arXiv subjects

Zinuo Li

Publications and source records attributed to Zinuo Li.

16 recordsLinked to original sources

arXiv2026

STEER: Structured Event Evidence for Video Reasoning via Multi-Objective Reinforcement Learning

Human understanding of video dynamics relies on forming structured representations of entities, actions, and temporal relations before engaging in abstract reasoning. In contrast, existing Video-LLMs apply unstructured chain-of-thought directly to raw visual tokens, where critical temporal cues are buried in verbose narration and event-level structure is largely overlooked. We propose Structured Event Evidence, which represents a video as a compact, time-ordered event schema capturing salient events with key attributes and inter-event temporal dependencies, enabling evidence-grounded reasoning through a constrained verification process. This design promotes concise, interpretable reasoning while reducing the drift typical of unconstrained chain-of-thought. To train models under this paradigm, we introduce STEER-60K, a dataset with a four-stage progressive pipeline: evidence training, format warm-start, thinking warm-start, and RL post-training. During RL, CoT length and task accuracy often conflict while rewards for hard samples are too sparse, causing the policy to neglect challenging instances. We formulate this as a multi-objective Pareto optimality problem and propose Pareto-Frontier guided Advantage Balancing (P-FAB), which dynamically resolves reward conflicts and identifies balanced optimization directions along the Pareto frontier. The resulting model STEER-4B rivals 7B-scale baselines on video understanding tasks with half the input frames Code and data will be released.

cs.CL↗

arXiv2026

Controllable Complex Human Motion Video Generation via Text-to-Skeleton Cascades

Generating videos of complex human motions such as flips, cartwheels, and martial arts remains challenging for current video diffusion models. Text-only conditioning is temporally ambiguous for fine-grained motion control, while explicit pose-based controls, though effective, require users to provide complete skeleton sequences that are costly to produce for long and dynamic actions. We propose a two-stage cascaded framework that addresses both limitations. First, an autoregressive text-to-skeleton model generates 2D pose sequences from natural language descriptions by predicting each joint conditioned on previously generated poses. This design captures long-range temporal dependencies and inter-joint coordination required for complex motions. Second, a pose-conditioned video diffusion model synthesizes videos from a reference image and the generated skeleton sequence. It employs DINO-ALF (Adaptive Layer Fusion), a multi-level reference encoder that preserves appearance and clothing details under large pose changes and self-occlusions. To address the lack of publicly available datasets for complex human motion video generation, we introduce a Blender-based synthetic dataset containing 2,000 videos with diverse characters performing acrobatic and stunt-like motions. The dataset provides full control over appearance, motion, and environment. It fills an important gap because existing benchmarks significantly under-represent acrobatic motions while web-collected datasets raise copyright and privacy concerns. Experiments on our synthetic dataset and the Motion-X Fitness benchmark show that our text-to-skeleton model outperforms prior methods on FID, R-precision, and motion diversity. Our pose-to-video model also achieves the best results among all compared methods on VBench metrics for temporal consistency, motion smoothness, and subject preservation.

cs.CV↗

arXiv2026

EvoFSM: Controllable Self-Evolution for Deep Research with Finite State Machines

While LLM-based agents have shown promise for deep research, most existing approaches rely on fixed workflows that struggle to adapt to real-world, open-ended queries. Recent work therefore explores self-evolution by allowing agents to rewrite their own code or prompts to improve problem-solving ability, but unconstrained optimization often triggers instability, hallucinations, and instruction drift. We propose EvoFSM, a structured self-evolving framework that achieves both adaptability and control by evolving an explicit Finite State Machine (FSM) instead of relying on free-form rewriting. EvoFSM decouples the optimization space into macroscopic Flow (state-transition logic) and microscopic Skill (state-specific behaviors), enabling targeted improvements under clear behavioral boundaries. Guided by a critic mechanism, EvoFSM refines the FSM through a small set of constrained operations, and further incorporates a self-evolving memory that distills successful trajectories as reusable priors and failure patterns as constraints for future queries. Extensive evaluations on five multi-hop QA benchmarks demonstrate the effectiveness of EvoFSM. In particular, EvoFSM reaches 58.0% accuracy on the DeepSearch benchmark. Additional results on interactive decision-making tasks further validate its generalization.

cs.AI↗

arXiv2025

AdaRD-key: Adaptive Relevance-Diversity Keyframe Sampling for Long-form Video understanding

