Qi Fan

Qi Fan, Wenjie Pei, Yu-Wing Tai et al.

Existing few-shot segmentation methods have achieved great progress based on the support-query matching framework. But they still heavily suffer from the limited coverage of intra-class variations from the few-shot supports provided. Motivated by the simple Gestalt principle that pixels belonging to the same object are more similar than those to different objects of same class, we propose a novel self-support matching strategy to alleviate this problem, which uses query prototypes to match query features, where the query prototypes are collected from high-confidence query predictions. This strategy can effectively capture the consistent underlying characteristics of the query objects, and thus fittingly match query features. We also propose an adaptive self-support background prototype generation module and self-support loss to further facilitate the self-support matching procedure. Our self-support network substantially improves the prototype quality, benefits more improvement from stronger backbones and more supports, and achieves SOTA on multiple datasets. Codes are at \url{https://github.com/fanq15/SSP}.

Peng Zheng, Huazhu Fu, Deng-Ping Fan et al.

In this paper, we present a novel end-to-end group collaborative learning network, termed GCoNet+, which can effectively and efficiently (250 fps) identify co-salient objects in natural scenes. The proposed GCoNet+ achieves the new state-of-the-art performance for co-salient object detection (CoSOD) through mining consensus representations based on the following two essential criteria: 1) intra-group compactness to better formulate the consistency among co-salient objects by capturing their inherent shared attributes using our novel group affinity module (GAM); 2) inter-group separability to effectively suppress the influence of noisy objects on the output by introducing our new group collaborating module (GCM) conditioning on the inconsistent consensus. To further improve the accuracy, we design a series of simple yet effective components as follows: i) a recurrent auxiliary classification module (RACM) promoting model learning at the semantic level; ii) a confidence enhancement module (CEM) assisting the model in improving the quality of the final predictions; and iii) a group-based symmetric triplet (GST) loss guiding the model to learn more discriminative features. Extensive experiments on three challenging benchmarks, i.e., CoCA, CoSOD3k, and CoSal2015, demonstrate that our GCoNet+ outperforms the existing 12 cutting-edge models. Code has been released at https://github.com/ZhengPeng7/GCoNet_plus.

Qi Fan, Xin Tao, Lei Ke et al.

The Segment Anything Model (SAM) achieves remarkable promptable segmentation given high-quality prompts which, however, often require good skills to specify. To make SAM robust to casual prompts, this paper presents the first comprehensive analysis on SAM's segmentation stability across a diverse spectrum of prompt qualities, notably imprecise bounding boxes and insufficient points. Our key finding reveals that given such low-quality prompts, SAM's mask decoder tends to activate image features that are biased towards the background or confined to specific object parts. To mitigate this issue, our key idea consists of calibrating solely SAM's mask attention by adjusting the sampling locations and amplitudes of image features, while the original SAM model architecture and weights remain unchanged. Consequently, our deformable sampling plugin (DSP) enables SAM to adaptively shift attention to the prompted target regions in a data-driven manner, facilitated by our effective robust training strategy (RTS). During inference, dynamic routing plugin (DRP) is proposed that toggles SAM between the deformable and regular grid sampling modes, conditioned on the input prompt quality. Thus, our solution, termed Stable-SAM, offers several advantages: 1) improved SAM's segmentation stability across a wide range of prompt qualities, while 2) retaining SAM's powerful promptable segmentation efficiency and generality, with 3) minimal learnable parameters (0.08 M) and fast adaptation (by 1 training epoch). Extensive experiments across multiple datasets validate the effectiveness and advantages of our approach, underscoring Stable-SAM as a more robust solution for segmenting anything. Codes will be released upon acceptance. https://github.com/fanq15/Stable-SAM

Qi Fan, Mattia Segu, Yu-Wing Tai et al.

Improving model's generalizability against domain shifts is crucial, especially for safety-critical applications such as autonomous driving. Real-world domain styles can vary substantially due to environment changes and sensor noises, but deep models only know the training domain style. Such domain style gap impedes model generalization on diverse real-world domains. Our proposed Normalization Perturbation (NP) can effectively overcome this domain style overfitting problem. We observe that this problem is mainly caused by the biased distribution of low-level features learned in shallow CNN layers. Thus, we propose to perturb the channel statistics of source domain features to synthesize various latent styles, so that the trained deep model can perceive diverse potential domains and generalizes well even without observations of target domain data in training. We further explore the style-sensitive channels for effective style synthesis. Normalization Perturbation only relies on a single source domain and is surprisingly effective and extremely easy to implement. Extensive experiments verify the effectiveness of our method for generalizing models under real-world domain shifts.

Zhengxuan Wei, Xu Guo, Xinghui Li et al.

Video world models aim to simulate controllable visual environments, but long-horizon rollouts depend on what the model remembers after observations leave its native context window. Explicit memories retain frames or online 3D reconstructions, which can suffer from heuristic retrieval errors, redundant appearance storage, or reconstruction artifacts. Implicit memories compress history into a compact state, but existing designs are not explicitly constrained to encode cross-view scene geometry. We propose GIM-World, a geometry-aware implicit memory framework for video world models. A lightweight transformer encoder compresses variable-length history into fixed-size memory tokens, a camera-queryable geometry head distills 3D scene structure from a frozen foundation model into the memory during training, and an information-guided pruning rule keeps encoding cost bounded as history grows. The geometry teacher is discarded at inference, leaving a lightweight memory module. Experiments on MIND show that GIM-World better preserves long-horizon geometric and visual consistency than both explicit- and implicit-memory baselines.

