ECCV 2026 Papers

Diffusion models (DMs), despite their impressive capabilitiesacross a wide range of generative tasks, have been shown to be vulnera-ble to backdoor attacks. However, existing backdoor methods face criticaltrade-offs among key factors: attack performance, stealthiness, time com-plexity, and required poison rates. For example, achieving high attackperformance typically demands a high poison rate and prolonged train-ing, which undermines stealthiness, making the attack more detectableby backdoor defenses. This paper proposes TooBad (trigger optimizationfor backdoor diffusion models), a backdoor framework which introducesa novel DM-tailored trigger optimization technique to dramatically en-hance the performance of backdoor attacks on DMs. Experiments on rep-resentative benchmarks such as CIFAR-10 show that TooBad can achievehigh ASRs (> 85%) at only 0.5% poison rate, significantly lower thanthe 10% typically required by prior work on the same datasets. At 5%poison rate, TooBad reaches nearly 100% ASR within just 3-5 backdoorinjection epochs1 , whereas existing methods need at least 30-50 epochsat double the poison rate for comparable results. Despite its potency,TooBad easily evades SOTA defenses and maintains high utility. Theseresults reveal a critical threat on DMs and highlight the need for morerobust defenses against such stealthy yet efficient attacks. Our code isavailable at https://github.com/tuanvu171/TooBad.

Text-conditioned diffusion models have advanced image andvideo super-resolution by using prompts as semantic priors, and modernsuper-resolution pipelines typically rely on latent tiling to scale to highresolutions. In practice, a single global caption is used with the latenttiling, often causing prompt misguidance. Specifically, a coarse globalprompt often misses localized details (errors of omission) and provideslocally irrelevant guidance (errors of commission) which leads to substan-dard results at the tile level. To solve this, we propose Tiled Prompts,a unified framework for image and video super-resolution that generatesa tile-specific prompt for each latent tile and performs super-resolutionunder locally text-conditioned posteriors to resolve prompt misguidancewith minimal overhead. Our experiments on high resolution real-worldimages and videos show that tiled prompts bring consistent gains in per-ceptual quality and fidelity, while reducing hallucinations and tile-levelartifacts that can be found in global-prompt baselines. Project Page:https://bryanswkim.github.io/tiled-prompts/.

We propose TacFiLM, a lightweight modality-fusion ap-proach that integrates visual-tactile signals into vision-language-action(VLA) models. While advances in VLAs have introduced robot policiesthat are both generalizable and semantically grounded, these modelsmainly rely on vision-based perception. Vision alone, however, cannotcapture the complex interaction dynamics that occur during contact-richmanipulation, including contact forces, surface friction, compliance, andshear. While recent attempts to integrate tactile signals into VLA modelsoften increase complexity through token concatenation or large-scale pre-training, the heavy computational demands of behaviour models necessi-tate lightweight fusion strategies. To address these challenges, TacFiLMoutlines a post-training finetuning approach that conditions intermedi-ate visual features on pretrained tactile representations using feature-wise linear modulation (FiLM). Experimental results on insertion anddrawer opening tasks demonstrate consistent improvements in successrate, direct task performance, completion time, and force stability acrossboth in-distribution and out-of-distribution tasks. Together, these resultssupport our method as an effective approach to integrating tactile sig-nals into VLA models, improving contact-rich manipulation behaviours.Project page: https://charliem7.github.io/projects/TacFilm/

Stylized video generation requires synthesizing dynamic con-tent that reflects the artistic characteristics of a reference image whilemaintaining semantic control and temporal coherence. Existing fine-tuning-based approaches suffer from two recurring limitations: motion degrada-tion caused by training on static image stylization datasets, which lacktemporal supervision and erode the temporal priors of pre-trained back-bones; and a representation bottleneck from encoder-based style injec-tion, which discards fine-grained stylistic details and bounds stylizationquality by the encoder’s representational capacity. We address both limi-tations through a unified pipeline that covers dataset construction, modelarchitecture, and training strategy. Specifically, we construct MoStyle-5K, a stylized video dataset of 5,000 videos spanning 27 artistic stylesand 4,000 preference pairs, built via a decoupled content-motion syn-thesis pipeline that explicitly preserves temporal priors in the supervi-sion signal. Building on this, we propose Implicit Gated Style To-ken Injection (IGST), which directly concatenates style patches as to-kens to bypass the encoder bottleneck, disables RoPE for style tokensto remove spatial alignment bias, and applies head-wise gated atten-tion to suppress residual structural correlations. The framework is opti-mized with a hybrid image-video objective and refined via Diffusion-DPO post-training using the structured preference pairs in MoStyle-5K. Experiments show that our method achieves superior style fidelityand motion dynamics compared with existing baselines. Project page:https://sixi111.github.io/MoStyle/.