Understanding long-form videos remains a significant challenge for vision--language models (VLMs) due to their extensive temporal length and high information density. Most current multimodal large language models (MLLMs) rely on uniform sampling, which often overlooks critical moments, leading to incorrect responses to queries. In parallel, many keyframe selection approaches impose rigid temporal spacing: once a frame is chosen, an exclusion window suppresses adjacent timestamps to reduce redundancy. While effective at limiting overlap, this strategy frequently misses short, fine-grained cues near important events. Other methods instead emphasize visual diversity but neglect query relevance. We propose AdaRD-Key, a training-free keyframe sampling module for query-driven long-form video understanding. AdaRD-Key maximizes a unified Relevance--Diversity Max-Volume (RD-MV) objective, combining a query-conditioned relevance score with a log-determinant diversity component to yield informative yet non-redundant frames. To handle broad queries with weak alignment to the video, AdaRD-Key employs a lightweight relevance-aware gating mechanism; when the relevance distribution indicates weak alignment, the method seamlessly shifts into a diversity-only mode, enhancing coverage without additional supervision. Our pipeline is training-free, computationally efficient (running in real time on a single GPU), and compatible with existing VLMs in a plug-and-play manner. Extensive experiments on LongVideoBench and Video-MME demonstrate state-of-the-art performance, particularly on long-form videos. Code available at https://github.com/Xian867/AdaRD-Key.

cs.CV↗

arXiv2025

LatentMove: Towards Complex Human Movement Video Generation

Image-to-video (I2V) generation seeks to produce realistic motion sequences from a single reference image. Although recent methods exhibit strong temporal consistency, they often struggle when dealing with complex, non-repetitive human movements, leading to unnatural deformations. To tackle this issue, we present LatentMove, a DiT-based framework specifically tailored for highly dynamic human animation. Our architecture incorporates a conditional control branch and learnable face/body tokens to preserve consistency as well as fine-grained details across frames. We introduce Complex-Human-Videos (CHV), a dataset featuring diverse, challenging human motions designed to benchmark the robustness of I2V systems. We also introduce two metrics to assess the flow and silhouette consistency of generated videos with their ground truth. Experimental results indicate that LatentMove substantially improves human animation quality--particularly when handling rapid, intricate movements--thereby pushing the boundaries of I2V generation. The code, the CHV dataset, and the evaluation metrics will be available at https://github.com/ --.

cs.CV↗

arXiv2025

Watch and Listen: Understanding Audio-Visual-Speech Moments with Multimodal LLM

Humans naturally understand moments in a video by integrating visual and auditory cues. For example, localizing a scene in the video like "A scientist passionately speaks on wildlife conservation as dramatic orchestral music plays, with the audience nodding and applauding" requires simultaneous processing of visual, audio, and speech signals. However, existing models often struggle to effectively fuse and interpret audio information, limiting their capacity for comprehensive video temporal understanding. To address this, we present TriSense, a triple-modality large language model designed for holistic video temporal understanding through the integration of visual, audio, and speech modalities. Central to TriSense is a Query-Based Connector that adaptively reweights modality contributions based on the input query, enabling robust performance under modality dropout and allowing flexible combinations of available inputs. To support TriSense's multimodal capabilities, we introduce TriSense-2M, a high-quality dataset of over 2 million curated samples generated via an automated pipeline powered by fine-tuned LLMs. TriSense-2M includes long-form videos and diverse modality combinations, facilitating broad generalization. Extensive experiments across multiple benchmarks demonstrate the effectiveness of TriSense and its potential to advance multimodal video analysis. Code and dataset will be publicly released.

cs.CL↗

arXiv2025

Unifying Interpretations of Phase Transitions in the Vicsek Model: Correlation Length as a Diagnostic Tool

Vicsek Model is widely used in simulations of dry active matter. We re-examined two typical phase transitions in the original Vicsek model by using the velocity correlation length. One is the noise-driven disordered-to-ordered phase transition driven by noise, which was initially considered as a second-order transition (continuous transition), but was later demonstrated by Chate's detailed study to be a first-order transition. The other one is the disordered-to-ordered phase transition driven by average distance between particles, which is a second-order transition and satisfies the hyper-scaling relation of continuous transitions. We have discovered the change of correlation length during transition indicates a critical point in continuous transition while not in the discontinuous situation. We have also provided a method to classify phase transitions in active matter systems by using the correlation length and summarized previous work within the same framework. Finally, we end up with a potential application in experiments of bactirial swarms and robotic swarms. We hope our work paves the way for both theory and experiment development of active matter.