Yanan Sun, Zihan Zhong, Qi Fan et al.

Large-scale joint training of multimodal models, e.g., CLIP, have demonstrated great performance in many vision-language tasks. However, image-text pairs for pre-training are restricted to the intersection of images and texts, limiting their ability to cover a large distribution of real-world data, where noise can also be introduced as misaligned pairs during pre-processing. Conversely, unimodal models trained on text or image data alone through unsupervised techniques can achieve broader coverage of diverse real-world data and are not constrained by the requirement of simultaneous presence of image and text. In this paper, we demonstrate that using large-scale unsupervised unimodal models as pre-training can enhance the zero-shot performance of image-text pair models. Our thorough studies validate that models pre-trained as such can learn rich representations of both modalities, improving their ability to understand how images and text relate to each other. Our experiments show that unimodal pre-training outperforms state-of-the-art CLIP-based models by 6.5% (52.3% $\rightarrow$ 58.8%) on PASCAL-5$^i$ and 6.2% (27.2% $\rightarrow$ 33.4%) on COCO-20$^i$ semantic segmentation under zero-shot setting respectively. By learning representations of both modalities, unimodal pre-training offers broader coverage, reduced misalignment errors, and the ability to capture more complex features and patterns in the real-world data resulting in better performance especially for zero-shot vision-language tasks.

Qi Fan, Yutong Li, Yi Xin et al.

The Multimodal Emotion Recognition challenge MER2024 focuses on recognizing emotions using audio, language, and visual signals. In this paper, we present our submission solutions for the Semi-Supervised Learning Sub-Challenge (MER2024-SEMI), which tackles the issue of limited annotated data in emotion recognition. Firstly, to address the class imbalance, we adopt an oversampling strategy. Secondly, we propose a modality representation combinatorial contrastive learning (MR-CCL) framework on the trimodal input data to establish robust initial models. Thirdly, we explore a self-training approach to expand the training set. Finally, we enhance prediction robustness through a multi-classifier weighted soft voting strategy. Our proposed method is validated to be effective on the MER2024-SEMI Challenge, achieving a weighted average F-score of 88.25% and ranking 6th on the leaderboard. Our project is available at https://github.com/WooyoohL/MER2024-SEMI.

Shangge Liu, Yuehan Yin, Lei Wang et al.

Task arithmetic provides an efficient, training-free way to edit pre-trained models, yet lacks a fundamental theoretical explanation for its success. The existing concept of ``weight disentanglement" describes the ideal outcome of non-interfering task composition but does not reveal its underlying cause. Crucially, what intrinsic properties of the pre-trained model ($θ_0$) or the task vectors ($τ_t$) enable this disentanglement remains underexplored. In this paper, we introduce Task-Feature Specialization (TFS), a model's ability to allocate distinct internal features to different tasks, as the fundamental principle. We first prove that TFS is a sufficient condition for weight disentanglement. More importantly, we find that TFS also gives rise to an observable geometric consequence: weight vector orthogonality. This positions TFS as the common cause for both the desired functional outcome (disentanglement) and a measurable geometric property (orthogonality). This relationship provides the key insight for our method: since the abstract TFS property is intractable to enforce directly, we can instead promote weight disentanglement by shaping its concrete geometric consequence, orthogonality. Therefore, we propose OrthoReg, a simple and effective regularization method that actively enforces an internal orthogonal structure on weight updates ($ΔW$) that constitute $τ_t$ during fine-tuning. And we theoretically prove that OrthoReg promotes disentanglement. Extensive experiments demonstrate that OrthoReg consistently and significantly enhances the performance of various task arithmetic methods. Code is available at \href{https://github.com/RL-MIND/OrthoReg}{https://github.com/RL-MIND/OrthoReg}.

Qihong Tang, Changhan Liu, Shaofeng Zhang et al.

Identifying potential objects is critical for object recognition and analysis across various computer vision applications. Existing methods typically localize potential objects by relying on exemplar images, predefined categories, or textual descriptions. However, their reliance on image and text prompts often limits flexibility, restricting adaptability in real-world scenarios. In this paper, we introduce a novel Prompt-Free Universal Region Proposal Network (PF-RPN), which identifies potential objects without relying on external prompts. First, the Sparse Image-Aware Adapter (SIA) module performs initial localization of potential objects using a learnable query embedding dynamically updated with visual features. Next, the Cascade Self-Prompt (CSP) module identifies the remaining potential objects by leveraging the self-prompted learnable embedding, autonomously aggregating informative visual features in a cascading manner. Finally, the Centerness-Guided Query Selection (CG-QS) module facilitates the selection of high-quality query embeddings using a centerness scoring network. Our method can be optimized with limited data (e.g., 5% of MS COCO data) and applied directly to various object detection application domains for identifying potential objects without fine-tuning, such as underwater object detection, industrial defect detection, and remote sensing image object detection. Experimental results across 19 datasets validate the effectiveness of our method. Code is available at https://github.com/tangqh03/PF-RPN.

Qi Fan, Haolin Zuo, Rui Liu et al.