Long-form audio description (AD) requires more than de-scribing visible actions: it must preserve characters, events, relationships,and story context across scenes so that blind and low-vision (BLV) audi-ences can follow a film. Modern video–language models (VLMs) are effec-tive on short clips, but they often treat each moment independently, pro-ducing descriptions that miss who characters are, why events matter, andhow the current scene connects to earlier narrative context. We proposeStoryTeller, a training-free framework for story-aware long-form AD. In-stead of relying only on local visual cues, StoryTeller maintains a verifiednarrative memory that carries forward story-relevant information acrossscenes, enabling later descriptions to remain coherent, grounded, andcontextually informative. Given only raw video and a movie title, Sto-ryTeller can optionally retrieve public movie metadata to resolve namesand story context, while accepting only facts that are supported by thevideo through semantic filtering and VLM verification. The method re-quires no subtitles, scripts, AD transcripts, aligned captions, characterbanks, precomputed face identities, or task-specific fine-tuning. To evalu-ate whether generated AD preserves narrative information, we introduceStoryAD-QA1 , a question-answering benchmark that tests whether alanguage model can answer story-context questions using only the gener-ated descriptions. Experiments on standard AD benchmarks and diverselong-form videos show that StoryTeller consistently improves narrativecoherence, factual grounding, and story comprehension over strong base-lines in automatic, QA-based, and human evaluations.

3D Gaussian Splatting (3DGS) achieves real-time novel-viewsynthesis by optimizing millions of anisotropic Gaussians, yet its trainingremains expensive, with the backward pass dominating runtime in thepost-densification refinement phase. We observe substantial update re-dundancy in this phase: many sampled views have near-plateaued lossesand provide diminishing gradient benefits, but standard training stillruns full backpropagation. We propose SkipGS with a novel view-adaptivebackward gating mechanism for efficient post-densification training. SkipGSalways performs the forward pass to update per-view loss statistics, andselectively skips backward passes when the sampled view’s loss is consis-tent with its recent per-view baseline, while enforcing a minimum back-ward budget for stable optimization. On Mip-NeRF 360, compared to3DGS, SkipGS reduces end-to-end training time by 23.1%, driven by a42.0% reduction in post-densification time, with comparable reconstruc-tion quality. Because it only changes when to backpropagate withoutmodifying the renderer, representation, or loss, SkipGS is plug-and-playand compatible with other complementary efficiency strategies, enablingadditive speedups. Code is available at https://github.com/ASU-ESIC-FAN-Lab/SkipGS.

Text-to-motion generation aims to synthesize realistic andsemantically aligned human motions from natural language descriptions.Recent diffusion-based approaches predominantly adopt Transformer orCNN-based denoisers; However, the former suffers from quadratic com-plexity with respect to sequence length, while the latter is fundamen-tally constrained by limited receptive fields. To bridge this gap, we pro-pose SFM, a state-space-enhanced diffusion framework that introducesa structured and efficient long-range modeling scheme tailored for mo-tion generation. At its core, SFM integrates two novel components into aUNet backbone: a Frequency State-Space Module (F-SSM), which trans-forms motion features into the frequency domain with text-gated multi-scale modulation, and a Spatial State-Space Module (S-SSM), whichcaptures bidirectional spatial dependencies through parallel state tran-sitions. This dual-axis decomposition enables SFM to model both theglobal structure and fine-grained variations of human motion while pre-serving alignment with linguistic prompts. Extensive experiments onHumanML3D and KIT-ML demonstrate that SFM outperforms strongTransformer- and CNN-based baselines in both fidelity and text-motionalignment, achieving superior R-Precision and FID with optimized speed.

Vision-language dataset distillation (VLDD) compresses alarge image-text paired dataset into a small set of synthetic pairs thatcan efficiently train contrastive vision-language models under strict dataand compute budgets. Most existing methods match expert trajectoriesor cross-modal statistics, yet still enforce full-dimensional alignment in aEuclidean embedding space. This is often overly restrictive due to rank-deficient image–text correlation, with shared semantics concentrated in alow-dimensional range and remaining variation spread across a weaklycorrelated residual subspace. LoRS relaxes alignment at the similaritylevel by low-rank factorization, but does not explicitly control domi-nant alignment capacity and structure in the representation space. Wethus propose a rank-aware hyperbolic alignment (RAHA) that combineshierarchical geometry with explicit alignment-capacity control. RAHAlifts multimodal representations to hyperbolic space and optimizes dis-tilled pairs with asymmetric objectives that enforce geodesic alignmentin the shared range while regularizing the residual subspace to preservemodality-private diversity and improve transfer robustness. Experimentson benchmarks show that RAHA demonstrates competitive cross-modalretrieval and improved transfer indicators under fixed budgets.