cond-mat.soft↗

arXiv2024

Layer-Wise Feature Metric of Semantic-Pixel Matching for Few-Shot Learning

In Few-Shot Learning (FSL), traditional metric-based approaches often rely on global metrics to compute similarity. However, in natural scenes, the spatial arrangement of key instances is often inconsistent across images. This spatial misalignment can result in mismatched semantic pixels, leading to inaccurate similarity measurements. To address this issue, we propose a novel method called the Layer-Wise Features Metric of Semantic-Pixel Matching (LWFM-SPM) to make finer comparisons. Our method enhances model performance through two key modules: (1) the Layer-Wise Embedding (LWE) Module, which refines the cross-correlation of image pairs to generate well-focused feature maps for each layer; (2)the Semantic-Pixel Matching (SPM) Module, which aligns critical pixels based on semantic embeddings using an assignment algorithm. We conducted extensive experiments to evaluate our method on four widely used few-shot classification benchmarks: miniImageNet, tieredImageNet, CUB-200-2011, and CIFAR-FS. The results indicate that LWFM-SPM achieves competitive performance across these benchmarks. Our code will be publicly available on https://github.com/Halo2Tang/Code-for-LWFM-SPM.

cs.CV↗

arXiv2024

Faster Image2Video Generation: A Closer Look at CLIP Image Embedding's Impact on Spatio-Temporal Cross-Attentions

This paper investigates the role of CLIP image embeddings within the Stable Video Diffusion (SVD) framework, focusing on their impact on video generation quality and computational efficiency. Our findings indicate that CLIP embeddings, while crucial for aesthetic quality, do not significantly contribute towards the subject and background consistency of video outputs. Moreover, the computationally expensive cross-attention mechanism can be effectively replaced by a simpler linear layer. This layer is computed only once at the first diffusion inference step, and its output is then cached and reused throughout the inference process, thereby enhancing efficiency while maintaining high-quality outputs. Building on these insights, we introduce the VCUT, a training-free approach optimized for efficiency within the SVD architecture. VCUT eliminates temporal cross-attention and replaces spatial cross-attention with a one-time computed linear layer, significantly reducing computational load. The implementation of VCUT leads to a reduction of up to 322T Multiple-Accumulate Operations (MACs) per video and a decrease in model parameters by up to 50M, achieving a 20% reduction in latency compared to the baseline. Our approach demonstrates that conditioning during the Semantic Binding stage is sufficient, eliminating the need for continuous computation across all inference steps and setting a new standard for efficient video generation.

cs.CV↗

arXiv2024

QEAN: Quaternion-Enhanced Attention Network for Visual Dance Generation

The study of music-generated dance is a novel and challenging Image generation task. It aims to input a piece of music and seed motions, then generate natural dance movements for the subsequent music. Transformer-based methods face challenges in time series prediction tasks related to human movements and music due to their struggle in capturing the nonlinear relationship and temporal aspects. This can lead to issues like joint deformation, role deviation, floating, and inconsistencies in dance movements generated in response to the music. In this paper, we propose a Quaternion-Enhanced Attention Network (QEAN) for visual dance synthesis from a quaternion perspective, which consists of a Spin Position Embedding (SPE) module and a Quaternion Rotary Attention (QRA) module. First, SPE embeds position information into self-attention in a rotational manner, leading to better learning of features of movement sequences and audio sequences, and improved understanding of the connection between music and dance. Second, QRA represents and fuses 3D motion features and audio features in the form of a series of quaternions, enabling the model to better learn the temporal coordination of music and dance under the complex temporal cycle conditions of dance generation. Finally, we conducted experiments on the dataset AIST++, and the results show that our approach achieves better and more robust performance in generating accurate, high-quality dance movements. Our source code and dataset can be available from https://github.com/MarasyZZ/QEAN and https://google.github.io/aistplusplus_dataset respectively.

cs.GR↗

arXiv2023

CLIP Guided Image-perceptive Prompt Learning for Image Enhancement

Image enhancement is a significant research area in the fields of computer vision and image processing. In recent years, many learning-based methods for image enhancement have been developed, where the Look-up-table (LUT) has proven to be an effective tool. In this paper, we delve into the potential of Contrastive Language-Image Pre-Training (CLIP) Guided Prompt Learning, proposing a simple structure called CLIP-LUT for image enhancement. We found that the prior knowledge of CLIP can effectively discern the quality of degraded images, which can provide reliable guidance. To be specific, We initially learn image-perceptive prompts to distinguish between original and target images using CLIP model, in the meanwhile, we introduce a very simple network by incorporating a simple baseline to predict the weights of three different LUT as enhancement network. The obtained prompts are used to steer the enhancement network like a loss function and improve the performance of model. We demonstrate that by simply combining a straightforward method with CLIP, we can obtain satisfactory results.

cs.CV↗

arXiv2023

A Large-scale Film Style Dataset for Learning Multi-frequency Driven Film Enhancement

