Shifeng Chen

Haipeng Luo, Qingfeng Sun, Can Xu et al. · microsoft-research

Large language models (LLMs), such as GPT-4, have shown remarkable performance in natural language processing (NLP) tasks, including challenging mathematical reasoning. However, most existing open-source models are only pre-trained on large-scale internet data and without math-related optimization. In this paper, we present WizardMath, which enhances the mathematical CoT reasoning abilities of LLMs without using external python tools, by applying our proposed Reinforcement Learning from Evol-Instruct Feedback (RLEIF) method to the domain of math. Through extensive experiments on two mathematical reasoning benchmarks, namely GSM8k and MATH, we reveal the extraordinary capabilities of our model. Remarkably, WizardMath-Mistral 7B surpasses top-tier open-source LLMs by a substantial margin with higher data efficiency. Furthermore, WizardMath 70B even outperforms GPT-3.5-Turbo, Claude 2, Gemini Pro and GPT-4-early-version. Additionally, our preliminary exploration highlights the pivotal role of instruction evolution and process supervision in achieving exceptional math performance. For more details refer to https://github.com/nlpxucan/WizardLM

Haipeng Luo, Qingfeng Sun, Can Xu et al. · microsoft-research

Assessing the effectiveness of large language models (LLMs) presents substantial challenges. The method of conducting human-annotated battles in an online Chatbot Arena is a highly effective evaluative technique. However, this approach is limited by the costs and time required for human annotation. In this paper, we introduce Arena Learning, an innovative offline strategy designed to simulate these arena battles using AI-driven annotations to evaluate battle outcomes, thus facilitating the continuous improvement of the target model through both supervised fine-tuning and reinforcement learning. Arena Learning comprises two key elements. First, it ensures precise evaluations and maintains consistency between offline simulations and online competitions via WizardArena, a pipeline developed to accurately predict the Elo rankings of various models using a meticulously designed offline test set. Our results demonstrate that WizardArena's predictions closely align with those from the online Arena. Second, it involves the continuous improvement of training data based on the battle results and the refined model. We establish a data flywheel to iteratively update the training data by highlighting the weaknesses of the target model based on its battle results, enabling it to learn from the strengths of multiple different models. We apply Arena Learning to train our target model, WizardLM-$β$, and demonstrate significant performance enhancements across various metrics. This fully automated training and evaluation pipeline sets the stage for continuous advancements in various LLMs via post-training. Notably, Arena Learning plays a pivotal role in the success of WizardLM-2, and this paper serves both as an exploration of its efficacy and a foundational study for future discussions related to WizardLM-2 and its derivatives.

Wanshui Gan, Hongbin Xu, Yi Huang et al.

Neural radiance fields have made a remarkable breakthrough in the novel view synthesis task at the 3D static scene. However, for the 4D circumstance (e.g., dynamic scene), the performance of the existing method is still limited by the capacity of the neural network, typically in a multilayer perceptron network (MLP). In this paper, we utilize 3D Voxel to model the 4D neural radiance field, short as V4D, where the 3D voxel has two formats. The first one is to regularly model the 3D space and then use the sampled local 3D feature with the time index to model the density field and the texture field by a tiny MLP. The second one is in look-up tables (LUTs) format that is for the pixel-level refinement, where the pseudo-surface produced by the volume rendering is utilized as the guidance information to learn a 2D pixel-level refinement mapping. The proposed LUTs-based refinement module achieves the performance gain with little computational cost and could serve as the plug-and-play module in the novel view synthesis task. Moreover, we propose a more effective conditional positional encoding toward the 4D data that achieves performance gain with negligible computational burdens. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance at a low computational cost.

Ningkai Mo, Wanshui Gan, Naoto Yokoya et al.

In this paper, a computation efficient regression framework is presented for estimating the 6D pose of rigid objects from a single RGB-D image, which is applicable to handling symmetric objects. This framework is designed in a simple architecture that efficiently extracts point-wise features from RGB-D data using a fully convolutional network, called XYZNet, and directly regresses the 6D pose without any post refinement. In the case of symmetric object, one object has multiple ground-truth poses, and this one-to-many relationship may lead to estimation ambiguity. In order to solve this ambiguity problem, we design a symmetry-invariant pose distance metric, called average (maximum) grouped primitives distance or A(M)GPD. The proposed A(M)GPD loss can make the regression network converge to the correct state, i.e., all minima in the A(M)GPD loss surface are mapped to the correct poses. Extensive experiments on YCB-Video and T-LESS datasets demonstrate the proposed framework's substantially superior performance in top accuracy and low computational cost.

Jiancheng Huang, Mingfu Yan, Songyan Chen et al.