Multimodal emotion recognition (MER) in practical scenarios is significantly challenged by the presence of missing or incomplete data across different modalities. To overcome these challenges, researchers have aimed to simulate incomplete conditions during the training phase to enhance the system's overall robustness. Traditional methods have often involved discarding data or substituting data segments with zero vectors to approximate these incompletenesses. However, such approaches neither accurately represent real-world conditions nor adequately address the issue of noisy data availability. For instance, a blurry image cannot be simply replaced with zero vectors, while still retaining information. To tackle this issue and develop a more precise MER system, we introduce a novel noise-robust MER model that effectively learns robust multimodal joint representations from noisy data. This approach includes two pivotal components: firstly, a noise scheduler that adjusts the type and level of noise in the data to emulate various realistic incomplete situations. Secondly, a Variational AutoEncoder (VAE)-based module is employed to reconstruct these robust multimodal joint representations from the noisy inputs. Notably, the introduction of the noise scheduler enables the exploration of an entirely new type of incomplete data condition, which is impossible with existing methods. Extensive experimental evaluations on the benchmark datasets IEMOCAP and CMU-MOSEI demonstrate the effectiveness of the noise scheduler and the excellent performance of our proposed model. Our project is publicly available on https://github.com/WooyoohL/Noise-robust_MER.

Yunkai Dang, Yizhu Jiang, Yifan Jiang et al.

Multimodal Large Language Models (MLLMs) suffer from substantial computational overhead due to the high redundancy in visual token sequences. Existing approaches typically address this issue using single-layer Vision Transformer (ViT) features and static pruning strategies. However, such fixed configurations are often brittle under diverse instructions. To overcome these limitations, we propose CLASP, a plug-and-play token reduction framework based on class-adaptive layer fusion and dual-stage pruning. Specifically, CLASP first constructs category-specific visual representations through multi-layer vision feature fusion. It then performs dual-stage pruning, allocating the token budget between attention-salient pivot tokens for relevance and redundancy-aware completion tokens for coverage. Through class-adaptive pruning, CLASP enables prompt-conditioned feature fusion and budget allocation, allowing aggressive yet robust visual token reduction. Extensive experiments demonstrate that CLASP consistently outperforms existing methods across a wide range of benchmarks, pruning ratios, and MLLM architectures. Code will be available at https://github.com/Yunkaidang/CLASP.

Qi Fan, Hongyu Yuan, Haolin Zuo et al.

Multimodal emotion recognition utilizes complete multimodal information and robust multimodal joint representation to gain high performance. However, the ideal condition of full modality integrity is often not applicable in reality and there always appears the situation that some modalities are missing. For example, video, audio, or text data is missing due to sensor failure or network bandwidth problems, which presents a great challenge to MER research. Traditional methods extract useful information from the complete modalities and reconstruct the missing modalities to learn robust multimodal joint representation. These methods have laid a solid foundation for research in this field, and to a certain extent, alleviated the difficulty of multimodal emotion recognition under missing modalities. However, relying solely on internal reconstruction and multimodal joint learning has its limitations, especially when the missing information is critical for emotion recognition. To address this challenge, we propose a novel framework of Retrieval Augment for Missing Modality Multimodal Emotion Recognition (RAMER), which introduces similar multimodal emotion data to enhance the performance of emotion recognition under missing modalities. By leveraging databases, that contain related multimodal emotion data, we can retrieve similar multimodal emotion information to fill in the gaps left by missing modalities. Various experimental results demonstrate that our framework is superior to existing state-of-the-art approaches in missing modality MER tasks. Our whole project is publicly available on https://github.com/WooyoohL/Retrieval_Augment_MER.

Zheng Cai, Maosong Cao, Haojiong Chen et al. · pku

The evolution of Large Language Models (LLMs) like ChatGPT and GPT-4 has sparked discussions on the advent of Artificial General Intelligence (AGI). However, replicating such advancements in open-source models has been challenging. This paper introduces InternLM2, an open-source LLM that outperforms its predecessors in comprehensive evaluations across 6 dimensions and 30 benchmarks, long-context modeling, and open-ended subjective evaluations through innovative pre-training and optimization techniques. The pre-training process of InternLM2 is meticulously detailed, highlighting the preparation of diverse data types including text, code, and long-context data. InternLM2 efficiently captures long-term dependencies, initially trained on 4k tokens before advancing to 32k tokens in pre-training and fine-tuning stages, exhibiting remarkable performance on the 200k ``Needle-in-a-Haystack" test. InternLM2 is further aligned using Supervised Fine-Tuning (SFT) and a novel Conditional Online Reinforcement Learning from Human Feedback (COOL RLHF) strategy that addresses conflicting human preferences and reward hacking. By releasing InternLM2 models in different training stages and model sizes, we provide the community with insights into the model's evolution.

Jiace Zhu, Wentao Chen, Qi Fan et al.

Recent studies have demonstrated the potential of Large Language Models (LLMs) in generating GPU Kernels. Current benchmarks focus on the translation of high-level languages into CUDA, overlooking the more general and challenging task of text-to-CUDA generation. Furthermore, given the hardware-specific and performance-critical features of GPU programming, accurately assessing the performance of LLM-generated GPU programs is nontrivial. In this work, we introduce CUDABench, a comprehensive benchmark designed to evaluate the text-to-CUDA capabilities of LLMs. First, we construct CUDABench-Set, which covers Breadth-Depth-Difficulty evaluation space in diverse application domains, including artificial intelligence, scientific computing, and data analytics, etc. Furthermore, we propose CUDABench-Score and Generative Verification Pipeline that assess (1) compilation correctness, (2) functional consistency through execution-based verification, and (3) a novel roofline-based metric, Performance-Score. Benchmarking state-of-the-art LLMs reveals insightful findings and challenges of text-to-CUDA, such as a notable mismatch between high compilation success rates and low functional correctness, a lack of domain-specific algorithmic knowledge, and suboptimal utilization of GPU hardware resources. Our benchmark is available at https://github.com/CUDA-Bench/CUDABench.

