NeurIPS 2025 Papers

Diffusion models have demonstrated promising performance in real-world video super-resolution (VSR). However, the dozens of sampling steps they require, make inference extremely slow. Sampling acceleration techniques, particularly single-step, provide a potential solution. Nonetheless, achieving one step in VSR remains challenging, due to the high training overhead on video data and stringent fidelity demands. To tackle the above issues, we propose DOVE, an efficient one-step diffusion model for real-world VSR. DOVE is obtained by fine-tuning a pretrained video diffusion model (*i.e.*, CogVideoX). To effectively train DOVE, we introduce the latent-pixel training strategy. The strategy employs a two-stage scheme to gradually adapt the model to the video super-resolution task. Meanwhile, we design a video processing pipeline to construct a high-quality dataset tailored for VSR, termed HQ-VSR. Fine-tuning on this dataset further enhances the restoration capability of DOVE. Extensive experiments show that DOVE exhibits comparable or superior performance to multi-step diffusion-based VSR methods. It also offers outstanding inference efficiency, achieving up to a **28$\times$** speed-up over existing methods such as MGLD-VSR. Code is available at: https://github.com/zhengchen1999/DOVE.

We present StreamBridge, a simple yet effective framework that seamlessly transforms offline Video-LLMs into streaming-capable models. It addresses two fundamental challenges in adapting existing models into online scenarios: (1) limited capability for multi-turn real-time understanding, and (2) lack of proactive response mechanisms. Specifically, StreamBridge incorporates (1) a memory buffer combined with a round-decayed compression strategy, supporting long-context multi-turn interactions, and (2) a decoupled, lightweight activation model that can be effortlessly integrated into existing Video-LLMs, enabling continuous proactive responses. To further support StreamBridge, we construct Stream-IT, a large-scale dataset tailored for streaming video understanding, featuring interleaved video-text sequences and diverse instruction formats. Extensive experiments show that StreamBridge significantly improves the streaming understanding capabilities of offline Video-LLMs across various tasks, outperforming even proprietary models such as GPT-4o and Gemini 1.5 Pro. Simultaneously, it achieves competitive or superior performance on standard video understanding benchmarks.

Latent action models (LAMs) aim to learn action-relevant changes from unlabeled videos by compressing changes between frames as latents. However, differences between video frames can be caused by \textit{controllable changes} as well as exogenous noise, leading to an important concern -- do latents capture the changes caused by actions or irrelevant noise? This paper studies this issue analytically, presenting a linear model that encapsulates the essence of LAM learning, while being tractable. This provides several insights, including connections between LAM and principal component analysis (PCA), desiderata of the data-generating policy, and justification of strategies to encourage learning controllable changes using data augmentation, data cleaning, and auxiliary action-prediction. We also provide illustrative results based on numerical simulation, shedding light on the specific structure of observations, actions, and noise in data that influence LAM learning.

Multimodal Large Language Models (MLLMs) have shown impressive video content understanding capabilities but struggle with fine-grained motion comprehension. To comprehensively assess the motion understanding ability of existing MLLMs, we introduce FAVOR-Bench, which comprises 1,776 videos from both ego-centric and third-person perspectives and enables assessment through both close-ended and open-ended tasks. For close-ended evaluation, we carefully design 8,184 multiple-choice question-answer pairs spanning six distinct sub-tasks. For open-ended evaluation, we employ the GPT-assisted evaluation and develop a novel cost-efficient LLM-free assessment method, where the latter can enhance benchmarking interpretability and accessibility. Comprehensive experiments with 21 state-of-the-art MLLMs reveal significant limitations in their ability to comprehend and describe detailed temporal dynamics in video motions. To alleviate this limitation, we further build FAVOR-Train, a dataset of 17,152 videos with fine-grained motion annotations. Finetuning Qwen2.5-VL on FAVOR-Train yields consistent improvements on motion-related tasks across TVBench, MotionBench and our FAVOR-Bench. Our assessment results demonstrate that the proposed FAVOR-Bench and FAVOR-Train provide valuable tools for the community to develop more powerful video understanding models.

