Lianghui Zhu

Lianghui Zhu, Yingyue Li, Jiemin Fang et al.

This paper explores the properties of the plain Vision Transformer (ViT) for Weakly-supervised Semantic Segmentation (WSSS). The class activation map (CAM) is of critical importance for understanding a classification network and launching WSSS. We observe that different attention heads of ViT focus on different image areas. Thus a novel weight-based method is proposed to end-to-end estimate the importance of attention heads, while the self-attention maps are adaptively fused for high-quality CAM results that tend to have more complete objects. Besides, we propose a ViT-based gradient clipping decoder for online retraining with the CAM results to complete the WSSS task. We name this plain Transformer-based Weakly-supervised learning framework WeakTr. It achieves the state-of-the-art WSSS performance on standard benchmarks, i.e., 78.4% mIoU on the val set of PASCAL VOC 2012 and 50.3% mIoU on the val set of COCO 2014. Code is available at https://github.com/hustvl/WeakTr.

Lianghui Zhu, Xinggang Wang, Xinlong Wang

Evaluating Large Language Models (LLMs) in open-ended scenarios is challenging because existing benchmarks and metrics can not measure them comprehensively. To address this problem, we propose to fine-tune LLMs as scalable judges (JudgeLM) to evaluate LLMs efficiently and effectively in open-ended benchmarks. We first propose a comprehensive, large-scale, high-quality dataset containing task seeds, LLMs-generated answers, and GPT-4-generated judgments for fine-tuning high-performance judges, as well as a new benchmark for evaluating the judges. We train JudgeLM at different scales from 7B, 13B, to 33B parameters, and conduct a systematic analysis of its capabilities and behaviors. We then analyze the key biases in fine-tuning LLM as a judge and consider them as position bias, knowledge bias, and format bias. To address these issues, JudgeLM introduces a bag of techniques including swap augmentation, reference support, and reference drop, which clearly enhance the judge's performance. JudgeLM obtains the state-of-the-art judge performance on both the existing PandaLM benchmark and our proposed new benchmark. Our JudgeLM is efficient and the JudgeLM-7B only needs 3 minutes to judge 5K samples with 8 A100 GPUs. JudgeLM obtains high agreement with the teacher judge, achieving an agreement exceeding 90% that even surpasses human-to-human agreement. JudgeLM also demonstrates extended capabilities in being judges of the single answer, multimodal models, multiple answers, multi-turn chat, etc. Code is available at https://github.com/baaivision/JudgeLM.

Lianghui Zhu, Yuxin Fang, Bencheng Liao et al.

Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2's efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .

Lianghui Zhu, Bencheng Liao, Qian Zhang et al.

Recently the state space models (SSMs) with efficient hardware-aware designs, i.e., the Mamba deep learning model, have shown great potential for long sequence modeling. Meanwhile building efficient and generic vision backbones purely upon SSMs is an appealing direction. However, representing visual data is challenging for SSMs due to the position-sensitivity of visual data and the requirement of global context for visual understanding. In this paper, we show that the reliance on self-attention for visual representation learning is not necessary and propose a new generic vision backbone with bidirectional Mamba blocks (Vim), which marks the image sequences with position embeddings and compresses the visual representation with bidirectional state space models. On ImageNet classification, COCO object detection, and ADE20k semantic segmentation tasks, Vim achieves higher performance compared to well-established vision transformers like DeiT, while also demonstrating significantly improved computation & memory efficiency. For example, Vim is 2.8$\times$ faster than DeiT and saves 86.8% GPU memory when performing batch inference to extract features on images with a resolution of 1248$\times$1248. The results demonstrate that Vim is capable of overcoming the computation & memory constraints on performing Transformer-style understanding for high-resolution images and it has great potential to be the next-generation backbone for vision foundation models. Code is available at https://github.com/hustvl/Vim.

Mingxi Ouyang, Yuqiu Fu, Renao Yan et al. · tsinghua

Recent advancements in computational pathology and artificial intelligence have significantly improved whole slide image (WSI) classification. However, the gigapixel resolution of WSIs and the scarcity of manual annotations present substantial challenges. Multiple instance learning (MIL) is a promising weakly supervised learning approach for WSI classification. Recently research revealed employing pseudo bag augmentation can encourage models to learn various data, thus bolstering models' performance. While directly inheriting the parents' labels can introduce more noise by mislabeling in training. To address this issue, we translate the WSI classification task from weakly supervised learning to semi-weakly supervised learning, termed SWS-MIL, where adaptive pseudo bag augmentation (AdaPse) is employed to assign labeled and unlabeled data based on a threshold strategy. Using the "student-teacher" pattern, we introduce a feature augmentation technique, MergeUp, which merges bags with low-priority bags to enhance inter-category information, increasing training data diversity. Experimental results on the CAMELYON-16, BRACS, and TCGA-LUNG datasets demonstrate the superiority of our method over existing state-of-the-art approaches, affirming its efficacy in WSI classification.