Film, a classic image style, is culturally significant to the whole photographic industry since it marks the birth of photography. However, film photography is time-consuming and expensive, necessitating a more efficient method for collecting film-style photographs. Numerous datasets that have emerged in the field of image enhancement so far are not film-specific. In order to facilitate film-based image stylization research, we construct FilmSet, a large-scale and high-quality film style dataset. Our dataset includes three different film types and more than 5000 in-the-wild high resolution images. Inspired by the features of FilmSet images, we propose a novel framework called FilmNet based on Laplacian Pyramid for stylizing images across frequency bands and achieving film style outcomes. Experiments reveal that the performance of our model is superior than state-of-the-art techniques. The link of code and data is \url{https://github.com/CXH-Research/FilmNet}.

cs.CV↗

arXiv2023

High-Resolution Document Shadow Removal via A Large-Scale Real-World Dataset and A Frequency-Aware Shadow Erasing Net

Shadows often occur when we capture the documents with casual equipment, which influences the visual quality and readability of the digital copies. Different from the algorithms for natural shadow removal, the algorithms in document shadow removal need to preserve the details of fonts and figures in high-resolution input. Previous works ignore this problem and remove the shadows via approximate attention and small datasets, which might not work in real-world situations. We handle high-resolution document shadow removal directly via a larger-scale real-world dataset and a carefully designed frequency-aware network. As for the dataset, we acquire over 7k couples of high-resolution (2462 x 3699) images of real-world document pairs with various samples under different lighting circumstances, which is 10 times larger than existing datasets. As for the design of the network, we decouple the high-resolution images in the frequency domain, where the low-frequency details and high-frequency boundaries can be effectively learned via the carefully designed network structure. Powered by our network and dataset, the proposed method clearly shows a better performance than previous methods in terms of visual quality and numerical results. The code, models, and dataset are available at: https://github.com/CXH-Research/DocShadow-SD7K

cs.CV↗

arXiv2023

Devignet: High-Resolution Vignetting Removal via a Dual Aggregated Fusion Transformer With Adaptive Channel Expansion

Vignetting commonly occurs as a degradation in images resulting from factors such as lens design, improper lens hood usage, and limitations in camera sensors. This degradation affects image details, color accuracy, and presents challenges in computational photography. Existing vignetting removal algorithms predominantly rely on ideal physics assumptions and hand-crafted parameters, resulting in the ineffective removal of irregular vignetting and suboptimal results. Moreover, the substantial lack of real-world vignetting datasets hinders the objective and comprehensive evaluation of vignetting removal. To address these challenges, we present Vigset, a pioneering dataset for vignetting removal. Vigset includes 983 pairs of both vignetting and vignetting-free high-resolution ($5340\times3697$) real-world images under various conditions. In addition, We introduce DeVigNet, a novel frequency-aware Transformer architecture designed for vignetting removal. Through the Laplacian Pyramid decomposition, we propose the Dual Aggregated Fusion Transformer to handle global features and remove vignetting in the low-frequency domain. Additionally, we propose the Adaptive Channel Expansion Module to enhance details in the high-frequency domain. The experiments demonstrate that the proposed model outperforms existing state-of-the-art methods. The code, models, and dataset are available at \url{https://github.com/CXH-Research/DeVigNet}.

cs.CV↗

arXiv2022

WavEnhancer: Unifying Wavelet and Transformer for Image Enhancement

Image enhancement is a technique that frequently utilized in digital image processing. In recent years, the popularity of learning-based techniques for enhancing the aesthetic performance of photographs has increased. However, the majority of current works do not optimize an image from different frequency domains and typically focus on either pixel-level or global-level enhancements. In this paper, we propose a transformer-based model in the wavelet domain to refine different frequency bands of an image. Our method focuses both on local details and high-level features for enhancement, which can generate superior results. On the basis of comprehensive benchmark evaluations, our method outperforms the state-of-the-art methods.

cs.CV↗

arXiv2022

Fast fluorescence lifetime imaging analysis via extreme learning machine

We present a fast and accurate analytical method for fluorescence lifetime imaging microscopy (FLIM) using the extreme learning machine (ELM). We used extensive metrics to evaluate ELM and existing algorithms. First, we compared these algorithms using synthetic datasets. Results indicate that ELM can obtain higher fidelity, even in low-photon conditions. Afterwards, we used ELM to retrieve lifetime components from human prostate cancer cells loaded with gold nanosensors, showing that ELM also outperforms the iterative fitting and non-fitting algorithms. By comparing ELM with a computational efficient neural network, ELM achieves comparable accuracy with less training and inference time. As there is no back-propagation process for ELM during the training phase, the training speed is much higher than existing neural network approaches. The proposed strategy is promising for edge computing with online training.

physics.bio-ph↗