalability to large-scale architectures. Furthermore, these fixed-step models lack the flexibility to trade off speed for quality. In this paper, we propose TEASR, a training-efficient anystep diffusion framework for Real-ISR that enables both one-step and multi-step restoration within a unified model. Our key idea is to perform self-adversarial distillation within a single diffusion model, eliminating the need for auxiliary teachers or discriminators. Specifically, we propose a timestep-aware rectification strategy that stabilizes one-step generation across noise levels. These two designs further enables the distillation of 20B-parameter diffusion models on a single GPU, significantly improving training efficiency. Moreover, we introduce a dual-branch diffusion transformer with decoupled timestep condition to separate the current noise state and the denoising target to enhance sampling quality. Extensive experiments demonstrate that TEASR supports seamless any-step sampling and consistently outperforms state-of-the-art methods across multiple datasets. Code: https://github.com/MediaX-SJTU/TEASR.

Multimodal Large Language Models (MLLMs) show promising results as decision-making engines for embodied agents operating in complex, physical environments. However, existing benchmarks often prioritize high-level planning or spatial reasoning, leaving the fine-grained action intelligence required for embodied physical interaction underexplored. To address this gap, we introduce CFG-Bench, a new benchmark designed to systematically evaluate this crucial capability. CFG-Bench consists of 1,368 curated videos paired with 19,562 question-answer pairs spanning three evaluation paradigms targeting four cognitive abilities: 1) Physical Interaction, 2) Temporal-Causal Relation, 3) Intentional Understanding, and 4) Evaluative Judgment. Together, these dimensions provide a systematic framework for assessing a model’s ability to translate visual observations into actionable knowledge, moving beyond mere surface-level recognition. Our comprehensive evaluation on CFG-Bench reveals that leading MLLMs struggle to produce detailed instructions for physical interactions and exhibit profound limitations in the higher-order reasoning of intention and evaluation. Moreover, supervised fine-tuning (SFT) on our data demonstrates that teaching an MLLMs to articulate fine-grained actions directly translates to significant performance gains on established embodied benchmarks. Our analysis highlights these limitations and offers insights for developing more capable and grounded embodied agents. Project page: https://cfg-bench.github.io/

Humans can effortlessly anticipate how objects might moveor change through interaction—imagining a cup being lifted, a knifeslicing, or a lid being closed. We aim to endow computational systemswith a similar ability to predict plausible future object motions directlyfrom passive visual observation. We introduce ObjectForesight, a 3Dobject-centric dynamics model that predicts future 6-DoF poses andtrajectories of rigid objects from short egocentric video sequences. Un-like conventional world/dynamics models that operate in pixel or latentspace, ObjectForesight represents the world explicitly in 3D at the objectlevel, enabling geometrically grounded and temporally coherent predic-tions that capture object affordances and trajectories. To train such amodel at scale, we leverage recent advances in segmentation, mesh re-construction, and 3D pose estimation to curate a dataset of 2+ millionshort clips with pseudo-ground-truth 3D object trajectories. Through ex-tensive experiments, we show that ObjectForesight achieves significantgains in accuracy, geometric consistency, and generalization to unseenobjects and scenes—establishing a scalable framework for learning phys-ically grounded, object-centric dynamics models directly from observa-tion. Project page: https://objectforesight.github.io

Modern Novel View Synthesis (NVS), such as 3D Gaussian Splatting (3DGS), introduces scene representations based on new parametrized primitives. These primitives are optimized through inverserendering optimization from color gradients, possibly with a depth prior. However, they fail to preserve geometric consistency and connectivity of the scene during the training phase. Furthermore, standard computer graphics (CG) pipelines do not easily handle these new primitives, thereby limiting rendering performance and integration into consumer applications. We propose a differentiable framework for Depth Image-Based Rendering (DIBR) using triangle primitives. Our method integrates seamlessly into traditional CG pipelines while remaining fully optimizable within modern NVS frameworks. Moreover, our primitives ensure geometric consistency and connectivity. Colors and depths are jointly optimized following a coarse-to-fine strategy. We compare our renderer to 3DGS, 2DGS, and Triangle Splatting across two forward-facing datasets. Although our approach yields slightly lower NVS quality (up to −3 dB), it achieves a 2× speed-up in rendering time (277 FPS) on a consumer-grade GPU, demonstrating its suitability for real-time or resource-constrained applications. Code: github.com/LISA-VR/ddibr.