Lianghui Zhu, Junwei Zhou, Yan Liu et al.

Weakly supervised visual recognition using inexact supervision is a critical yet challenging learning problem. It significantly reduces human labeling costs and traditionally relies on multi-instance learning and pseudo-labeling. This paper introduces WeakSAM and solves the weakly-supervised object detection (WSOD) and segmentation by utilizing the pre-learned world knowledge contained in a vision foundation model, i.e., the Segment Anything Model (SAM). WeakSAM addresses two critical limitations in traditional WSOD retraining, i.e., pseudo ground truth (PGT) incompleteness and noisy PGT instances, through adaptive PGT generation and Region of Interest (RoI) drop regularization. It also addresses the SAM's problems of requiring prompts and category unawareness for automatic object detection and segmentation. Our results indicate that WeakSAM significantly surpasses previous state-of-the-art methods in WSOD and WSIS benchmarks with large margins, i.e. average improvements of 7.4% and 8.5%, respectively. The code is available at \url{https://github.com/hustvl/WeakSAM}.

Qiming He, Jing Li, Tian Guan et al.

Glomerular pathology is central to the diagnosis and prognosis of renal diseases, yet the heterogeneity of glomerular morphology and fine-grained lesion patterns remain challenging for current AI approaches. We present GloPath, an entity-centric foundation model trained on over one million glomeruli extracted from 14,049 renal biopsy specimens using multi-scale and multi-view self-supervised learning. GloPath addresses two major challenges in nephropathology: glomerular lesion assessment and clinicopathological insights discovery. For lesion assessment, GloPath was benchmarked across three independent cohorts on 52 tasks, including lesion recognition, grading, few-shot classification, and cross-modality diagnosis-outperforming state-of-the-art methods in 42 tasks (80.8%). In the large-scale real-world study, it achieved an ROC-AUC of 91.51% for lesion recognition, demonstrating strong robustness in routine clinical settings. For clinicopathological insights, GloPath systematically revealed statistically significant associations between glomerular morphological parameters and clinical indicators across 224 morphology-clinical variable pairs, demonstrating its capacity to connect tissue-level pathology with patient-level outcomes. Together, these results position GloPath as a scalable and interpretable platform for glomerular lesion assessment and clinicopathological discovery, representing a step toward clinically translatable AI in renal pathology.

Lianghui Zhu, Bin Ouyang, Yuxuan Zhang et al.

Text-prompted image segmentation enables fine-grained visual understanding and is critical for applications such as human-computer interaction and robotics. However, existing supervised fine-tuning methods typically ignore explicit chain-of-thought (CoT) reasoning at test time, which limits their ability to generalize to unseen prompts and domains. To address this issue, we introduce LENS, a scalable reinforcement-learning framework that jointly optimizes the reasoning process and segmentation in an end-to-end manner. We propose unified reinforcement-learning rewards that span sentence-, box-, and segment-level cues, encouraging the model to generate informative CoT rationales while refining mask quality. Using a publicly available 3-billion-parameter vision-language model, i.e., Qwen2.5-VL-3B-Instruct, LENS achieves an average cIoU of 81.2% on the RefCOCO, RefCOCO+, and RefCOCOg benchmarks, outperforming the strong fine-tuned method, i.e., GLaMM, by up to 5.6%. These results demonstrate that RL-driven CoT reasoning significantly enhances text-prompted segmentation and offers a practical path toward more generalizable Segment Anything models (SAM). Code is available at https://github.com/hustvl/LENS.

Minxi Ouyang, Lianghui Zhu, Yaqing Bao et al.

Multimodal large models have shown great potential in automating pathology image analysis. However, current multimodal models for gastrointestinal pathology are constrained by both data quality and reasoning transparency: pervasive noise and incomplete annotations in public datasets predispose vision language models to factual hallucinations when generating diagnostic text, while the absence of explicit intermediate reasoning chains renders the outputs difficult to audit and thus less trustworthy in clinical practice. To address these issues, we construct a large scale gastrointestinal pathology dataset containing both microscopic descriptions and diagnostic conclusions, and propose a prompt argumentation strategy that incorporates lesion classification and anatomical site information. This design guides the model to better capture image specific features and maintain semantic consistency in generation. Furthermore, we employ a post training pipeline that combines supervised fine tuning with Group Relative Policy Optimization (GRPO) to improve reasoning quality and output structure. Experimental results on real world pathology report generation tasks demonstrate that our approach significantly outperforms state of the art open source and proprietary baselines in terms of generation quality, structural completeness, and clinical relevance. Our solution outperforms state of the art models with 18.7% higher clinical relevance, 32.4% improved structural completeness, and 41.2% fewer diagnostic errors, demonstrating superior accuracy and clinical utility compared to existing solutions.