Amid the surge in generic text-to-video generation, the field of personalized human video generation has witnessed notable advancements, primarily concentrated on single-person scenarios. However, to our knowledge, the domain of two-person interactions, particularly in the context of martial arts combat, remains uncharted. We identify a significant gap: existing models for single-person dancing generation prove insufficient for capturing the subtleties and complexities of two engaged fighters, resulting in challenges such as identity confusion, anomalous limbs, and action mismatches. To address this, we introduce a pioneering new task, Personalized Martial Arts Combat Video Generation. Our approach, MagicFight, is specifically crafted to overcome these hurdles. Given this pioneering task, we face a lack of appropriate datasets. Thus, we generate a bespoke dataset using the game physics engine Unity, meticulously crafting a multitude of 3D characters, martial arts moves, and scenes designed to represent the diversity of combat. MagicFight refines and adapts existing models and strategies to generate high-fidelity two-person combat videos that maintain individual identities and ensure seamless, coherent action sequences, thereby laying the groundwork for future innovations in the realm of interactive video content creation. Website: https://MingfuYAN.github.io/MagicFight/ Dataset: https://huggingface.co/datasets/MingfuYAN/KungFu-Fiesta

Mingfu Yan, Jiancheng Huang, Shifeng Chen

Seal-related tasks in document processing-such as seal segmentation, authenticity verification, seal removal, and text recognition under seals-hold substantial commercial importance. However, progress in these areas has been hindered by the scarcity of labeled document seal datasets, which are essential for supervised learning. To address this limitation, we propose Seal2Real, a novel generative framework designed to synthesize large-scale labeled document seal data. As part of this work, we also present Seal-DB, a comprehensive dataset containing 20,000 labeled images to support seal-related research. Seal2Real introduces a prompt prior learning architecture built upon a pre-trained Stable Diffusion model, effectively transferring its generative capability to the unsupervised domain of seal image synthesis. By producing highly realistic synthetic seal images, Seal2Real significantly enhances the performance of downstream seal-related tasks on real-world data. Experimental evaluations on the Seal-DB dataset demonstrate the effectiveness and practical value of the proposed framework. The dataset is available at https://github.com/liuyifan6613/DocBank-Document-Enhancement-Dataset.

Jiaxi Lv, Liang Zhang, Yi Huang et al.

Recently, more and more research has focused on using Graph Neural Networks (GNN) to solve the Graph Similarity Computation problem (GSC), i.e., computing the Graph Edit Distance (GED) between two graphs. These methods treat GSC as an end-to-end learnable task, and the core of their architecture is the feature fusion modules to interact with the features of two graphs. Existing methods consider that graph-level embedding is difficult to capture the differences in local small structures between two graphs, and thus perform fine-grained feature fusion on node-level embedding can improve the accuracy, but leads to greater time and memory consumption in the training and inference phases. However, this paper proposes a novel graph-level fusion module Different Attention (DiffAtt), and demonstrates that graph-level fusion embeddings can substantially outperform these complex node-level fusion embeddings. We posit that the relative difference structure of the two graphs plays an important role in calculating their GED values. To this end, DiffAtt uses the difference between two graph-level embeddings as an attentional mechanism to capture the graph structural difference of the two graphs. Based on DiffAtt, a new GSC method, named Graph Edit Distance Learning via Different Attention (REDRAFT), is proposed, and experimental results demonstrate that REDRAFT achieves state-of-the-art performance in 23 out of 25 metrics in five benchmark datasets. Especially on MSE, it respectively outperforms the second best by 19.9%, 48.8%, 29.1%, 31.6%, and 2.2%. Moreover, we propose a quantitative test Remaining Subgraph Alignment Test (RESAT) to verify that among all graph-level fusion modules, the fusion embedding generated by DiffAtt can best capture the structural differences between two graphs.

Yufeng Yang, Xianfang Zeng, Zhangqi Jiang et al.

Image restoration under real-world degradations is critical for downstream tasks such as autonomous driving and object detection. However, existing restoration models are often limited by the scale and distribution of their training data, resulting in poor generalization to real-world scenarios. Recently, large-scale image editing models have shown strong generalization ability in restoration tasks, especially for closed-source models like Nano Banana Pro, which can restore images while preserving consistency. Nevertheless, achieving such performance with those large universal models requires substantial data and computational costs. To address this issue, we construct a large-scale dataset covering nine common real-world degradation types and train a state-of-the-art open-source model to narrow the gap with closed-source alternatives. Furthermore, we introduce RealIR-Bench, which contains 464 real-world degraded images and tailored evaluation metrics focusing on degradation removal and consistency preservation. Extensive experiments demonstrate our model ranks first among open-source methods, achieving state-of-the-art performance.

Jisheng Chu, Wenrui Li, Rui Zhao et al.

Generating immersive 3D scenes from texts is a core task in computer vision, crucial for applications in virtual reality and game development. Despite the promise of leveraging 2D diffusion priors, existing methods suffer from spatial blindness and rely on predefined trajectories that fail to exploit the inner relationships among salient objects. Consequently, these approaches are unable to comprehend the semantic layout, preventing them from exploring the scene adaptively to infer occluded content. Moreover, current inpainting models operate in 2D image space, struggling to plausibly fill holes caused by camera motion. To address these limitations, we propose RoamScene3D, a novel framework that bridges the gap between semantic guidance and spatial generation. Our method reasons about the semantic relations among objects and produces consistent and photorealistic scenes. Specifically, we employ a vision-language model (VLM) to construct a scene graph that encodes object relations, guiding the camera to perceive salient object boundaries and plan an adaptive roaming trajectory. Furthermore, to mitigate the limitations of static 2D priors, we introduce a Motion-Injected Inpainting model that is fine-tuned on a synthetic panoramic dataset integrating authentic camera trajectories, making it adaptive to camera motion. Extensive experiments demonstrate that with semantic reasoning and geometric constraints, our method significantly outperforms state-of-the-art approaches in producing consistent and photorealistic scenes. Our code is available at https://github.com/JS-CHU/RoamScene3D.

Jiaxi Lv, Yi Huang, Mingfu Yan et al.