Yunkai Dang, Meiyi Zhu, Donghao Wang et al.

Multimodal large language models (MLLMs) demonstrate strong perception and reasoning performance on existing remote sensing (RS) benchmarks. However, most prior benchmarks rely on low-resolution imagery, and some high-resolution benchmarks suffer from flawed reasoning-task designs. We show that text-only LLMs can perform competitively with multimodal vision-language models on RS reasoning tasks without access to images, revealing a critical mismatch between current benchmarks and the intended evaluation of visual understanding. To enable faithful assessment, we introduce RSHR-Bench, a super-high-resolution benchmark for RS visual understanding and reasoning. RSHR-Bench contains 5,329 full-scene images with a long side of at least 4,000 pixels, with up to about 3 x 10^8 pixels per image, sourced from widely used RS corpora and UAV collections. We design four task families: multiple-choice VQA, open-ended VQA, image captioning, and single-image evaluation. These tasks cover nine perception categories and four reasoning types, supporting multi-turn and multi-image dialog. To reduce reliance on language priors, we apply adversarial filtering with strong LLMs followed by rigorous human verification. Overall, we construct 3,864 VQA tasks, 3,913 image captioning tasks, and 500 fully human-written or verified single-image evaluation VQA pairs. Evaluations across open-source, closed-source, and RS-specific VLMs reveal persistent performance gaps in super-high-resolution scenarios. Code: https://github.com/Yunkaidang/RSHR

Xueqing Deng, Qi Fan, Xiaojie Jin et al.

Fine-tuning pre-trained transformer models, e.g., Swin Transformer, are successful in numerous downstream for dense prediction vision tasks. However, one major issue is the cost/storage of their huge amount of parameters, which becomes increasingly challenging to handle with the growing amount of vision tasks. In this paper, we propose an effective approach to alleviate the issue, namely selective feature adapter (SFA). It achieves state-of-the-art (SoTA) performance under any given budget of trainable parameters, and demonstrates comparable or better performance than fully fine-tuned models across various dense tasks. Specifically, SFA consists of external adapters and internal adapters which are sequentially operated over a transformer model. For external adapters, we properly select the places and amount of additional multilayer perception (MLP). For internal adapters, we transform a few task-important parameters inside the transformer, which are automatically discovered through a simple yet effective lottery ticket algorithm. Our experiments show that the dual adapter module, a.k.a SFA, is essential to achieve the best trade-off on dense vision tasks, such as segmentation, detection and depth-estimation, outperforming other adapters with a single module.

Yunkai Dang, Donghao Wang, Jiacheng Yang et al.

Large vision-language models (VLMs) exhibit strong performance across various tasks. However, these VLMs encounter significant challenges when applied to the remote sensing domain due to the inherent differences between remote sensing images and natural images. Existing remote sensing VLMs often fail to extract fine-grained visual features and suffer from visual forgetting during deep language processing. To address this, we introduce MF-RSVLM, a Multi-Feature Fusion Remote Sensing Vision--Language Model that effectively extracts and fuses visual features for RS understanding. MF-RSVLM learns multi-scale visual representations and combines global context with local details, improving the capture of small and complex structures in RS scenes. A recurrent visual feature injection scheme ensures the language model remains grounded in visual evidence and reduces visual forgetting during generation. Extensive experiments on diverse RS benchmarks show that MF-RSVLM achieves state-of-the-art or highly competitive performance across remote sensing classification, image captioning, and VQA tasks. Our code is publicly available at https://github.com/Yunkaidang/RSVLM.

Shaofeng Zhang, Xuanqi Chen, Ning Liao et al.

The dominance of denoising generative models (e.g., diffusion, flow-matching) in visual synthesis is tempered by their substantial training costs and inefficiencies in representation learning. While injecting discriminative representations via auxiliary alignment has proven effective, this approach still faces key limitations: the reliance on external, pre-trained encoders introduces overhead and domain shift. A dispersed-based strategy that encourages strong separation among in-batch latent representations alleviates this specific dependency. To assess the effect of the number of negative samples in generative modeling, we propose {\mname}, a plug-and-play training framework that requires no external encoders. Our method integrates a memory bank mechanism that maintains a large, dynamically updated queue of negative samples across training iterations. This decouples the number of negatives from the mini-batch size, providing abundant and high-quality negatives for a contrastive objective without a multiplicative increase in computational cost. A low-dimensional projection head is used to further minimize memory and bandwidth overhead. {\mname} offers three principal advantages: (1) it is self-contained, eliminating dependency on pretrained vision foundation models and their associated forward-pass overhead; (2) it introduces no additional parameters or computational cost during inference; and (3) it enables substantially faster convergence, achieving superior generative quality more efficiently. On ImageNet-256, {\mname} achieves a state-of-the-art FID of \textbf{2.40} within 400k steps, significantly outperforming comparable methods.

Jiapeng Su, Qi Fan, Guangming Lu et al.