Rui Hu, Lianghui Zhu, Yuxuan Zhang et al.

Pixel grounding, encompassing tasks such as Referring Expression Segmentation (RES), has garnered considerable attention due to its immense potential for bridging the gap between vision and language modalities. However, advancements in this domain are currently constrained by limitations inherent in existing datasets, including limited object categories, insufficient textual diversity, and a scarcity of high-quality annotations. To mitigate these limitations, we introduce GroundingSuite, which comprises: (1) an automated data annotation framework leveraging multiple Vision-Language Model (VLM) agents; (2) a large-scale training dataset encompassing 9.56 million diverse referring expressions and their corresponding segmentations; and (3) a meticulously curated evaluation benchmark consisting of 3,800 images. The GroundingSuite training dataset facilitates substantial performance improvements, enabling models trained on it to achieve state-of-the-art results. Specifically, a cIoU of 68.9 on gRefCOCO and a gIoU of 55.3 on RefCOCOm. Moreover, the GroundingSuite annotation framework demonstrates superior efficiency compared to the current leading data annotation method, i.e., $4.5 \times$ faster than GLaMM.

Yuxuan Chen, Jiawen Li, Huijuan Shi et al.

Bone metastasis analysis is a significant challenge in pathology and plays a critical role in determining patient quality of life and treatment strategies. The microenvironment and specific tissue structures are essential for pathologists to predict the primary bone cancer origins and primary bone cancer subtyping. By digitizing bone tissue sections into whole slide images (WSIs) and leveraging deep learning to model slide embeddings, this analysis can be enhanced. However, tumor metastasis involves complex multivariate interactions with diverse bone tissue structures, which traditional WSI analysis methods such as multiple instance learning (MIL) fail to capture. Moreover, graph neural networks (GNNs), limited to modeling pairwise relationships, are hard to represent high-order biological associations. To address these challenges, we propose a dynamic hypergraph neural network (DyHG) that overcomes the edge construction limitations of traditional graph representations by connecting multiple nodes via hyperedges. A low-rank strategy is used to reduce the complexity of parameters in learning hypergraph structures, while a Gumbel-Softmax-based sampling strategy optimizes the patch distribution across hyperedges. An MIL aggregator is then used to derive a graph-level embedding for comprehensive WSI analysis. To evaluate the effectiveness of DyHG, we construct two large-scale datasets for primary bone cancer origins and subtyping classification based on real-world bone metastasis scenarios. Extensive experiments demonstrate that DyHG significantly outperforms state-of-the-art (SOTA) baselines, showcasing its ability to model complex biological interactions and improve the accuracy of bone metastasis analysis.

Gangwei Xu, Haotong Lin, Hongcheng Luo et al.

This paper presents Pixel-Perfect Depth, a monocular depth estimation model based on pixel-space diffusion generation that produces high-quality, flying-pixel-free point clouds from estimated depth maps. Current generative depth estimation models fine-tune Stable Diffusion and achieve impressive performance. However, they require a VAE to compress depth maps into latent space, which inevitably introduces \textit{flying pixels} at edges and details. Our model addresses this challenge by directly performing diffusion generation in the pixel space, avoiding VAE-induced artifacts. To overcome the high complexity associated with pixel-space generation, we introduce two novel designs: 1) Semantics-Prompted Diffusion Transformers (SP-DiT), which incorporate semantic representations from vision foundation models into DiT to prompt the diffusion process, thereby preserving global semantic consistency while enhancing fine-grained visual details; and 2) Cascade DiT Design that progressively increases the number of tokens to further enhance efficiency and accuracy. Our model achieves the best performance among all published generative models across five benchmarks, and significantly outperforms all other models in edge-aware point cloud evaluation.

Zongming Li, Lianghui Zhu, Haocheng Shen et al.