Recent advances in text-to-video generation have harnessed the power of diffusion models to create visually compelling content conditioned on text prompts. However, they usually encounter high computational costs and often struggle to produce videos with coherent physical motions. To tackle these issues, we propose GPT4Motion, a training-free framework that leverages the planning capability of large language models such as GPT, the physical simulation strength of Blender, and the excellent image generation ability of text-to-image diffusion models to enhance the quality of video synthesis. Specifically, GPT4Motion employs GPT-4 to generate a Blender script based on a user textual prompt, which commands Blender's built-in physics engine to craft fundamental scene components that encapsulate coherent physical motions across frames. Then these components are inputted into Stable Diffusion to generate a video aligned with the textual prompt. Experimental results on three basic physical motion scenarios, including rigid object drop and collision, cloth draping and swinging, and liquid flow, demonstrate that GPT4Motion can generate high-quality videos efficiently in maintaining motion coherency and entity consistency. GPT4Motion offers new insights in text-to-video research, enhancing its quality and broadening its horizon for further explorations.

Yi Huang, Jiancheng Huang, Yifan Liu et al.

Denoising diffusion models have emerged as a powerful tool for various image generation and editing tasks, facilitating the synthesis of visual content in an unconditional or input-conditional manner. The core idea behind them is learning to reverse the process of gradually adding noise to images, allowing them to generate high-quality samples from a complex distribution. In this survey, we provide an exhaustive overview of existing methods using diffusion models for image editing, covering both theoretical and practical aspects in the field. We delve into a thorough analysis and categorization of these works from multiple perspectives, including learning strategies, user-input conditions, and the array of specific editing tasks that can be accomplished. In addition, we pay special attention to image inpainting and outpainting, and explore both earlier traditional context-driven and current multimodal conditional methods, offering a comprehensive analysis of their methodologies. To further evaluate the performance of text-guided image editing algorithms, we propose a systematic benchmark, EditEval, featuring an innovative metric, LMM Score. Finally, we address current limitations and envision some potential directions for future research. The accompanying repository is released at https://github.com/SiatMMLab/Awesome-Diffusion-Model-Based-Image-Editing-Methods.

Jiancheng Huang, Yifan Liu, Jin Qin et al.

Text-conditioned image editing is a recently emerged and highly practical task, and its potential is immeasurable. However, most of the concurrent methods are unable to perform action editing, i.e. they can not produce results that conform to the action semantics of the editing prompt and preserve the content of the original image. To solve the problem of action editing, we propose KV Inversion, a method that can achieve satisfactory reconstruction performance and action editing, which can solve two major problems: 1) the edited result can match the corresponding action, and 2) the edited object can retain the texture and identity of the original real image. In addition, our method does not require training the Stable Diffusion model itself, nor does it require scanning a large-scale dataset to perform time-consuming training.

Yufeng Yang, Jianzhuang Liu, Jisheng Chu et al.

Existing deep learning-based low-light enhancement methods are typically trained on limited datasets with single enhancement targets, which restricts their generalization ability and controllability in real-world applications. To overcome these limitations, we propose ControlLight, a controllable, consistent, and generalizable framework for low-light enhancement. We first construct a large-scale dataset of real-world degraded images with continuous illumination-strength supervision. To further ensure consistent outputs under different control strengths, we introduce a misalignment-aware weighted flow matching loss that preserves image structure across continuous enhancement strengths. ControlLight allows users to edit real-world degraded low-light images toward satisfactory enhancement results by flexibly controlling the strength while preserving visual consistency and realism. Extensive experiments show that ControlLight achieves state-of-the-art performance against existing low-light enhancement approaches while demonstrating strong continuous controllability and generalization to real-world scenarios.

Linxiao Shi, Siming Zheng, Zerong Wang et al.

Existing mobile devices are constrained by compact optical designs, such as small apertures, which make it difficult to produce natural, optically realistic bokeh effects. Although recent learning-based methods have shown promising results, they still struggle with photos captured under high digital zoom levels, which often suffer from reduced resolution and loss of fine details. A naive solution is to enhance image quality before applying bokeh rendering, yet this two-stage pipeline reduces efficiency and introduces unnecessary error accumulation. To overcome these limitations, we propose MagicBokeh, a unified diffusion-based framework designed for high-quality and efficient bokeh rendering. Through an alternative training strategy and a focus-aware masked attention mechanism, our method jointly optimizes bokeh rendering and super-resolution, substantially improving both controllability and visual fidelity. Furthermore, we introduce degradation-aware depth module to enable more accurate depth estimation from low-quality inputs. Experimental results demonstrate that MagicBokeh efficiently produces photorealistic bokeh effects, particularly on real-world low-resolution images, paving the way for future advancements in bokeh rendering. Our code and models are available at https://github.com/vivoCameraResearch/MagicBokeh.