Few-shot semantic segmentation (FSS) has achieved great success on segmenting objects of novel classes, supported by only a few annotated samples. However, existing FSS methods often underperform in the presence of domain shifts, especially when encountering new domain styles that are unseen during training. It is suboptimal to directly adapt or generalize the entire model to new domains in the few-shot scenario. Instead, our key idea is to adapt a small adapter for rectifying diverse target domain styles to the source domain. Consequently, the rectified target domain features can fittingly benefit from the well-optimized source domain segmentation model, which is intently trained on sufficient source domain data. Training domain-rectifying adapter requires sufficiently diverse target domains. We thus propose a novel local-global style perturbation method to simulate diverse potential target domains by perturbating the feature channel statistics of the individual images and collective statistics of the entire source domain, respectively. Additionally, we propose a cyclic domain alignment module to facilitate the adapter effectively rectifying domains using a reverse domain rectification supervision. The adapter is trained to rectify the image features from diverse synthesized target domains to align with the source domain. During testing on target domains, we start by rectifying the image features and then conduct few-shot segmentation on the domain-rectified features. Extensive experiments demonstrate the effectiveness of our method, achieving promising results on cross-domain few-shot semantic segmentation tasks. Our code is available at https://github.com/Matt-Su/DR-Adapter.

Junyuan Ma, Xunzhi Xiang, Wenbin Li et al.

Vision foundation models (VFMs) have achieved strong performance across various vision tasks. However, it still remains challenging to apply VFMs for cross-domain few-shot segmentation (CD-FSS), which segments objects of novel classes under domain shifts using only a few labeled exemplars. The challenge is mainly driven by two factors: (1) limited labeled exemplars per novel class relative to the scale of VFM pre-training, making the model prone to overfitting during retraining, and (2) target-domain shifts underrepresented during pre-training, inducing cross-domain inconsistency and layer-wise sensitivity. To address these issues, we propose Hierarchical Exemplar Representation Adaptation (HERA), a three-stage select-regularize-calibrate VFM-based segmentation framework that learns effectively from limited labels and adapts to novel domains without source-data retraining. We first design Hierarchical Layer Selection (HLS) to adaptively identify the most informative VFM layer using a data-dependent Exemplar Transfer Risk (ETR) computed for each candidate layer. Then, Prior-Guided Regularization (PGR) regularizes interactions on the selected representation, yielding well-structured local signals for the subsequent stage. Furthermore, Pixelwise Adaptive Calibration (PAC) combines the selected representation with the refined interaction maps to calibrate pixel-wise predictions, producing consistent masks. Together, these stages form a hierarchical select-regularize-calibrate pipeline that guides frozen VFM features in new domains while fine-tuning less than 2.7% of parameters at test time. Extensive experiments show that HERA surpasses the state of the art by more than 4.1 mIoU across multiple CD-FSS benchmarks.

Cheng Liang, Haoxian Chen, Liang Hou et al.

The quadratic complexity of the attention mechanism severely limits the context scalability of Video Diffusion Transformers (DiTs). We find that the highly sparse spatio-temporal attention patterns exhibited in Video DiTs can be naturally represented by the Monarch matrix. It is a class of structured matrices with flexible sparsity, enabling sub-quadratic attention via an alternating minimization algorithm. Accordingly, we propose VMonarch, a novel attention mechanism for Video DiTs that enables efficient computation over the dynamic sparse patterns with structured Monarch matrices. First, we adapt spatio-temporal Monarch factorization to explicitly capture the intra-frame and inter-frame correlations of the video data. Second, we introduce a recomputation strategy to mitigate artifacts arising from instabilities during alternating minimization of Monarch matrices. Third, we propose a novel online entropy algorithm fused into FlashAttention, enabling fast Monarch matrix updates for long sequences. Extensive experiments demonstrate that VMonarch achieves comparable or superior generation quality to full attention on VBench after minimal tuning. It overcomes the attention bottleneck in Video DiTs, reduces attention FLOPs by a factor of 17.5, and achieves a speedup of over 5x in attention computation for long videos, surpassing state-of-the-art sparse attention methods at 90% sparsity.

Xunzhi Xiang, Zixuan Duan, Guiyu Zhang et al.

Distilled autoregressive diffusion models facilitate real-time short video synthesis but suffer from severe error accumulation during long-sequence generation. While existing Test-Time Optimization (TTO) methods prove effective for images or short clips, we identify that they fail to mitigate drift in extended sequences due to unstable reward landscapes and the hypersensitivity of distilled parameters. To overcome these limitations, we introduce Test-Time Correction (TTC), a training-free alternative. Specifically, TTC utilizes the initial frame as a stable reference anchor to calibrate intermediate stochastic states along the sampling trajectory. Extensive experiments demonstrate that our method seamlessly integrates with various distilled models, extending generation lengths with negligible overhead while matching the quality of resource-intensive training-based methods on 30-second benchmarks.

Qi Fan, An Zou, Yehan Ma

Large Language Models (LLMs) are increasingly deployed in time-critical systems, such as robotics, autonomous driving, embodied intelligence, and industrial automation, where generating accurate responses within a given time budget is crucial for decision-making, control, or safety-critical tasks. However, the auto-regressive generation process of LLMs makes it challenging to model and estimate the end-to-end execution time. Furthermore, existing efficient inference methods based on a fixed key-value (KV) cache eviction ratio struggle to adapt to varying tasks with diverse time budgets, where an improper eviction ratio may lead to incomplete inference or a drop in response performance. In this paper, we propose TimeBill, a novel time-budgeted inference framework for LLMs that balances the inference efficiency and response performance. To be more specific, we propose a fine-grained response length predictor (RLP) and an execution time estimator (ETE) to accurately predict the end-to-end execution time of LLMs. Following this, we develop a time-budgeted efficient inference approach that adaptively adjusts the KV cache eviction ratio based on execution time prediction and the given time budget. Finally, through extensive experiments, we demonstrate the advantages of TimeBill in improving task completion rate and maintaining response performance under various overrun strategies.