Despite the remarkable generation quality of video Diffusion Transformer (DiT) models, their practical deployment is severely hindered by extensive computational requirements. This inefficiency stems from two key challenges: the quadratic complexity of self-attention with respect to token length and the multi-step nature of diffusion models. To address these limitations, we present Jenga, a novel inference pipeline that combines dynamic attention carving with progressive resolution generation. Our approach leverages two key insights: (1) early denoising steps do not require high-resolution latents, and (2) later steps do not require dense attention. Jenga introduces a block-wise attention mechanism that dynamically selects relevant token interactions using 3D space-filling curves, alongside a progressive resolution strategy that gradually increases latent resolution during generation. Experimental results demonstrate that Jenga achieves substantial speedups across multiple state-of-the-art video diffusion models while maintaining comparable generation quality (8.83$\times$ speedup with 0.01\% performance drop on VBench). As a plug-and-play solution, Jenga enables practical, high-quality video generation on modern hardware by reducing inference time from minutes to seconds---without requiring model retraining.

As both model and dataset sizes continue to scale rapidly, conventional pretraining strategies with fixed compute budgets—such as cosine learning rate schedules—are increasingly inadequate for large-scale training. Recent alternatives, including warmup-stable-decay (WSD) schedules and weight averaging, offer greater flexibility. However, WSD relies on explicit decay phases to track progress, while weight averaging addresses this limitation at the cost of additional memory. In search of a more principled and scalable alternative, we revisit the Schedule-Free (SF) method [Defazio et al., 2024], which has shown strong empirical performance across diverse settings. We show that SF-AdamW effectively navigates the "river" structure of the loss landscape without decay phases or auxiliary averaging, making it particularly suitable for continuously scaling training workloads. To understand this behavior, we conduct a theoretical and empirical analysis of SF dynamics, revealing that it implicitly performs weight averaging without memory overhead. Guided by this analysis, we propose a refined variant of SF that improves robustness to momentum and performs better under large batch sizes, addressing key limitations of the original method. Together, these results establish SF as a practical, scalable, and theoretically grounded approach for language model training.

Distinguishing visually similar objects by their motion remains a critical challenge in computer vision. Although supervised trackers show promise, contemporary self-supervised trackers struggle when visual cues become ambiguous, limiting their scalability and generalization without extensive labeled data. We find that pre-trained video diffusion models inherently learn motion representations suitable for tracking without task-specific training. This ability arises because their denoising process isolates motion in early, high-noise stages, distinct from later appearance refinement. Capitalizing on this discovery, our self-supervised tracker significantly improves performance in distinguishing visually similar objects, an underexplored failure point for existing methods. Our method achieves up to a 6-point improvement over recent self-supervised approaches on established benchmarks and our newly introduced tests focused on tracking visually similar items. Visualizations confirm that these diffusion-derived motion representations enable robust tracking of even identical objects across challenging viewpoint changes and deformations.

$360^\circ$ videos offer a hyper-immersive experience that allows the viewers to explore a dynamic scene from full 360 degrees. To achieve more accessible and personalized content creation in $360^\circ$ video format, we seek to lift standard perspective videos into $360^\circ$ equirectangular videos. To this end, we introduce **Imagine360**, the first perspective-to-$360^\circ$ video generation framework that creates high-quality $360^\circ$ videos with rich and diverse motion patterns from video anchors. Imagine360 learns fine-grained spherical visual and motion patterns from limited $360^\circ$ video data with several key designs. **1)** Firstly we adopt the dual-branch design, including a perspective and a panorama video denoising branch to provide local and global constraints for $360^\circ$ video generation, with motion module and spatial LoRA layers fine-tuned on $360^\circ$ videos. **2)** Additionally, an antipodal mask is devised to capture long-range motion dependencies, enhancing the reversed camera motion between antipodal pixels across hemispheres. **3)** To handle diverse perspective video inputs, we propose rotation-aware designs that adapt to varying video masking due to changing camera poses across frames. **4)** Lastly, we introduce a new 360 video dataset featuring 10K high-quality, trimmed 360 video clips with structured motion to facilitate training. Extensive experiments show Imagine360 achieves superior graphics quality and motion coherence with our curated dataset among state-of-the-art $360^\circ$ video generation methods. We believe Imagine360 holds promise for advancing personalized, immersive $360^\circ$ video creation.