Manipulating fast and dynamically moving targets in un-structured 3D environments remains challenging for embodied AI. Ex-isting visual-language-action models and world models struggle with ac-curate 3D geometry and physically meaningful forecasting. We proposePhysMani, a framework that couples a physics-principled 3D Gaussianworld model with a future-aware action policy model. The world modellearns a divergence-free Gaussian velocity field via online optimizationfor fast and physically grounded future dynamics prediction. The pol-icy model integrates the predicted 3D scene future dynamics through alearnable token based cross-attention module. We introduce PhysMani-Bench, a dynamic manipulation benchmark with 16 tasks, and demon-strate a superior success rate over strong baselines in both simulationand real-world robot experiments. Our code and data are available athttps://github.com/vLAR-group/PhysMani

Continual learning enables vision systems to adapt to ever-changing data distributions. Despite significant advances, existing ap-proaches fail to capture continuous and concurrent shifts in classes anddomains, a critical capability for real-world deployment. This work in-troduces Online VIL (Online Versatile Incremental Learning),a novel scenario where class concepts and visual domains evolve simul-taneously online without explicit boundaries. To better adapt to thechallenges of such dynamic environments that more closely resemblereal-world conditions, we propose a novel framework TopFlow, Topologypreservation with Flow matching representation that contains two com-plementary mechanisms: Domain-agnostic Flow Matching (DFM)and Global Topology Preservation (GTP). DFM guides the modelto have domain-agnostic representations by integrating the geodesic flowkernel into contrastive learning. In contrast, GTP maintains the globalstructure of the feature space without explicitly storing past examples. Ourextensive experiments demonstrate that TopFlow e!ectively addresses thelimitations of existing methods within the Online VIL scenario, achievingstate-of-the-art performance in challenging Online VIL. The proposedmethods suggest potential directions for building continual learning sys-tems in realistic dynamic environments. Our implementation code isavailable at https://github.com/KU-VGI/Online-VIL.

We study the challenging problem of novel view video syn-thesis from single images or monocular videos. Existing methods, whichoperate under the assumption that pre-trained video models lack nativenovel view synthesis capability and enforce view alignment via cameraconditioning, task-specific fine-tuning, or stepwise hard denoising guid-ance, often suffer from artifacts and compromised global scene consis-tency. In this paper, we introduce NeoMap, a novel training-free frame-work designed to locate high-fidelity, view-consistent novel view solutionsfrom general pre-trained video models. The key to our approach is thecore insight that promising novel view solutions are inherently encodedwithin the natural video data manifold learned by pre-trained models,and the core challenge is simply to locate this optimal solution. We solvethis via our core mechanism: convergent manifold alternating projectioniterations that optimize the initial noise. Extensive experiments demon-strate that NeoMap significantly outperforms all existing methods across3 standard novel view synthesis benchmarks, including the challengingTanks-and-Temples, LLFF and DAVIS datasets, achieving state-of-the-art generation fidelity and top-tier view consistency. Our code and dataare available at https://github.com/vLAR-group/NeoMap

Event cameras provide sequential visual data with spatialsparsity and high temporal resolution, making them attractive for low-latency object detection. Existing asynchronous event-based neural net-works exploit this low-latency advantage by updating predictions eventby event, but still suffer from two bottlenecks: recurrent architecturesare difficult to train efficiently on long sequences, and improving accu-racy often increases per-event computation and latency. Linear atten-tion is appealing because it enables parallel training and recurrent infer-ence. However, its dense state updates make per-event computation scalewith the state size, yielding a poor accuracy-efficiency trade-off for ob-ject detection, where accurate localization requires fine-grained spatialstates. The key challenge is therefore to introduce sparse state activa-tion that exploits the spatial sparsity of events while preserving efficientparallel training. We propose Spatially-Sparse Linear Attention (SSLA),which introduces a mixture-of-spaces state decomposition and a scatter-compute-gather training procedure, enabling state-level sparsity as wellas training parallelism. Building on SSLA, we develop an end-to-endasynchronous linear attention model, SSLA-Det, for low-latency event-based object detection. On Gen1 and N-Caltech101, SSLA-Det achievesstate-of-the-art accuracy among asynchronous methods, reaching 0.375mAP and 0.515 mAP, respectively, while reducing per-event computationby over 20