Rui Liu, Haolin Zuo, Zheng Lian et al.

Missing modalities have recently emerged as a critical research direction in multimodal emotion recognition (MER). Conventional approaches typically address this issue through missing modality reconstruction. However, these methods fail to account for variations in reconstruction difficulty across different samples, consequently limiting the model's ability to handle hard samples effectively. To overcome this limitation, we propose a novel Hardness-Aware Dynamic Curriculum Learning framework, termed HARDY-MER. Our framework operates in two key stages: first, it estimates the hardness level of each sample, and second, it strategically emphasizes hard samples during training to enhance model performance on these challenging instances. Specifically, we first introduce a Multi-view Hardness Evaluation mechanism that quantifies reconstruction difficulty by considering both Direct Hardness (modality reconstruction errors) and Indirect Hardness (cross-modal mutual information). Meanwhile, we introduce a Retrieval-based Dynamic Curriculum Learning strategy that dynamically adjusts the training curriculum by retrieving samples with similar semantic information and balancing the learning focus between easy and hard instances. Extensive experiments on benchmark datasets demonstrate that HARDY-MER consistently outperforms existing methods in missing-modality scenarios. Our code will be made publicly available at https://github.com/HARDY-MER/HARDY-MER.

Qi Fan, Chi-Keung Tang, Yu-Wing Tai

We introduce Few-Shot Video Object Detection (FSVOD) with three contributions to real-world visual learning challenge in our highly diverse and dynamic world: 1) a large-scale video dataset FSVOD-500 comprising of 500 classes with class-balanced videos in each category for few-shot learning; 2) a novel Tube Proposal Network (TPN) to generate high-quality video tube proposals for aggregating feature representation for the target video object which can be highly dynamic; 3) a strategically improved Temporal Matching Network (TMN+) for matching representative query tube features with better discriminative ability thus achieving higher diversity. Our TPN and TMN+ are jointly and end-to-end trained. Extensive experiments demonstrate that our method produces significantly better detection results on two few-shot video object detection datasets compared to image-based methods and other naive video-based extensions. Codes and datasets are released at \url{https://github.com/fanq15/FewX}.

Qi Fan, Wei Zhuo, Chi-Keung Tang et al.

Conventional methods for object detection typically require a substantial amount of training data and preparing such high-quality training data is very labor-intensive. In this paper, we propose a novel few-shot object detection network that aims at detecting objects of unseen categories with only a few annotated examples. Central to our method are our Attention-RPN, Multi-Relation Detector and Contrastive Training strategy, which exploit the similarity between the few shot support set and query set to detect novel objects while suppressing false detection in the background. To train our network, we contribute a new dataset that contains 1000 categories of various objects with high-quality annotations. To the best of our knowledge, this is one of the first datasets specifically designed for few-shot object detection. Once our few-shot network is trained, it can detect objects of unseen categories without further training or fine-tuning. Our method is general and has a wide range of potential applications. We produce a new state-of-the-art performance on different datasets in the few-shot setting. The dataset link is https://github.com/fanq15/Few-Shot-Object-Detection-Dataset.

Zhe Gao, Shiyu Shen, Taifeng Chai et al.

Existing Multimodal Large Language Models (MLLMs) often suffer from hallucinations in long video understanding (LVU), primarily due to the imbalance between textual and visual tokens. Observing that MLLMs handle short visual inputs well, recent LVU works alleviate hallucinations by automatically parsing the vast visual data into manageable segments that can be effectively processed by MLLMs. SFT-based tool-calling methods can serve this purpose, but they typically require vast amounts of fine-grained, high-quality data and suffer from constrained tool-calling trajectories. We propose a novel VideoTIR that leverages Reinforcement Learning (RL) to encourage proper usage of comprehensive multi-level toolkits for efficient long video understanding. VideoTIR explores both Zero-RL and SFT cold-starting to enable MLLMs to retrieve and focus on meaningful video segments/images/regions, enhancing long video understanding both accurately and efficiently. To reduce redundant tool-calling, we propose Toolkit Action Grouped Policy Optimization (TAGPO), which enhances the efficiency of the calling process through stepwise reward assignment and reuse of failed rollouts. Additionally, we develop a sandbox-based trajectory synthesis framework to generate high-quality trajectories data. Extensive experiments on three long-video QA benchmarks demonstrate the effectiveness and efficiency of our method.

Wentao Chen, Jiace Zhu, Qi Fan et al.

Large Language Models (LLMs) have demonstrated strong capabilities in general-purpose code generation. However, generating the code which is deeply hardware-specific, architecture-aware, and performance-critical, especially for massively parallel GPUs, remains a complex challenge. In this work, we explore the use of LLMs for the automated generation and optimization of CUDA programs, with the goal of producing high-performance GPU kernels that fully exploit the underlying hardware. To address this challenge, we propose a novel framework called \textbf{Feature Search and Reinforcement (FSR)}. FSR jointly optimizes compilation and functional correctness, as well as the runtime performance, which are validated through extensive and diverse test cases, and measured by actual kernel execution latency on the target GPU, respectively. This approach enables LLMs not only to generate syntactically and semantically correct CUDA code but also to iteratively refine it for efficiency, tailored to the characteristics of the GPU architecture. We evaluate FSR on representative CUDA kernels, covering AI workloads and computational intensive algorithms. Our results show that LLMs augmented with FSR consistently guarantee correctness rates. Meanwhile, the automatically generated kernels can outperform general human-written code by a factor of up to 179$\times$ in execution speeds. These findings highlight the potential of combining LLMs with performance reinforcement to automate GPU programming for hardware-specific, architecture-sensitive, and performance-critical applications.