Large Multimodal Models (LMMs) have achieved impressive progress in visual perception and reasoning. However, when confronted with visually ambiguous or non-semantic scene text, they often struggle to accurately spot and understand the content, frequently generating semantically plausible yet visually incorrect answers, which we refer to as semantic hallucination. In this work, we investigate the underlying causes of semantic hallucination and identify a key finding: Transformer layers in LLM with stronger attention focus on scene text regions are less prone to producing semantic hallucinations. Thus, we propose a training-free semantic hallucination mitigation framework comprising two key components: (1) ZoomText, a coarse-to-fine strategy that identifies potential text regions without external detectors; and (2) Grounded Layer Correction, which adaptively leverages the internal representations from layers less prone to hallucination to guide decoding, correcting hallucinated outputs for non-semantic samples while preserving the semantics of meaningful ones. To enable rigorous evaluation, we introduce TextHalu-Bench, a benchmark of 1,740 samples spanning both semantic and non-semantic cases, with manually curated question–answer pairs designed to probe model hallucinations. Extensive experiments demonstrate that our method not only effectively mitigates semantic hallucination but also achieves strong performance on public benchmarks for scene text spotting and understanding.

Recent advancements in low-light video enhancement (LLVE) have increasingly leveraged both RGB and event cameras to improve video quality under challenging conditions. However, existing approaches share two key drawbacks. First, they are tuned for steady low-light scenes, so their performance drops when illumination varies. Second, they assume every sensing modality is always available, while real systems may lose or corrupt one of them. These limitations make the methods brittle in dynamic, real-world settings. In this paper, we propose EVDiffuser, a novel framework for consistent LLVE that integrates RGB and event data through a modality-adaptive diffusion pipeline. By harnessing the powerful priors of video diffusion models, EVDiffuser enables consistent video enhancement and generalization to diverse scenarios under varying illumination, where RGB or events may even be absent. Specifically, we first design a modality-agnostic conditioning mechanism based on a diffusion pipeline by treating the two modalities as optional conditions, which is fine-tuned using augmented and integrated datasets. Furthermore, we introduce a modality-adaptive guidance rescaling that dynamically adjusts the contribution of each modality according to sensor-specific characteristics. Additionally, we establish a benchmark that accounts for varying illumination and diverse real-world scenarios, facilitating future research on consistent event-guided LLVE. Our experiments demonstrate state-of-the-art performance across challenging scenarios (i.e., varying illumination) and sensor-based settings (e.g., event-only, RGB-only), highlighting the generalization of our framework.

Digitizing the physical world into accurate simulation‑ready virtual environments offers significant opportunities in a variety of fields such as augmented and virtual reality, gaming, and robotics. However, current 3D reconstruction and scene-understanding methods commonly fall short in one or more critical aspects, such as geometry completeness, object interactivity, physical plausibility, photorealistic rendering, or realistic physical properties for reliable dynamic simulation. To address these limitations, we introduce HoloScene, a novel interactive 3D reconstruction framework that simultaneously achieves these requirements. HoloScene leverages a comprehensive interactive scene-graph representation, encoding object geometry, appearance, and physical properties alongside hierarchical and inter-object relationships. Reconstruction is formulated as an energy-based optimization problem, integrating observational data, physical constraints, and generative priors into a unified, coherent objective. Optimization is efficiently performed via a hybrid approach combining sampling-based exploration with gradient-based refinement. The resulting digital twins exhibit complete and precise geometry, physical stability, and realistic rendering from novel viewpoints. Evaluations conducted on multiple benchmark datasets demonstrate superior performance, while practical use-cases in interactive gaming and real-time digital-twin manipulation illustrate HoloScene's broad applicability and effectiveness.