Prompt-conditioned foundation segmenters rely on explicit spatial prompts, such as points, boxes, or masks, to guide mask decoding. In prompt-absent deployment, however, no external/user prompts or detector-generated prompts are available at test time, creating a mismatch between the expected prompt-conditioned interface and automatic inference. We propose IP-SAM, a prompt-space adaptation framework that restores prompt-conditioned decoding without external prompts. Instead of bypassing SAM2’s prompt interface through direct feature-space adaptation, IP-SAM synthesizes complementary intrinsic foreground/background prompts and routes their dense logits through SAM2’s frozen prompt encoder, translating task-specific spatial cues into the native prompt embedding space. To reduce background leakage under severe camouflage, Prompt-Space Gating uses the intrinsic background prompt as an asymmetric suppressive constraint before decoding. Under a deterministic noexternal-prompt protocol, IP-SAM achieves state-of-the-art performance across four COD benchmarks, including MAE 0.017 on COD10K, with only 21.26M trainable parameters from SPG, PSG, a SAM2-initialized downstream mask decoder, and image-encoder LoRA, while keeping the prompt encoder frozen. Medical polyp experiments in the main paper and supplementary SOD results further show that the same prompt-space adaptation route transfers within foreground–background segmentation.

Text-conditioned human interaction generation must cap-ture both long-range temporal causality within each individual and tightlycoupled coordination between partners. Existing interaction diffusionmodels typically denoise full sequences using bidirectional attention, whichobscures causality and hinders streaming and long-horizon generation.Autoregressive alternatives enforce causality but often suffer from tem-poral drift, leading to coordination degradation and unstable interactiondynamics over time. We propose InterCMDM, a block-causal latent diffu-sion framework for autoregressive two-person interaction generation. In-terCMDM introduces a Dual-Stream Causal Diffusion Transformer thatmaintains separate causal streams for each person while modeling inter-person dependencies via unified dual-stream attention with multi-taskattention masks. These masks unify interaction modeling within a sin-gle attention mechanism and support diverse coordination behaviors, in-cluding simultaneous actions, reactive responses, leader–follower dynam-ics, and independent motion. By training a single model across thesemask configurations as a form of data augmentation, InterCMDM en-ables controllable interaction generation by simply selecting the desiredattention mask at inference time. Finally, a block-wise diffusion objec-tive enables stable latent rollout over long sequences without repeateddecode–encode cycles. InterCMDM achieves state-of-the-art performanceon InterHuman and Inter-X, improving text–motion alignment, realism,and long-horizon continuity.

Learning a generalist control policy for robotic manipula-tion typically relies on large-scale datasets. Given the high cost of real-world data collection, a practical alternative is to generate synthetic datathrough simulation. However, the resulting synthetic data often exhibitsa significant gap from real-world distributions. While many prior studieshave proposed algorithms to bridge the Sim-to-Real discrepancy, thereremains a lack of principled research that grounds these methods in real-world manipulation tasks, particularly their performance on generalistpolicies such as Vision-Language-Action (VLA) models. In this study,we empirically examine the primary determinants of Sim-to-Real gen-eralization across four dimensions: multi-level domain randomization,photorealistic rendering, physics-realistic modeling, and reinforcementlearning updates. To support this study, we design a comprehensive eval-uation protocol to quantify the real-world performance of manipulationtasks. The protocol accounts for key variations in background, light-ing, distractors, object types, and spatial features. Through experimentsinvolving over 10k real-world trials, we derive critical insights into Sim-to-Real transfer. To inform and advance future studies, we release boththe robotic platforms and the evaluation protocol for public access tofacilitate independent verification, thereby establishing a realistic andstandardized benchmark for robotic manipulation policies.

We address the problem of training on long-tailed data forvideo action recognition. We propose to augment the training set usinga text-to-video generative model, conditioned on diverse text promptsgrounded in action profiles and training exemplars. Our approach, calledGen2Balance, converts an imbalanced training set into a balanced com-bination of real and generated video clips. To effectively learn from suchdata, we employ a two-stage training strategy that mitigates domainshift and yields significant improvements.We evaluate on long-tailed versions of standard benchmarks: UCF-101(UCF-LT) and a 100-class subset of Kinetics (K100-LT) selected to pri-oritise temporally challenging actions. Gen2Balance improves accuracyover the strongest baselines for long-tailed learning by 5.1% and 7.0% onthe respective datasets. On rare actions from the RareAct dataset (e.g.,cut keyboard ), Gen2Balance improves accuracy by 31.9%, demonstrat-ing effectiveness for scarce actions. By varying the amount of syntheticdata added, we show that partial balancing already achieves 79% of theperformance gains at 27% of the compute cost on K100-LT, highlightingthe practical scalability of Gen2Balance.