Jiancheng Huang, Donghao Zhou, Shifeng Chen

Single-shot face anti-spoofing (FAS) is a key technique for securing face recognition systems, and it requires only static images as input. However, single-shot FAS remains a challenging and under-explored problem due to two main reasons: 1) on the data side, learning FAS from RGB images is largely context-dependent, and single-shot images without additional annotations contain limited semantic information. 2) on the model side, existing single-shot FAS models are infeasible to provide proper evidence for their decisions, and FAS methods based on depth estimation require expensive per-pixel annotations. To address these issues, a large binocular NIR image dataset (BNI-FAS) is constructed and published, which contains more than 300,000 real face and plane attack images, and an Interpretable FAS Transformer (IFAST) is proposed that requires only weak supervision to produce interpretable predictions. Our IFAST can produce pixel-wise disparity maps by the proposed disparity estimation Transformer with Dynamic Matching Attention (DMA) block. Besides, a well-designed confidence map generator is adopted to cooperate with the proposed dual-teacher distillation module to obtain the final discriminant results. The comprehensive experiments show that our IFAST can achieve state-of-the-art results on BNI-FAS, proving the effectiveness of the single-shot FAS based on binocular NIR images.

Jiancheng Huang, Yifan Liu, Shifeng Chen

Learning-based methods have attracted a lot of research attention and led to significant improvements in low-light image enhancement. However, most of them still suffer from two main problems: expensive computational cost in high resolution images and unsatisfactory performance in simultaneous enhancement and denoising. To address these problems, we propose BDCE, a bootstrap diffusion model that exploits the learning of the distribution of the curve parameters instead of the normal-light image itself. Specifically, we adopt the curve estimation method to handle the high-resolution images, where the curve parameters are estimated by our bootstrap diffusion model. In addition, a denoise module is applied in each iteration of curve adjustment to denoise the intermediate enhanced result of each iteration. We evaluate BDCE on commonly used benchmark datasets, and extensive experiments show that it achieves state-of-the-art qualitative and quantitative performance.

Yuqi Peng, Lingtao Zheng, Yufeng Yang et al.

Personalized text-to-image generation aims to synthesize novel images of a specific subject or style using only a few reference images. Recent methods based on Low-Rank Adaptation (LoRA) enable efficient single-concept customization by injecting lightweight, concept-specific adapters into pre-trained diffusion models. However, combining multiple LoRA modules for multi-concept generation often leads to identity missing and visual feature leakage. In this work, we identify two key issues behind these failures: (1) token-wise interference among different LoRA modules, and (2) spatial misalignment between the attention map of a rare token and its corresponding concept-specific region. To address these issues, we propose Token-Aware LoRA (TARA), which introduces a token mask to explicitly constrain each module to focus on its associated rare token to avoid interference, and a training objective that encourages the spatial attention of a rare token to align with its concept region. Our method enables training-free multi-concept composition by directly injecting multiple independently trained TARA modules at inference time. Experimental results demonstrate that TARA enables efficient multi-concept inference and effectively preserving the visual identity of each concept by avoiding mutual interference between LoRA modules. The code and models are available at https://github.com/YuqiPeng77/TARA.

Yuxiang Zhao, Wenhao Wu, Yue He et al.

The detection of traffic anomalies is a critical component of the intelligent city transportation management system. Previous works have proposed a variety of notable insights and taken a step forward in this field, however, dealing with the complex traffic environment remains a challenge. Moreover, the lack of high-quality data and the complexity of the traffic scene, motivate us to study this problem from a hand-crafted perspective. In this paper, we propose a straightforward and efficient framework that includes pre-processing, a dynamic track module, and post-processing. With video stabilization, background modeling, and vehicle detection, the pro-processing phase aims to generate candidate anomalies. The dynamic tracking module seeks and locates the start time of anomalies by utilizing vehicle motion patterns and spatiotemporal status. Finally, we use post-processing to fine-tune the temporal boundary of anomalies. Not surprisingly, our proposed framework was ranked $1^{st}$ in the NVIDIA AI CITY 2021 leaderboard for traffic anomaly detection. The code is available at: https://github.com/Endeavour10020/AICity2021-Anomaly-Detection .

Yiyu Zhuang, Jiaxi Lv, Hao Wen et al.

Creating a high-fidelity, animatable 3D full-body avatar from a single image is a challenging task due to the diverse appearance and poses of humans and the limited availability of high-quality training data. To achieve fast and high-quality human reconstruction, this work rethinks the task from the perspectives of dataset, model, and representation. First, we introduce a large-scale HUman-centric GEnerated dataset, HuGe100K, consisting of 100K diverse, photorealistic sets of human images. Each set contains 24-view frames in specific human poses, generated using a pose-controllable image-to-multi-view model. Next, leveraging the diversity in views, poses, and appearances within HuGe100K, we develop a scalable feed-forward transformer model to predict a 3D human Gaussian representation in a uniform space from a given human image. This model is trained to disentangle human pose, body shape, clothing geometry, and texture. The estimated Gaussians can be animated without post-processing. We conduct comprehensive experiments to validate the effectiveness of the proposed dataset and method. Our model demonstrates the ability to efficiently reconstruct photorealistic humans at 1K resolution from a single input image using a single GPU instantly. Additionally, it seamlessly supports various applications, as well as shape and texture editing tasks. Project page: https://yiyuzhuang.github.io/IDOL/.

Bin Ren, Hang Guo, Lei Sun et al.

This paper presents a comprehensive review of the NTIRE 2025 Challenge on Single-Image Efficient Super-Resolution (ESR). The challenge aimed to advance the development of deep models that optimize key computational metrics, i.e., runtime, parameters, and FLOPs, while achieving a PSNR of at least 26.90 dB on the $\operatorname{DIV2K\_LSDIR\_valid}$ dataset and 26.99 dB on the $\operatorname{DIV2K\_LSDIR\_test}$ dataset. A robust participation saw \textbf{244} registered entrants, with \textbf{43} teams submitting valid entries. This report meticulously analyzes these methods and results, emphasizing groundbreaking advancements in state-of-the-art single-image ESR techniques. The analysis highlights innovative approaches and establishes benchmarks for future research in the field.