Fine-grained video action recognition can be formulated as a video–text matching problem. Previous approaches primarily rely on global video semantics to consolidate video embeddings, often leading to misaligned video–text pairs due to inaccurate atomic-level action understanding. This inaccuracy arises due to i) videos with distinct global semantics may share similar atomic actions or visual appearances, and ii) atomic actions can be momentary, gradual, or not directly aligned with overarching video semantics. Inspired by storyboarding, where a script is segmented into individual shots, we propose a multi-granularity framework, SFAR. SFAR generates fine-grained descriptions of common atomic actions for each global semantic using a large language model. Unlike existing works that refine global semantics with auxiliary video frames, SFAR introduces a filtering metric to ensure correspondence between the descriptions and the global semantics, eliminating the need for direct video involvement and thereby enabling more nuanced recognition of subtle actions. By leveraging both global semantics and fine-grained descriptions, our SFAR effectively identifies prominent frames within videos, thereby improving the accuracy of embedding aggregation. Extensive experiments on various video action recognition datasets demonstrate the competitive performance of our SFAR in supervised, few-shot, and zero-shot settings.

In this work, we address dynamic view synthesis from monocular videos as an inverse problem in a training-free setting. By redesigning the noise initialization phase of a pre-trained video diffusion model, we enable high-fidelity dynamic view synthesis without any weight updates or auxiliary modules. We begin by identifying a fundamental obstacle to deterministic inversion arising from zero-terminal signal-to-noise ratio (SNR) schedules and resolve it by introducing a novel noise representation, termed K-order Recursive Noise Representation. We derive a closed form expression for this representation, enabling precise and efficient alignment between the VAE-encoded and the DDIM inverted latents. To synthesize newly visible regions resulting from camera motion, we introduce Stochastic Latent Modulation, which performs visibility aware sampling over the latent space to complete occluded regions. Comprehensive experiments demonstrate that dynamic view synthesis can be effectively performed through structured latent manipulation in the noise initialization phase.

Diffusion models, though originally designed for generative tasks, have demonstrated impressive self-supervised representation learning capabilities. A particularly intriguing phenomenon in these models is the emergence of unimodal representation dynamics, where the quality of learned features peaks at an intermediate noise level. In this work, we conduct a comprehensive theoretical and empirical investigation of this phenomenon. Leveraging the inherent low-dimensionality structure of image data, we theoretically demonstrate that the unimodal dynamic emerges when the diffusion model successfully captures the underlying data distribution. The unimodality arises from an interplay between denoising strength and class confidence across noise scales. Empirically, we further show that, in classification tasks, the presence of unimodal dynamics reliably reflects the diffusion model’s generalization: it emerges when the model generate novel images and gradually transitions to a monotonically decreasing curve as the model begins to memorize the training data.

Despite recent advances in Vision-Language Models (VLMs), long-video understanding remains a challenging problem. Although state-of-the-art long-context VLMs can process around 1000 input frames, they still struggle to effectively leverage this sequence length, and succumb to irrelevant distractors within the context window. We present Dynamic Context Aggregation, an inference strategy for video question-answering that curates the model's input context. We use the VLM itself to iteratively identify and extract the most relevant frames from the video, which are then used for answering. We demonstrate how leveraging more computation at inference-time to select the most relevant context leads to improvements in accuracy, in agreement with recent work on inference-time scaling of LLMs. Moreover, we achieve state-of-the-art results on 4 diverse video question-answering datasets, showing consistent improvements with 3 different VLMs. In particular, our method shines on longer videos which would not otherwise fit in the model's context window: On longer videos of more than 1 hour on LVBench, our approach using a context window of 32K outperforms the same VLM using standard inference with a 700K context window by 2.8 points.