Most existing illumination-editing approaches fail to simultaneously provide customized control of light effects and preserve content integrity. This makes them less effective for practical lighting stylization requirements, especially in the challenging task of transferring complex light effects from a reference image to a user-specified target image. To address this problem, we propose TransLight, a novel framework that enables high-fidelity and high-freedom transfer of light effects. Extracting the light effect from the reference image is the most critical and challenging step in our method. The difficulty lies in the complex geometric structure features embedded in light effects that are highly coupled with content in real-world scenarios. To achieve this, we first present Generative Decoupling, where two fine-tuned diffusion models are used to accurately separate image content and light effects, generating a newly curated, million-scale dataset of image-content-light triplets. Then, we employ IC-Light as the generative model and train our model with our triplets, injecting the reference lighting image as an additional conditioning signal. The resulting TransLight model enables customized and natural transfer of diverse light effects. Notably, by thoroughly disentangling light effects from reference images, our generative decoupling strategy endows TransLight with highly flexible illumination control. Experimental results establish TransLight as the first method to successfully transfer light effects across disparate images, delivering more customized illumination control than existing techniques and charting new directions for research in illumination harmonization and editing.

Yizhi Wang, Li Chen, Qiang Huang et al.

Cervical cancer remains a major malignancy, necessitating extensive and complex histopathological assessments and comprehensive support tools. Although deep learning shows promise, these models still lack accuracy and generalizability. General foundation models offer a broader reach but remain limited in capturing subspecialty-specific features and task adaptability. We introduce the Cervical Subspecialty Pathology (CerS-Path) diagnostic system, developed through two synergistic pretraining stages: self-supervised learning on approximately 190 million tissue patches from 140,000 slides to build a cervical-specific feature extractor, and multimodal enhancement with 2.5 million image-text pairs, followed by integration with multiple downstream diagnostic functions. Supporting eight diagnostic functions, including rare cancer classification and multimodal Q&A, CerS-Path surpasses prior foundation models in scope and clinical applicability. Comprehensive evaluations demonstrate a significant advance in cervical pathology, with prospective testing on 3,173 cases across five centers maintaining 99.38% screening sensitivity and excellent generalizability, highlighting its potential for subspecialty diagnostic translation and cervical cancer screening.

Mingxi Fu, Fanglei Fu, Xitong Ling et al.

Pathological image segmentation faces numerous challenges, particularly due to ambiguous semantic boundaries and the high cost of pixel-level annotations. Although recent semi-supervised methods based on consistency regularization (e.g., UniMatch) have made notable progress, they mainly rely on perturbation-based consistency within the image modality, making it difficult to capture high-level semantic priors, especially in structurally complex pathology images. To address these limitations, we propose MPAMatch - a novel segmentation framework that performs pixel-level contrastive learning under a multimodal prototype-guided supervision paradigm. The core innovation of MPAMatch lies in the dual contrastive learning scheme between image prototypes and pixel labels, and between text prototypes and pixel labels, providing supervision at both structural and semantic levels. This coarse-to-fine supervisory strategy not only enhances the discriminative capability on unlabeled samples but also introduces the text prototype supervision into segmentation for the first time, significantly improving semantic boundary modeling. In addition, we reconstruct the classic segmentation architecture (TransUNet) by replacing its ViT backbone with a pathology-pretrained foundation model (Uni), enabling more effective extraction of pathology-relevant features. Extensive experiments on GLAS, EBHI-SEG-GLAND, EBHI-SEG-CANCER, and KPI show MPAMatch's superiority over state-of-the-art methods, validating its dual advantages in structural and semantic modeling.

Mingxi Fu, Xitong Ling, Yuxuan Chen et al.

Accurate classification of Whole Slide Images (WSIs) and Regions of Interest (ROIs) is a fundamental challenge in computational pathology. While mainstream approaches often adopt Multiple Instance Learning (MIL), they struggle to capture the spatial dependencies among tissue structures. Graph Neural Networks (GNNs) have emerged as a solution to model inter-instance relationships, yet most rely on static graph topologies and overlook the physical spatial positions of tissue patches. Moreover, conventional attention mechanisms lack specificity, limiting their ability to focus on structurally relevant regions. In this work, we propose a novel GNN framework with deformable attention for pathology image analysis. We construct a dynamic weighted directed graph based on patch features, where each node aggregates contextual information from its neighbors via attention-weighted edges. Specifically, we incorporate learnable spatial offsets informed by the real coordinates of each patch, enabling the model to adaptively attend to morphologically relevant regions across the slide. This design significantly enhances the contextual field while preserving spatial specificity. Our framework achieves state-of-the-art performance on four benchmark datasets (TCGA-COAD, BRACS, gastric intestinal metaplasia grading, and intestinal ROI classification), demonstrating the power of deformable attention in capturing complex spatial structures in WSIs and ROIs.