Jiancheng Huang, Yi Huang, Jianzhuang Liu et al.

Text-conditional image editing is a practical AIGC task that has recently emerged with great commercial and academic value. For real image editing, most diffusion model-based methods use DDIM Inversion as the first stage before editing. However, DDIM Inversion often results in reconstruction failure, leading to unsatisfactory performance for downstream editing. To address this problem, we first analyze why the reconstruction via DDIM Inversion fails. We then propose a new inversion and sampling method named Dual-Schedule Inversion. We also design a classifier to adaptively combine Dual-Schedule Inversion with different editing methods for user-friendly image editing. Our work can achieve superior reconstruction and editing performance with the following advantages: 1) It can reconstruct real images perfectly without fine-tuning, and its reversibility is guaranteed mathematically. 2) The edited object/scene conforms to the semantics of the text prompt. 3) The unedited parts of the object/scene retain the original identity.

Yong Zhang, Rui Zhu, Shifeng Zhang et al.

Self-supervised contrastive learning heavily relies on the view variance brought by data augmentation, so that it can learn a view-invariant pre-trained representation. Beyond increasing the view variance for contrast, this work focuses on improving the diversity of training data, to improve the generalization and robustness of the pre-trained models. To this end, we propose a unified framework to conduct data augmentation in the feature space, known as feature augmentation. This strategy is domain-agnostic, which augments similar features to the original ones and thus improves the data diversity. We perform a systematic investigation of various feature augmentation architectures, the gradient-flow skill, and the relationship between feature augmentation and traditional data augmentation. Our study reveals some practical principles for feature augmentation in self-contrastive learning. By integrating feature augmentation on the instance discrimination or the instance similarity paradigm, we consistently improve the performance of pre-trained feature learning and gain better generalization over the downstream image classification and object detection task.

Yi Huang, Wei Xiong, He Zhang et al.

Building on the success of diffusion models in image generation and editing, video editing has recently gained substantial attention. However, maintaining temporal consistency and motion alignment still remains challenging. To address these issues, this paper proposes DINO-guided Video Editing (DIVE), a framework designed to facilitate subject-driven editing in source videos conditioned on either target text prompts or reference images with specific identities. The core of DIVE lies in leveraging the powerful semantic features extracted from a pretrained DINOv2 model as implicit correspondences to guide the editing process. Specifically, to ensure temporal motion consistency, DIVE employs DINO features to align with the motion trajectory of the source video. For precise subject editing, DIVE incorporates the DINO features of reference images into a pretrained text-to-image model to learn Low-Rank Adaptations (LoRAs), effectively registering the target subject's identity. Extensive experiments on diverse real-world videos demonstrate that our framework can achieve high-quality editing results with robust motion consistency, highlighting the potential of DINO to contribute to video editing. Project page: https://dino-video-editing.github.io

Shifeng Chen, Yihui Li, Jun Liao et al.

Recent advances in 3D scene editing using NeRF and 3DGS enable high-quality static scene editing. In contrast, dynamic scene editing remains challenging, as methods that directly extend 2D diffusion models to 4D often produce motion artifacts, temporal flickering, and inconsistent style propagation. We introduce Catalyst4D, a framework that transfers high-quality 3D edits to dynamic 4D Gaussian scenes while maintaining spatial and temporal coherence. At its core, Anchor-based Motion Guidance (AMG) builds a set of structurally stable and spatially representative anchors from both original and edited Gaussians. These anchors serve as robust region-level references, and their correspondences are established via optimal transport to enable consistent deformation propagation without cross-region interference or motion drift. Complementarily, Color Uncertainty-guided Appearance Refinement (CUAR) preserves temporal appearance consistency by estimating per-Gaussian color uncertainty and selectively refining regions prone to occlusion-induced artifacts. Extensive experiments demonstrate that Catalyst4D achieves temporally stable, high-fidelity dynamic scene editing and outperforms existing methods in both visual quality and motion coherence.

Yuqi Peng, Pengfei Wang, Jianzhuang Liu et al.

Pre-trained vision-language models, such as CLIP, show impressive zero-shot recognition ability and can be easily transferred to specific downstream tasks via prompt tuning, even with limited training data. However, existing prompt tuning methods face two main challenges: (1) In few-shot scenarios, data scarcity often leads to overfitting, making the model sensitive to changes in the input domain. (2) To mitigate overfitting, these methods typically rely on complex task-specific model architectures and sensitive hyperparameter tuning, severely restricting their general applicability. To address these issues, we propose a simpler and more general framework called GLAD (Generalizable LoRA tuning with RegulArized GraDient). We show that merely applying LoRA achieves performance in downstream tasks comparable to current state-of-the-art prompt-based methods. While LoRA is effective and easy to use, it remains susceptible to overfitting in few-shot learning scenarios. To mitigate this risk, we introduce a gradient-based regularization technique. This technique effectively steers the optimization trajectory, encouraging the model to find a more stable parameter region that is robust to variations in data distribution. Through extensive experiments conducted on 15 benchmark datasets, we demonstrate that GLAD outperforms previous tuning approaches in terms of base-to-novel class generalization, image domain generalization, and cross-dataset generalization. The code will be publicly available.