Dongdong Ren, Wenbin Li, Tianyu Ding et al.

Recent advancements in model pruning have focused on developing new algorithms and improving upon benchmarks. However, the practical application of these algorithms across various models and platforms remains a significant challenge. To address this challenge, we propose ONNXPruner, a versatile pruning adapter designed for the ONNX format models. ONNXPruner streamlines the adaptation process across diverse deep learning frameworks and hardware platforms. A novel aspect of ONNXPruner is its use of node association trees, which automatically adapt to various model architectures. These trees clarify the structural relationships between nodes, guiding the pruning process, particularly highlighting the impact on interconnected nodes. Furthermore, we introduce a tree-level evaluation method. By leveraging node association trees, this method allows for a comprehensive analysis beyond traditional single-node evaluations, enhancing pruning performance without the need for extra operations. Experiments across multiple models and datasets confirm ONNXPruner's strong adaptability and increased efficacy. Our work aims to advance the practical application of model pruning.

Haoran Fan, Qi Fan, Maurice Pagnucco et al.

Few-shot segmentation (FSS) aims to segment novel classes in a query image by using only a small number of supporting images from base classes. However, in cross-domain few-shot segmentation (CD-FSS), leveraging features from label-rich domains for resource-constrained domains poses challenges due to domain discrepancies. This work presents a Dynamically Adaptive Refine (DARNet) method that aims to balance generalization and specificity for CD-FSS. Our method includes the Channel Statistics Disruption (CSD) strategy, which perturbs feature channel statistics in the source domain, bolstering generalization to unknown target domains. Moreover, recognizing the variability across target domains, an Adaptive Refine Self-Matching (ARSM) method is also proposed to adjust the matching threshold and dynamically refine the prediction result with the self-matching method, enhancing accuracy. We also present a Test-Time Adaptation (TTA) method to refine the model's adaptability to diverse feature distributions. Our approach demonstrates superior performance against state-of-the-art methods in CD-FSS tasks.

Yikun Ji, Hong Yan, Jun Lan et al.

The rapid advancement of image generation technologies intensifies the demand for interpretable and robust detection methods. Although existing approaches often attain high accuracy, they typically operate as black boxes without providing human-understandable justifications. Multi-modal Large Language Models (MLLMs), while not originally intended for forgery detection, exhibit strong analytical and reasoning capabilities. When properly fine-tuned, they can effectively identify AI-generated images and offer meaningful explanations. However, existing MLLMs still struggle with hallucination and often fail to align their visual interpretations with actual image content and human reasoning. To bridge this gap, we construct a dataset of AI-generated images annotated with bounding boxes and descriptive captions that highlight synthesis artifacts, establishing a foundation for human-aligned visual-textual grounded reasoning. We then finetune MLLMs through a multi-stage optimization strategy that progressively balances the objectives of accurate detection, visual localization, and coherent textual explanation. The resulting model achieves superior performance in both detecting AI-generated images and localizing visual flaws, significantly outperforming baseline methods.

Xunzhi Xiang, Yabo Chen, Guiyu Zhang et al.

Current autoregressive diffusion models excel at video generation but are generally limited to short temporal durations. Our theoretical analysis indicates that the autoregressive modeling typically suffers from temporal drift caused by error accumulation and hinders parallelization in long video synthesis. To address these limitations, we propose a novel planning-then-populating framework centered on Macro-from-Micro Planning (MMPL) for long video generation. MMPL sketches a global storyline for the entire video through two hierarchical stages: Micro Planning and Macro Planning. Specifically, Micro Planning predicts a sparse set of future keyframes within each short video segment, offering motion and appearance priors to guide high-quality video segment generation. Macro Planning extends the in-segment keyframes planning across the entire video through an autoregressive chain of micro plans, ensuring long-term consistency across video segments. Subsequently, MMPL-based Content Populating generates all intermediate frames in parallel across segments, enabling efficient parallelization of autoregressive generation. The parallelization is further optimized by Adaptive Workload Scheduling for balanced GPU execution and accelerated autoregressive video generation. Extensive experiments confirm that our method outperforms existing long video generation models in quality and stability. Generated videos and comparison results are in our project page.

Jiahua Chen, Qihong Tang, Weinong Wang et al.

Although Multimodal Large Language Models have achieved remarkable progress, they still struggle with complex 3D spatial reasoning due to the reliance on 2D visual priors. Existing approaches typically mitigate this limitation either through computationally expensive post-training procedures on limited 3D datasets or through rigid tool-calling mechanisms that lack explicit geometric understanding and viewpoint flexibility. To address these challenges, we propose a \textit{training-free} framework that introduces a Visual Chain-of-Thought mechanism grounded in explicit 3D reconstruction. The proposed pipeline first reconstructs a high-fidelity 3D mesh from a single image using MLLM-guided keyword extraction and mask generation at multiple granularities. Subsequently, the framework leverages an external knowledge base to iteratively compute optimal camera extrinsic parameters and synthesize novel views, thereby emulating human perspective-taking. Extensive experiments demonstrate that the proposed approach significantly enhances spatial comprehension. Specifically, the framework outperforms specialized spatial models and general-purpose MLLMs, including \textit{GPT-5.2} and \textit{Gemini-2.5-Flash}, on major benchmarks such as 3DSRBench and Rel3D.