Recent advances in text-to-image (T2I) generation have led to models capable of producing highly realistic images. Yet, reliably evaluating their outputs remains challenging, especially at scale. Existing automatic evaluators, often relying on a static prompt set, struggle to capture subtle failure modes such as partial prompt misalignment, compositional errors or visually plausible but semantically incorrect generations. In this work, we introduce DynEval, a Dynamic Evaluation framework designed to jointly assess text-to-image alignment and image quality of T2I models. To support scalable training beyond limited humanannotated data, we construct two large datasets. First, we build GenDB, a collection of 500K prompt-image pairs generated from human-written prompts drawn from DiffusionDB using a tiered prompt-model generation strategy. Second, building upon GenDB, we construct DynEvalInstruct, a 250K instruction dataset comprising prompt-image-response triplets distilled from a structured evaluation pipeline that decomposes evaluation into text-image alignment and visual quality reasoning. Using this dataset, we perform full fine-tuning of a compact evaluator through a curriculum learning strategy to effectively distill the superior evaluation capabilities of a larger teacher vision-language model, resulting in DynEval-2B and DynEval-4B. In extensive comparisons against existing evaluators across 11 benchmarks, our evaluator achieves a higher overall correlation with human judgments. Furthermore, it provides finegrained analysis of the capabilities and failure modes of 36 T2I models across 42 subcategories and 9 semantic dimensions.

We focus on the foundational task of Scene Staging: given a reference scene image and a text condition specifying an actor category to be generated in the scene and its spatial relation to the scene, the goal is to synthesize an output image that preserves the same scene identity as the reference image while correctly generating the actor according to the spatial relation described in the text. Existing methods struggle with this task, largely due to the scarcity of high-quality paired data and unconstrained generation objectives. To overcome the data bottleneck, we propose a novel data construction pipeline that combines real-world photographs, entity removal, and image-to-video diffusion models to generate training pairs with diverse scenes, viewpoints and correct entity-scene relationships. Furthermore, we introduce a novel correspondence-guided attention loss that leverages cross-view cues to enforce spatial alignment with the reference scene. Experiments on our scene-consistent benchmark show that our approach achieves better scene alignment and text–image alignment than state-of-the-art baselines, according to both automatic metrics and human preference studies. Our method generates images with diverse viewpoints and compositions while faithfully following the textual instructions and preserving the reference scene identity.

Traditional image stitching methods estimate warps from hand-crafted geometric features, whereas recent learning-based solutions leverage semantic features from neural networks instead. These two lines of research have largely diverged along separate evolution, with virtually no meaningful convergence to date. In this paper, we take a pioneering step to bridge this gap by unifying semantic and geometric features with UniStitch, a unified image stitching framework from multimodal features. To align discrete geometric features (i.e., keypoint) with continuous semantic feature maps, we present a Neural Point Transformer (NPT) module, which transforms unordered, sparse 1D geometric keypoints into ordered, dense 2D semantic maps. Then, to integrate the advantages of both representations, an Adaptive Mixture of Experts (AMoE) module is designed to fuse geometric and semantic representations. It dynamically shifts focus toward more reliable features during the fusion process, allowing the model to handle complex scenes, especially when either modality might be compromised. The fused representation can be adopted into common deep stitching pipelines, delivering significant performance gains over any single feature. Experiments show that UniStitch outperforms existing state-of-the-art methods with a large margin, paving the way for a unified paradigm between traditional and learning-based image stitching. Our project page is available at https://mmelodyy.github.io/projects/unistitch/.

When LiDAR scans a scene with reflective materials, it of-ten produces erroneous virtual points, which resemble real structuresand degrade downstream 3D perception. Conventional methods rely onmulti-stage pipelines that combine 3D surface reconstruction with heuris-tic symmetry detection, making them highly sensitive to surface noise,reflection irregularities, and accumulated estimation errors across stages.To address this, we propose an end-to-end framework that reformu-lates virtual point removal as ray-aligned geometric reasoning in the 2Dequirectangular projection domain, eliminating the dependency on ex-plicit surface modeling. Specifically, we introduce a Layer-Wise NearestPoint Map, which organizes multi-return LiDAR points by layer corre-spondence rather than fixed distance intervals, preserving geometric con-tinuity between real and reflected surfaces. Building on this representa-tion, we propose a Ray-Path-Aware Attention mechanism that constrainsfeature aggregation along physically plausible reflection trajectories de-rived from surface normals and incident ray directions. This physicallyguided attention enables the network to learn reflection-induced symme-try between real and reflected regions without stage-wise surface estima-tion. Extensive experiments on real-world LiDAR datasets demonstratethat our method achieves more accurate and stable virtual point removalthan prior symmetry-based approaches.