Jiyu Wu, Yifan Liu, Jiancheng Huang et al.

Existing shadow removal methods often rely on shadow masks, which are challenging to acquire in real-world scenarios. Exploring intrinsic image cues, such as local contrast information, presents a potential alternative for guiding shadow removal in the absence of explicit masks. However, the cue's inherent ambiguity becomes a critical limitation in complex scenes, where it can fail to distinguish true shadows from low-reflectance objects and intricate background textures. To address this motivation, we propose the Adaptive Gated Dual-Branch Attention (AGBA) mechanism. AGBA dynamically filters and re-weighs the contrast prior to effectively disentangle shadow features from confounding visual elements. Furthermore, to tackle the persistent challenge of restoring soft shadow boundaries and fine-grained details, we introduce a diffusion-based Frequency-Contrast Fusion Network (FCFN) that leverages high-frequency and contrast cues to guide the generative process. Extensive experiments demonstrate that our method achieves state-of-the-art results among mask-free approaches while maintaining competitive performance relative to mask-based methods.

Mingfu Yan, Jiancheng Huang, Yifan Liu et al.

Generalized Category Discovery (GCD) tackles the challenging problem of categorizing unlabeled images into both known and novel classes within a partially labeled dataset, without prior knowledge of the number of unknown categories. Traditional methods often rely on rigid assumptions, such as predefining the number of classes, which limits their ability to handle the inherent variability and complexity of real-world data. To address these shortcomings, we propose AdaGCD, a cluster-centric contrastive learning framework that incorporates Adaptive Slot Attention (AdaSlot) into the GCD framework. AdaSlot dynamically determines the optimal number of slots based on data complexity, removing the need for predefined slot counts. This adaptive mechanism facilitates the flexible clustering of unlabeled data into known and novel categories by dynamically allocating representational capacity. By integrating adaptive representation with dynamic slot allocation, our method captures both instance-specific and spatially clustered features, improving class discovery in open-world scenarios. Extensive experiments on public and fine-grained datasets validate the effectiveness of our framework, emphasizing the advantages of leveraging spatial local information for category discovery in unlabeled image datasets.

Tong Li, Long Liu, Yihang Hu et al.

Knowledge distillation (KD) provides an effective way to improve the performance of a student network under the guidance of pre-trained teachers. However, this approach usually brings in a large capacity gap between teacher and student networks, limiting the distillation gains. Previous methods addressing this problem either discard accurate knowledge representation or fail to dynamically adjust the transferred knowledge, which is less effective in addressing the capacity gap problem and hinders students from achieving comparable performance with the pre-trained teacher. In this work, we extend the ideology of prompt-based learning to address the capacity gap problem, and propose Dual-Forward Path Teacher Knowledge Distillation (DFPT-KD), which replaces the pre-trained teacher with a novel dual-forward path teacher to supervise the learning of student. The key to DFPT-KD is prompt-based tuning, i.e., establishing an additional prompt-based forward path within the pre-trained teacher and optimizing it with the pre-trained teacher frozen to make the transferred knowledge compatible with the representation ability of the student. Extensive experiments demonstrate that DFPT-KD leads to trained students performing better than the vanilla KD. To make the transferred knowledge better compatible with the representation abilities of the student, we further fine-tune the whole prompt-based forward path, yielding a novel distillation approach dubbed DFPT-KD+. By extensive experiments, it is shown that DFPT-KD+ improves upon DFPT-KD and achieves state-of-the-art accuracy performance.

Yifan Liu, Jiancheng Huang, Na Liu et al.

Document shadows are a major obstacle in the digitization process. Due to the dense information in text and patterns covered by shadows, document shadow removal requires specialized methods. Existing document shadow removal methods, although showing some progress, still rely on additional information such as shadow masks or lack generalization and effectiveness across different shadow scenarios. This often results in incomplete shadow removal or loss of original document content and tones. Moreover, these methods tend to underutilize the information present in the original shadowed document image. In this paper, we refocus our approach on the document images themselves, which inherently contain rich information.We propose an end-to-end document shadow removal method guided by contrast representation, following a coarse-to-fine refinement approach. By extracting document contrast information, we can effectively and quickly locate shadow shapes and positions without the need for additional masks. This information is then integrated into the refined shadow removal process, providing better guidance for network-based removal and feature fusion. Extensive qualitative and quantitative experiments show that our method achieves state-of-the-art performance.

Yi Huang, Jiancheng Huang, Jianzhuang Liu et al.

Latest diffusion-based methods for many image restoration tasks outperform traditional models, but they encounter the long-time inference problem. To tackle it, this paper proposes a Wavelet-Based Diffusion Model (WaveDM). WaveDM learns the distribution of clean images in the wavelet domain conditioned on the wavelet spectrum of degraded images after wavelet transform, which is more time-saving in each step of sampling than modeling in the spatial domain. To ensure restoration performance, a unique training strategy is proposed where the low-frequency and high-frequency spectrums are learned using distinct modules. In addition, an Efficient Conditional Sampling (ECS) strategy is developed from experiments, which reduces the number of total sampling steps to around 5. Evaluations on twelve benchmark datasets including image raindrop removal, rain steaks removal, dehazing, defocus deblurring, demoiréing, and denoising demonstrate that WaveDM achieves state-of-the-art performance with the efficiency that is comparable to traditional one-pass methods and over 100$\times$ faster than existing image restoration methods using vanilla diffusion models.