Xunzhi Xiang, Qi Fan

Autoregressive conditional image generation models have emerged as a dominant paradigm in text-to-image synthesis. These methods typically convert images into one-dimensional token sequences and leverage the self-attention mechanism, which has achieved remarkable success in natural language processing, to capture long-range dependencies, model global context, and ensure semantic coherence. However, excessively long contexts during inference lead to significant memory overhead caused by KV-cache and computational delays. To alleviate these challenges, we systematically analyze how global semantics, spatial layouts, and fine-grained textures are formed during inference, and propose a novel training-free context optimization method called Adaptive Dynamic Sparse Attention (ADSA). Conceptually, ADSA dynamically identifies historical tokens crucial for maintaining local texture consistency and those essential for ensuring global semantic coherence, thereby efficiently streamlining attention computation. Additionally, we introduce a dynamic KV-cache update mechanism tailored for ADSA, reducing GPU memory consumption during inference by approximately $50\%$. Extensive qualitative and quantitative experiments demonstrate the effectiveness and superiority of our approach in terms of both generation quality and resource efficiency.

Xunzhi Xiang, Xingye Tian, Guiyu Zhang et al.

Variational autoencoders (VAEs) typically encode images into a compact latent space, reducing computational cost but introducing an optimization dilemma: a higher-dimensional latent space improves reconstruction fidelity but often hampers generative performance. Recent methods attempt to address this dilemma by regularizing high-dimensional latent spaces using external vision foundation models (VFMs). However, it remains unclear how high-dimensional VAE latents affect the optimization of generative models. To our knowledge, our analysis is the first to reveal that redundant high-frequency components in high-dimensional latent spaces hinder the training convergence of diffusion models and, consequently, degrade generation quality. To alleviate this problem, we propose a spectral self-regularization strategy to suppress redundant high-frequency noise while simultaneously preserving reconstruction quality. The resulting Denoising-VAE, a ViT-based autoencoder that does not rely on VFMs, produces cleaner, lower-noise latents, leading to improved generative quality and faster optimization convergence. We further introduce a spectral alignment strategy to facilitate the optimization of Denoising-VAE-based generative models. Our complete method enables diffusion models to converge approximately 2$\times$ faster than with SD-VAE, while achieving state-of-the-art reconstruction quality (rFID = 0.28, PSNR = 27.26) and competitive generation performance (gFID = 1.82) on the ImageNet 256$\times$256 benchmark.

Qi Fan, Kaiqi Liu, Nian Liu et al.

Cross-domain few-shot segmentation (CD-FSS) aims to segment objects of novel classes in new domains, which is often challenging due to the diverse characteristics of target domains and the limited availability of support data. Most CD-FSS methods redesign and retrain in-domain FSS models using various domain-generalization techniques, which are effective but costly to train. To address these issues, we propose adapting informative model structures of the well-trained FSS model for target domains by learning domain characteristics from few-shot labeled support samples during inference, thereby eliminating the need for retraining. Specifically, we first adaptively identify domain-specific model structures by measuring parameter importance using a novel structure Fisher score in a data-dependent manner. Then, we progressively train the selected informative model structures with hierarchically constructed training samples, progressing from fewer to more support shots. The resulting Informative Structure Adaptation (ISA) method effectively addresses domain shifts and equips existing well-trained in-domain FSS models with flexible adaptation capabilities for new domains, eliminating the need to redesign or retrain CD-FSS models on base data. Extensive experiments validate the effectiveness of our method, demonstrating superior performance across multiple CD-FSS benchmarks.

Qi Fan, Deng-Ping Fan, Huazhu Fu et al.

We present a novel group collaborative learning framework (GCoNet) capable of detecting co-salient objects in real time (16ms), by simultaneously mining consensus representations at group level based on the two necessary criteria: 1) intra-group compactness to better formulate the consistency among co-salient objects by capturing their inherent shared attributes using our novel group affinity module; 2) inter-group separability to effectively suppress the influence of noisy objects on the output by introducing our new group collaborating module conditioning the inconsistent consensus. To learn a better embedding space without extra computational overhead, we explicitly employ auxiliary classification supervision. Extensive experiments on three challenging benchmarks, i.e., CoCA, CoSOD3k, and Cosal2015, demonstrate that our simple GCoNet outperforms 10 cutting-edge models and achieves the new state-of-the-art. We demonstrate this paper's new technical contributions on a number of important downstream computer vision applications including content aware co-segmentation, co-localization based automatic thumbnails, etc.

Qi Fan, Lei Ke, Wenjie Pei et al.

Partially supervised instance segmentation aims to perform learning on limited mask-annotated categories of data thus eliminating expensive and exhaustive mask annotation. The learned models are expected to be generalizable to novel categories. Existing methods either learn a transfer function from detection to segmentation, or cluster shape priors for segmenting novel categories. We propose to learn the underlying class-agnostic commonalities that can be generalized from mask-annotated categories to novel categories. Specifically, we parse two types of commonalities: 1) shape commonalities which are learned by performing supervised learning on instance boundary prediction; and 2) appearance commonalities which are captured by modeling pairwise affinities among pixels of feature maps to optimize the separability between instance and the background. Incorporating both the shape and appearance commonalities, our model significantly outperforms the state-of-the-art methods on both partially supervised setting and few-shot setting for instance segmentation on COCO dataset.