Lianghui Zhu, Xitong Ling, Minxi Ouyang et al.

Gastrointestinal (GI) diseases represent a clinically significant burden, necessitating precise diagnostic approaches to optimize patient outcomes. Conventional histopathological diagnosis suffers from limited reproducibility and diagnostic variability. To overcome these limitations, we develop Digepath, a specialized foundation model for GI pathology. Our framework introduces a dual-phase iterative optimization strategy combining pretraining with fine-screening, specifically designed to address the detection of sparsely distributed lesion areas in whole-slide images. Digepath is pretrained on over 353 million multi-scale images from 210,043 H&E-stained slides of GI diseases. It attains state-of-the-art performance on 33 out of 34 tasks related to GI pathology, including pathological diagnosis, protein expression status prediction, gene mutation prediction, and prognosis evaluation. We further translate the intelligent screening module for early GI cancer and achieve near-perfect 99.70% sensitivity across nine independent medical institutions. This work not only advances AI-driven precision pathology for GI diseases but also bridge critical gaps in histopathological practice.

Xitong Ling, Yifeng Ping, Jiawen Li et al.

Multiple Instance Learning (MIL) plays a significant role in computational pathology, enabling weakly supervised analysis of Whole Slide Image (WSI) datasets. The field of WSI analysis is confronted with a severe long-tailed distribution problem, which significantly impacts the performance of classifiers. Long-tailed distributions lead to class imbalance, where some classes have sparse samples while others are abundant, making it difficult for classifiers to accurately identify minority class samples. To address this issue, we propose an ensemble learning method based on MIL, which employs expert decoders with shared aggregators and consistency constraints to learn diverse distributions and reduce the impact of class imbalance on classifier performance. Moreover, we introduce a multimodal distillation framework that leverages text encoders pre-trained on pathology-text pairs to distill knowledge and guide the MIL aggregator in capturing stronger semantic features relevant to class information. To ensure flexibility, we use learnable prompts to guide the distillation process of the pre-trained text encoder, avoiding limitations imposed by specific prompts. Our method, MDE-MIL, integrates multiple expert branches focusing on specific data distributions to address long-tailed issues. Consistency control ensures generalization across classes. Multimodal distillation enhances feature extraction. Experiments on Camelyon+-LT and PANDA-LT datasets show it outperforms state-of-the-art methods.

Yuxuan Zhang, Tianheng Cheng, Lianghui Zhu et al.

Segment Anything Model (SAM) has attracted widespread attention for its superior interactive segmentation capabilities with visual prompts while lacking further exploration of text prompts. In this paper, we empirically investigate what text prompt encoders (e.g., CLIP or LLM) are good for adapting SAM for referring expression segmentation and introduce the Early Vision-language Fusion-based SAM (EVF-SAM). EVF-SAM is a simple yet effective referring segmentation method which exploits multimodal prompts (i.e., image and text) and comprises a pre-trained vision-language model to generate referring prompts and a SAM model for segmentation. Surprisingly, we observe that: (1) multimodal prompts and (2) vision-language models with early fusion (e.g., BEIT-3) are beneficial for prompting SAM for accurate referring segmentation. Our experiments show that the proposed EVF-SAM based on BEIT-3 can obtain state-of-the-art performance on RefCOCO/+/g for referring expression segmentation and demonstrate the superiority of prompting SAM with early vision-language fusion. In addition, the proposed EVF-SAM with 1.32B parameters achieves remarkably higher performance while reducing nearly 82% of parameters compared to previous SAM methods based on large multimodal models.

Xuenian Wang, Shanshan Shi, Renao Yan et al.

In the field of whole slide image (WSI) classification, multiple instance learning (MIL) serves as a promising approach, commonly decoupled into feature extraction and aggregation. In this paradigm, our observation reveals that discriminative embeddings are crucial for aggregation to the final prediction. Among all feature updating strategies, task-oriented ones can capture characteristics specifically for certain tasks. However, they can be prone to overfitting and contaminated by samples assigned with noisy labels. To address this issue, we propose a heuristic clustering-driven feature fine-tuning method (HC-FT) to enhance the performance of multiple instance learning by providing purified positive and hard negative samples. Our method first employs a well-trained MIL model to evaluate the confidence of patches. Then, patches with high confidence are marked as positive samples, while the remaining patches are used to identify crucial negative samples. After two rounds of heuristic clustering and selection, purified positive and hard negative samples are obtained to facilitate feature fine-tuning. The proposed method is evaluated on both CAMELYON16 and BRACS datasets, achieving an AUC of 97.13% and 85.85%, respectively, consistently outperforming all compared methods.