Video question answering plays a vital role in computer vision, and recent advances in large language models have further propelled the development of this field. However, existing video question answering techniques often face limitations in grasping fine-grained video content in spatial dimensions. It mainly stems from the fixed and low-resolution input of video frames. While some approaches using high-resolution inputs partially alleviate this problem, they introduce excessive computational burdens by encoding the entire high-resolution image. In this work, we propose a granularity-adaptive spatial evidence tokenization model for video question answering. Our method introduces multi-granular visual tokenization in the spatial dimension to produce video tokens at various granularities based on the question. It highlights spatially activated patches at low resolutions through a granularity weighting module and then adaptively encodes these activated patches at high resolution for detail supplementation. To mitigate the computational overhead associated with high-resolution frame encoding, a masking and acceleration module is developed for efficient visual tokenization. Moreover, a granularity compression module is designed to dynamically select and compress visual tokens of varying granularities based on questions. We conduct extensive experiments on 11 mainstream video question answering datasets and the experimental results demonstrate the effectiveness of our proposed method.

Video generation requires modeling a vast spatiotemporal space, which demands significant computational resources and data usage. To reduce the complexity, the prevailing approaches employ a cascaded architecture to avoid direct training with full resolution latent. Despite reducing computational demands, the separate optimization of each sub-stage hinders knowledge sharing and sacrifices flexibility. This work introduces a unified pyramidal flow matching algorithm. It reinterprets the original denoising trajectory as a series of pyramid stages, where only the final stage operates at the full resolution, thereby enabling more efficient video generative modeling. Through our sophisticated design, the flows of different pyramid stages can be interlinked to maintain continuity. Moreover, we craft autoregressive video generation with a temporal pyramid to compress the full-resolution history. The entire framework can be optimized in an end-to-end manner and with a single unified Diffusion Transformer (DiT). Extensive experiments demonstrate that our method supports generating high-quality 5-second (up to 10-second) videos at 768p resolution and 24 FPS within 20.7k A100 GPU training hours. All code and models are open-sourced at https://pyramid-flow.github.io.

In the endeavor to make autonomous robots take actions, task planning is a major challenge that requires translating high-level task descriptions to long-horizon action sequences. Despite recent advances in language model agents, they remain prone to planning errors and limited in their ability to plan ahead. To address these limitations in robotic planning, we advocate a self-refining scheme that iteratively refines a draft plan until an equilibrium is reached. Remarkably, this process can be optimized end-to-end from an analytical perspective without the need to curate additional verifiers or reward models, allowing us to train self-refining planners in a simple supervised learning fashion. Meanwhile, a nested equilibrium sequence modeling procedure is devised for efficient closed-loop planning that incorporates useful feedback from the environment (or an internal world model). Our method is evaluated on the VirtualHome-Env benchmark, showing advanced performance with improved scaling w.r.t. inference-time computation. Code is available at https://github.com/anonymous-icml-2025/equilibrium-planner.

Flow-based generative models have gained popularity for image generation and editing. For instruction-based image editing, it is critical to ensure that modifications are confined to the targeted regions. Yet existing methods often fail to maintain consistency in non-targeted regions between the original / edited images. Our primary contribution is to identify the cause of this limitation as the error accumulation across individual editing steps and to address it by incorporating the historical editing trajectory. Specifically, we formulate image editing as a control problem and leverage the Kalman filter to integrate the historical editing trajectory. Our proposed algorithm, dubbed Kalman-Edit, reuses early-stage details from the historical trajectory to enhance the structural consistency of the editing results. To speed up editing, we introduce a shortcut technique based on approximate vector field velocity estimation. Extensive experiments on several datasets demonstrate its superior performance compared to previous state-of-the-art methods.

Open world object detection aims to identify objects of unseen categories and incrementally recognize them once their annotations are provided. In distinction to the traditional paradigm that is limited to predefined categories, this setting promises a continual and generalizable way of estimating objectness using class-agnostic information. However, achieving such decorrelation between objectness and class information proves challenging. Without explicit consideration, existing methods usually exhibit low recall on unknown objects and can misclassify them into known classes. To address this problem, we exploit three levels of orthogonality in the detection process: First, the objectness and classification heads are disentangled by operating on separate sets of features that are orthogonal to each other in a devised polar coordinate system. Secondly, a prediction decorrelation loss is introduced to guide the detector towards more general and class-independent prediction. Furthermore, we propose a calibration scheme that helps maintain orthogonality throughout the training process to mitigate catastrophic interference and facilitate incremental learning of previously unseen objects. Our method is comprehensively evaluated on open world and incremental object detection benchmarks, demonstrating its effectiveness in detecting both known and unknown objects. Code and models are available at this link.

State-of-the-art personalized text-to-image generation systems are usually trained on a few reference images to learn novel visual representations. However, this is likely to incur infringement of copyright for the reference image owners, when these images are personal and publicly available. Recent progress has been made in protecting these images from unauthorized use by adding protective noises. Yet current protection methods work under the assumption that these protected images are not changed, which is in contradiction to the fact that most public platforms intend to modify user-uploaded content, e.g., image compression. This paper introduces a robust watermarking method, namely InMark, to protect images from unauthorized learning. Inspired by influence functions, the proposed method forges protective watermarks on more important pixels for these reference images from both heuristic and statistical perspectives. In this way, the personal semantics of these images are under protection even if these images are modified to some extent. Extensive experiments demonstrate that the proposed InMark outperforms previous state-of-the-art methods in both protective performance and robustness.