Diffusion-based image editing models have achieved signifi-cant progress in real world applications. However, conventional modelstypically rely on natural language prompts, which often lack the pre-cision required to localize target objects. Consequently, these modelsstruggle to maintain background consistency due to their global im-age regeneration paradigm. Recognizing that visual cues provide an in-tuitive means for users to highlight specific areas of interest, we uti-lize bounding boxes as guidance to explicitly define the editing tar-get. This approach ensures that the diffusion model can accurately lo-calize the target while preserving background consistency. To achievethis, we propose FineEdit, a multi-level bounding box injection methodthat enables the model to utilize spatial conditions more effectively.To support this high precision guidance, we present FineEdit-1.2M,a large scale, fine-grained dataset comprising 1.2 million image edit-ing pairs with precise bounding box annotations. Furthermore, we con-struct a comprehensive benchmark, termed FineEdit-Bench, which in-cludes 1,000 images across 10 subjects to effectively evaluate region basedediting capabilities. Evaluations on FineEdit-Bench demonstrate thatour model significantly outperforms state-of-the-art open-source models(e.g., Qwen-Image-Edit and LongCat-Image-Edit) in instruction compli-ance and background preservation. Further assessments on open bench-marks (GEdit and ImgEdit Bench) confirm its superior generalizationand robustness. The dataset and code will be publicly available throughthe project page at https://laulampaul.github.io/FineEdit.

High-fidelity and expressive controllable human animation isessential for content creation and digital avatar applications. However,existing methods face a dilemma between expressiveness and disentangle-ment. Mainstream 2D pose-conditioned approaches suffer from "motion-shape entanglement", leading to the leakage of the driving subject’s bodyshape. Conversely, methods relying on 3D priors (e.g., SMPL) achievegeometric disentanglement but struggle to capture facial expressions andcomplex gestures, resulting in rigid animations. To this end, we proposeEMOSH, a novel framework for high-fidelity controllable human videogeneration. First, an Expressive Human Model (EHM) is introduced asthe core control representation. By explicitly disentangling shape andpose parameters, we fundamentally resolve the body shape leakage issue.Alongside this, a robust motion tracker is designed to accurately estimateEHM parameters from video. Second, we propose a Coarse-to-Fine Hy-brid Motion Injection strategy, enabling more fine-grained control overexpressions and gestures. Furthermore, we introduce a Spatially-AlignedConditioning mechanism to bridge the domain gap between training and∗ †Intern at WeChat Vision. Corresponding authors.inference, improving identity consistency. Extensive experiments demon-strate that EMOSH outperforms previous methods in both self-drivenand cross-driven scenarios, producing high-fidelity videos with vivid ex-pressions while maintaining shape disentanglement. Video demos andadditional results are available at our Project page.

Anticipating human motion from an egocentric perspectiveis fundamental for proactive assistance in AR/VR, human-robot collab-oration, and embodied AI. While recent works incorporate language asa semantic prior to reduce the ill-posed nature of egocentric forecasting,they largely neglect the 3D spatial and semantic context that governshow motion unfolds, and treat pose and language prediction as separateinference streams. We introduce Ego3DLM, built on two core principles:accurate motion forecasting requires explicit spatial and semantic un-derstanding of the 3D environment, and pose and language must be pre-dicted holistically in a single pass, since motion is inherently tied to thesemantic interpretation of actions being performed. Given three-pointtracking, 3D scene features, and egocentric video, Ego3DLM simultane-ously decodes past pose, future pose, past narration, and future narrationin a single autoregressive pass, grounding predicted poses and descrip-tions in one another to enforce cross-modal and temporal consistency. Weadopt a three-stage training scheme: (1) spatial-semantic scene aware-ness pretraining; (2) holistic instruction tuning over all four outputs ina single pass; and (3) GRPO-based reinforcement finetuning with intra-and inter-modal rewards that directly optimize pose-language fidelity.Experiments on the Nymeria benchmark demonstrate that Ego3DLMachieves state-of-the-art performance across future motion prediction,past motion tracking, and motion description, showing that 3D scenegrounding and holistic cross-modal prediction yield physically plausibleand semantically coherent motion forecasts. The project page is availableat https://jaewoo97.github.io/Ego3DLM/.