Perturbation-based explanations are widely utilized to enhance the transparency of machine-learning models in practice. However, their reliability is often compromised by the unknown model behavior under the specific perturbations used. This paper investigates the relationship between uncertainty calibration - the alignment of model confidence with actual accuracy - and perturbation-based explanations. We show that models systematically produce unreliable probability estimates when subjected to explainability-specific perturbations and theoretically prove that this directly undermines global and local explanation quality. To address this, we introduce ReCalX, a novel approach to recalibrate models for improved explanations while preserving their original predictions. Empirical evaluations across diverse models and datasets demonstrate that ReCalX consistently reduces perturbation-specific miscalibration most effectively while enhancing explanation robustness and the identification of globally important input features.

Multimodal Large Language Models (MLLMs) have recently achieved remarkable progress in video understanding. However, their effectiveness in real-time streaming scenarios remains limited due to storage constraints of historical visual features and insufficient real-time spatiotemporal reasoning. To address these challenges, we propose StreamForest, a novel architecture specifically designed for streaming video understanding. Central to StreamForest is the Persistent Event Memory Forest, a memory mechanism that adaptively organizes video frames into multiple event-level tree structures. This process is guided by penalty functions based on temporal distance, content similarity, and merge frequency, enabling efficient long-term memory retention under limited computational resources. To enhance real-time perception, we introduce a Fine-grained Spatiotemporal Window, which captures detailed short-term visual cues to improve current scene perception. Additionally, we present OnlineIT, an instruction-tuning dataset tailored for streaming video tasks. OnlineIT significantly boosts MLLM performance in both real-time perception and future prediction. To evaluate generalization in practical applications, we introduce ODV-Bench, a new benchmark focused on real-time streaming video understanding in autonomous driving scenarios. Experimental results demonstrate that StreamForest achieves the state-of-the-art performance, with accuracies of 77.3% on StreamingBench, 60.5% on OVBench, and 55.6% on OVO-Bench. In particular, even under extreme visual token compression (limited to 1024 tokens), the model retains 96.8% of its average accuracy in eight benchmarks relative to the default setting. These results underscore the robustness, efficiency, and generalizability of StreamForest for streaming video understanding.

Egocentric video reasoning centers on an unobservable agent behind the camera who dynamically shapes the environment, requiring inference of hidden intentions and recognition of fine-grained interactions. This core challenge limits current multimodal large language models (MLLMs), which excel at visible event reasoning but lack embodied, first-person understanding. To bridge this gap, we introduce EgoThinker, a novel framework that endows MLLMs with robust egocentric reasoning capabilities through spatio-temporal chain-of-thought supervision and a two-stage learning curriculum. First, we introduce EgoRe-5M, a large-scale egocentric QA dataset constructed from 13M diverse egocentric video clips. This dataset features multi-minute segments annotated with detailed CoT rationales and dense hand–object grounding. Second, we employ SFT on EgoRe-5M to instill reasoning skills, followed by reinforcement fine-tuning (RFT) to further enhance spatio-temporal localization. Experimental results show that EgoThinker outperforms existing methods across multiple egocentric benchmarks, while achieving substantial improvements in fine-grained spatio-temporal localization tasks.

Recent advances in large multimodal models have significantly advanced video comprehension, yet their performance remains limited in first-person scenarios. The interactive nature of egocentric videos is critical for applications like embodied intelligence, but introduces complex visual contexts that conventional models struggle to capture. To bridge this gap, we introduce OpenMMEgo with innovations across three dimensions: data, model, and training strategy. To provide rich spatiotemporal visual knowledge, we curate a large-scale, high-quality dataset named OME10M, comprising over 8.2M egocentric video QA pairs synthesized from Ego4D series. We also establish OMEBench, a comprehensive benchmark for rigorous egocentric understanding assessment. To alleviate the frequent viewpoint shifts inherent in egocentric videos, we implement semantic-aware visual token compression. Further, a curriculum learning strategy is complemented to foster stable learning across various data complexities. OpenMMEgo consistently improves the performance of LMMs on egocentric benchmarks without sacrificing general video understanding performance. Notably, Qwen2.5-VL tuned with OpenMMEgo substantially outperforms other models of the same size in egocentric video understanding. The data, weights and training code will be put at https://github.com/BeingBeyond/OpenMMEgo.