Vision-language models (VLMs) have achieved remarkable generalization across diverse multimodal tasks through large-scale pretraining, yet their rapidly increasing computational and memory requirements pose significant challenges for deployment in constrained environments. Existing pruning strategies often depend on task-specific criteria or LLM-oriented importance measures, making them unsuitable for task-agnostic pruning, where no task-specific samples are available at pruning time and the pruned model remains broadly applicable. We introduce a retraining-free VLM pruning framework called PORTA that derives a taskand modality-agnostic importance formulation based on activation variation, estimated from generic calibration data, which reliably captures feature-level representation utility across modalities. PORTA further incorporates an adaptive sparsity allocation mechanism that assigns layer-wise pruning ratios based on output feature variability, avoiding the limitations of uniform sparsity and reducing performance degradation at high compression levels. Extensive experiments across VLM architectures, such as CLIP, BLIP, and Qwen2-VL, demonstrate that PORTA achieves competitive downstream performance under high sparsity without requiring any retraining, supporting efficient VLM compression. Code is available at https://github.com/cau-hailab/PORTA.git.

Editing portrait videos is a challenging task that requires flexible yet precise control over a wide range of modifications, such as appearance changes, expression edits, or the addition of objects. The key difficulty lies in preserving the subject’s original temporal behavior, demanding that every edited frame remains precisely synchronized with the corresponding source frame. We present Sync-LoRA, a method for editing portrait videos that achieves high-quality visual modifications while maintaining frame-accurate synchronization and identity consistency. Our approach uses an image-to-video diffusion model, where the edit is defined by modifying the first frame and then propagated to the entire sequence. To enable accurate synchronization, we train an in-context LoRA using paired videos that depict identical motion trajectories but differ in appearance. These pairs are automatically generated and curated through a synchronization-based filtering process that selects only the most temporally aligned examples for training. This training setup teaches the model to combine motion cues from the source video with the visual changes introduced in the edited first frame. Trained on a compact, highly curated set of synchronized human portraits, SyncLoRA generalizes to unseen identities and diverse edits (e.g., modifying appearance, adding objects, or changing backgrounds), robustly handling variations in pose and expression. Our results demonstrate high visual fidelity and strong temporal coherence, achieving a robust balance between edit fidelity and precise motion preservation. The project page is available at http://sagipolaczek.github.io/Sync-LoRA.

Video fundamentally intertwines two crucial axes: the dynamic content of a scene and the camera motion through which it is observed. However, existing generation models often entangle these factors, limiting independent control. In this work, we introduce OmniCamera, a unified framework designed to explicitly disentangle and command these two dimensions. This compositional approach enables flexible video generation by allowing arbitrary pairings of camera and content conditions, unlocking unprecedented creative control. To overcome the fundamental challenges of modality conflict and data scarcity inherent in such a system, we present two key innovations. First, we construct OmniCAM, a novel hybrid dataset combining curated real-world videos with synthetic data that provides diverse paired examples for robust multi-task learning. Second, we propose a Dual-level Curriculum Co-Training strategy that mitigates modality interference and synergistically learns from diverse data sources. This strategy operates on two levels: first, it progressively introduces control modalities by difficulties (condition-level), and second, trains for precise control on synthetic data before adapting to real data for photorealism (data-level). As a result, OmniCamera achieves state-of-the-art performance, enabling flexible control for complex camera movements while maintaining superior visual quality.

Recent zero-shot 3D visual grounding methods leverage vision–language models (VLMs) to localize objects in 3D scenes from natural language queries. However, these methods typically rely on heuristic rules to select which camera views are provided to the VLM, often prioritizing object visibility rather than grounding relevance. We present IVSGround, a framework that learns Influential View Selection for VLM-based 3D visual grounding. Instead of using fixed heuristics, a lightweight view selector is trained to identify views that provide discriminative evidence for grounding. To obtain supervision signals, we generate training signals using feedback from a reasoning VLM through a two-stage rejection sampling process. During inference, the learned selector predicts query-conditioned influential views for each candidate object, which are then evaluated by a frozen reasoning VLM through comparative grounding. Experiments on ScanRefer and NR3D show that IVSGround consistently improves grounding accuracy over existing zeroshot pipelines, demonstrating that selecting where to look is crucial for effective 3D visual grounding.