Guotai Wang, Shuwei Zhai, Giovanni Lasio et al.

Computed Tomography (CT) plays an important role in monitoring radiation-induced Pulmonary Fibrosis (PF), where accurate segmentation of the PF lesions is highly desired for diagnosis and treatment follow-up. However, the task is challenged by ambiguous boundary, irregular shape, various position and size of the lesions, as well as the difficulty in acquiring a large set of annotated volumetric images for training. To overcome these problems, we propose a novel convolutional neural network called PF-Net and incorporate it into a semi-supervised learning framework based on Iterative Confidence-based Refinement And Weighting of pseudo Labels (I-CRAWL). Our PF-Net combines 2D and 3D convolutions to deal with CT volumes with large inter-slice spacing, and uses multi-scale guided dense attention to segment complex PF lesions. For semi-supervised learning, our I-CRAWL employs pixel-level uncertainty-based confidence-aware refinement to improve the accuracy of pseudo labels of unannotated images, and uses image-level uncertainty for confidence-based image weighting to suppress low-quality pseudo labels in an iterative training process. Extensive experiments with CT scans of Rhesus Macaques with radiation-induced PF showed that: 1) PF-Net achieved higher segmentation accuracy than existing 2D, 3D and 2.5D neural networks, and 2) I-CRAWL outperformed state-of-the-art semi-supervised learning methods for the PF lesion segmentation task. Our method has a potential to improve the diagnosis of PF and clinical assessment of side effects of radiotherapy for lung cancers.

Zhiwei Zhang, Yu Dong, Hanyu Peng et al.

One-class novelty detection is conducted to identify anomalous instances, with different distributions from the expected normal instances. In this paper, the Generative Adversarial Network based on the Encoder-Decoder-Encoder scheme (EDE-GAN) achieves state-of-the-art performance. The two factors bellow serve the above purpose: 1) The EDE-GAN calculates the distance between two latent vectors as the anomaly score, which is unlike the previous methods by utilizing the reconstruction error between images. 2) The model obtains best results when the batch size is set to 1. To illustrate their superiority, we design a new GAN architecture, and compare performances according to different batch sizes. Moreover, with experimentation leads to discovery, our result implies there is also evidence of just how beneficial constraint on the latent space are when engaging in model training. In an attempt to learn compact and fast models, we present a new technology, Progressive Knowledge Distillation with GANs (P-KDGAN), which connects two standard GANs through the designed distillation loss. Two-step progressive learning continuously augments the performance of student GANs with improved results over single-step approach. Our experimental results on CIFAR-10, MNIST, and FMNIST datasets illustrate that P-KDGAN improves the performance of the student GAN by 2.44%, 1.77%, and 1.73% when compressing the computationat ratios of 24.45:1, 311.11:1, and 700:1, respectively.

Zhiwei Zhang, Shifeng Chen, Lei Sun

One-class novelty detection is to identify anomalous instances that do not conform to the expected normal instances. In this paper, the Generative Adversarial Networks (GANs) based on encoder-decoder-encoder pipeline are used for detection and achieve state-of-the-art performance. However, deep neural networks are too over-parameterized to deploy on resource-limited devices. Therefore, Progressive Knowledge Distillation with GANs (PKDGAN) is proposed to learn compact and fast novelty detection networks. The P-KDGAN is a novel attempt to connect two standard GANs by the designed distillation loss for transferring knowledge from the teacher to the student. The progressive learning of knowledge distillation is a two-step approach that continuously improves the performance of the student GAN and achieves better performance than single step methods. In the first step, the student GAN learns the basic knowledge totally from the teacher via guiding of the pretrained teacher GAN with fixed weights. In the second step, joint fine-training is adopted for the knowledgeable teacher and student GANs to further improve the performance and stability. The experimental results on CIFAR-10, MNIST, and FMNIST show that our method improves the performance of the student GAN by 2.44%, 1.77%, and 1.73% when compressing the computation at ratios of 24.45:1, 311.11:1, and 700:1, respectively.

Wenhao Wu, Dongliang He, Xiao Tan et al.

Though action recognition in videos has achieved great success recently, it remains a challenging task due to the massive computational cost. Designing lightweight networks is a possible solution, but it may degrade the recognition performance. In this paper, we innovatively propose a general dynamic inference idea to improve inference efficiency by leveraging the variation in the distinguishability of different videos. The dynamic inference approach can be achieved from aspects of the network depth and the number of input video frames, or even in a joint input-wise and network depth-wise manner. In a nutshell, we treat input frames and network depth of the computational graph as a 2-dimensional grid, and several checkpoints are placed on this grid in advance with a prediction module. The inference is carried out progressively on the grid by following some predefined route, whenever the inference process comes across a checkpoint, an early prediction can be made depending on whether the early stop criteria meets. For the proof-of-concept purpose, we instantiate three dynamic inference frameworks using two well-known backbone CNNs. In these instances, we overcome the drawback of limited temporal coverage resulted from an early prediction by a novel frame permutation scheme, and alleviate the conflict between progressive computation and video temporal relation modeling by introducing an online temporal shift module. Extensive experiments are conducted to thoroughly analyze the effectiveness of our ideas and to inspire future research efforts. Results on various datasets also evident the superiority of our approach.

Yu Dong, Yihao Liu, He Zhang et al.