In light of recent advances in multimodal Large Language Models (LLMs), there is increasing attention to scaling them from image-text data to more informative real-world videos. Compared to static images, video poses unique challenges for effective large-scale pre-training due to the modeling of its spatiotemporal dynamics. In this paper, we address such limitations in video-language pre-training with an efficient video decomposition that represents each video as keyframes and temporal motions. These are then adapted to an LLM using well-designed tokenizers that discretize visual and temporal information as a few tokens, thus enabling unified generative pre-training of videos, images, and text. At inference, the generated tokens from the LLM are carefully recovered to the original continuous pixel space to create various video content. Our proposed framework is both capable of comprehending and generating image and video content, as demonstrated by its competitive performance across 13 multimodal benchmarks in image and video understanding and generation. Our code and models are available at https://video-lavit.github.io.

Customizing diffusion models to generate identity-preserving images from user-provided reference images is an intriguing new problem. The prevalent approaches typically require training on extensive domain-specific images to achieve identity preservation, which lacks flexibility across different use cases. To address this issue, we exploit classifier guidance, a training-free technique that steers diffusion models using an existing classifier, for personalized image generation. Our study shows that based on a recent rectified flow framework, the major limitation of vanilla classifier guidance in requiring a special classifier can be resolved with a simple fixed-point solution, allowing flexible personalization with off-the-shelf image discriminators. Moreover, its solving procedure proves to be stable when anchored to a reference flow trajectory, with a convergence guarantee. The derived method is implemented on rectified flow with different off-the-shelf image discriminators, delivering advantageous personalization results for human faces, live subjects, and certain objects. Code is available at https://github.com/feifeiobama/RectifID.

Continual learning aims to learn on non-stationary data streams without catastrophically forgetting previous knowledge. Prevalent replay-based methods address this challenge by rehearsing on a small buffer holding the seen data, for which a delicate sample selection strategy is required. However, existing selection schemes typically seek only to maximize the utility of the ongoing selection, overlooking the interference between successive rounds of selection. Motivated by this, we dissect the interaction of sequential selection steps within a framework built on influence functions. We manage to identify a new class of second-order influences that will gradually amplify incidental bias in the replay buffer and compromise the selection process. To regularize the second-order effects, a novel selection objective is proposed, which also has clear connections to two widely adopted criteria. Furthermore, we present an efficient implementation for optimizing the proposed criterion. Experiments on multiple continual learning benchmarks demonstrate the advantage of our approach over state-of-the-art methods. Code is available at https://github.com/feifeiobama/InfluenceCL.

The human brain rewires itself for neuroplasticity in the presence of new tasks. We are inspired to harness this key process in continual reinforcement learning, prioritizing adaptation to non-stationary environments. In distinction to existing rewiring approaches that rely on pruning or dynamic routing, which may limit network capacity and plasticity, this work presents a novel rewiring scheme by permuting hidden neurons. Specifically, the neuron permutation is parameterized to be end-to-end learnable and can rearrange all available synapses to explore a large span of weight space, thereby promoting adaptivity. In addition, we introduce two main designs to steer the rewiring process in continual reinforcement learning: first, a multi-mode rewiring strategy is proposed which diversifies the policy and encourages exploration when encountering new environments. Secondly, to ensure stability on history tasks, the network is devised to cache each learned wiring while subtly updating its weights, allowing for retrospective recovery of any previous state appropriate for the task. Meanwhile, an alignment mechanism is curated to achieve better plasticity-stability tradeoff by jointly optimizing cached wirings and weights. Our proposed method is comprehensively evaluated on 18 continual reinforcement learning scenarios ranging from locomotion to manipulation, demonstrating its advantages over state-of-the-art competitors in performance-efficiency tradeoffs. Code is available at https://github.com/feifeiobama/RewireNeuron.

The task of lifelong person re-identification aims to match a person across multiple cameras given continuous data streams. Similar to other lifelong learning tasks, it severely suffers from the so-called catastrophic forgetting problem, which refers to the notable performance degradation on previously-seen data after adapting the model to some newly incoming data. To alleviate it, a few existing methods have utilized knowledge distillation to enforce consistency between the original and adapted models. However, the effectiveness of such a strategy can be largely reduced facing the data distribution discrepancy between seen and new data. The hallmark of our work is using adaptively-chosen patches (rather than whole images as in other works) to pilot the forgetting-resistant distillation. Specifically, the technical contributions of our patch-based new solution are two-fold: first, a novel patch sampler is proposed. It is fully differentiable and trained to select a diverse set of image patches that stay crucial and discriminative under streaming data. Secondly, with those patches we curate a novel knowledge distillation framework. Valuable patch-level knowledge within individual patch features and mutual relations is well preserved by the two newly introduced distillation modules, further mitigating catastrophic forgetting. Extensive experiments on twelve person re-identification datasets clearly validate the superiority of our method over state-of-the-art competitors by large performance margins.