Recently, convolutional neural networks (CNNs) have achieved great improvements in single image dehazing and attained much attention in research. Most existing learning-based dehazing methods are not fully end-to-end, which still follow the traditional dehazing procedure: first estimate the medium transmission and the atmospheric light, then recover the haze-free image based on the atmospheric scattering model. However, in practice, due to lack of priors and constraints, it is hard to precisely estimate these intermediate parameters. Inaccurate estimation further degrades the performance of dehazing, resulting in artifacts, color distortion and insufficient haze removal. To address this, we propose a fully end-to-end Generative Adversarial Networks with Fusion-discriminator (FD-GAN) for image dehazing. With the proposed Fusion-discriminator which takes frequency information as additional priors, our model can generator more natural and realistic dehazed images with less color distortion and fewer artifacts. Moreover, we synthesize a large-scale training dataset including various indoor and outdoor hazy images to boost the performance and we reveal that for learning-based dehazing methods, the performance is strictly influenced by the training data. Experiments have shown that our method reaches state-of-the-art performance on both public synthetic datasets and real-world images with more visually pleasing dehazed results.

Wenhao Wu, Dongliang He, Xiao Tan et al.

Video Recognition has drawn great research interest and great progress has been made. A suitable frame sampling strategy can improve the accuracy and efficiency of recognition. However, mainstream solutions generally adopt hand-crafted frame sampling strategies for recognition. It could degrade the performance, especially in untrimmed videos, due to the variation of frame-level saliency. To this end, we concentrate on improving untrimmed video classification via developing a learning-based frame sampling strategy. We intuitively formulate the frame sampling procedure as multiple parallel Markov decision processes, each of which aims at picking out a frame/clip by gradually adjusting an initial sampling. Then we propose to solve the problems with multi-agent reinforcement learning (MARL). Our MARL framework is composed of a novel RNN-based context-aware observation network which jointly models context information among nearby agents and historical states of a specific agent, a policy network which generates the probability distribution over a predefined action space at each step and a classification network for reward calculation as well as final recognition. Extensive experimental results show that our MARL-based scheme remarkably outperforms hand-crafted strategies with various 2D and 3D baseline methods. Our single RGB model achieves a comparable performance of ActivityNet v1.3 champion submission with multi-modal multi-model fusion and new state-of-the-art results on YouTube Birds and YouTube Cars.

Ming Zhang, Xuefei Zhe, Le Ou-Yang et al.

Deep hashing models have been proposed as an efficient method for large-scale similarity search. However, most existing deep hashing methods only utilize fine-level labels for training while ignoring the natural semantic hierarchy structure. This paper presents an effective method that preserves the classwise similarity of full-level semantic hierarchy for large-scale image retrieval. Experiments on two benchmark datasets show that our method helps improve the fine-level retrieval performance. Moreover, with the help of the semantic hierarchy, it can produce significantly better binary codes for hierarchical retrieval, which indicates its potential of providing more user-desired retrieval results.

Xuefei Zhe, Shifeng Chen, Hong Yan

Deep supervised hashing has emerged as an influential solution to large-scale semantic image retrieval problems in computer vision. In the light of recent progress, convolutional neural network based hashing methods typically seek pair-wise or triplet labels to conduct the similarity preserving learning. However, complex semantic concepts of visual contents are hard to capture by similar/dissimilar labels, which limits the retrieval performance. Generally, pair-wise or triplet losses not only suffer from expensive training costs but also lack in extracting sufficient semantic information. In this regard, we propose a novel deep supervised hashing model to learn more compact class-level similarity preserving binary codes. Our deep learning based model is motivated by deep metric learning that directly takes semantic labels as supervised information in training and generates corresponding discriminant hashing code. Specifically, a novel cubic constraint loss function based on Gaussian distribution is proposed, which preserves semantic variations while penalizes the overlap part of different classes in the embedding space. To address the discrete optimization problem introduced by binary codes, a two-step optimization strategy is proposed to provide efficient training and avoid the problem of gradient vanishing. Extensive experiments on four large-scale benchmark databases show that our model can achieve the state-of-the-art retrieval performance. Moreover, when training samples are limited, our method surpasses other supervised deep hashing methods with non-negligible margins.

Xuefei Zhe, Shifeng Chen, Hong Yan

Deep distance metric learning (DDML), which is proposed to learn image similarity metrics in an end-to-end manner based on the convolution neural network, has achieved encouraging results in many computer vision tasks.$L2$-normalization in the embedding space has been used to improve the performance of several DDML methods. However, the commonly used Euclidean distance is no longer an accurate metric for $L2$-normalized embedding space, i.e., a hyper-sphere. Another challenge of current DDML methods is that their loss functions are usually based on rigid data formats, such as the triplet tuple. Thus, an extra process is needed to prepare data in specific formats. In addition, their losses are obtained from a limited number of samples, which leads to a lack of the global view of the embedding space. In this paper, we replace the Euclidean distance with the cosine similarity to better utilize the $L2$-normalization, which is able to attenuate the curse of dimensionality. More specifically, a novel loss function based on the von Mises-Fisher distribution is proposed to learn a compact hyper-spherical embedding space. Moreover, a new efficient learning algorithm is developed to better capture the global structure of the embedding space. Experiments for both classification and retrieval tasks on several standard datasets show that our method achieves state-of-the-art performance with a simpler training procedure. Furthermore, we demonstrate that, even with a small number of convolutional layers, our model can still obtain significantly better classification performance than the widely used softmax loss.