Guangzhi Wang

Bohao Li, Yuying Ge, Yixiao Ge et al. · tencent-ai

Multimodal large language models (MLLMs), building upon the foundation of powerful large language models (LLMs), have recently demonstrated exceptional capabilities in generating not only texts but also images given interleaved multimodal inputs (acting like a combination of GPT-4V and DALL-E 3). However, existing MLLM benchmarks remain limited to assessing only models' comprehension ability of single image-text inputs, failing to keep up with the strides made in MLLMs. A comprehensive benchmark is imperative for investigating the progress and uncovering the limitations of current MLLMs. In this work, we categorize the capabilities of MLLMs into hierarchical levels from $L_0$ to $L_4$ based on the modalities they can accept and generate, and propose SEED-Bench-2, a comprehensive benchmark that evaluates the \textbf{hierarchical} capabilities of MLLMs. Specifically, SEED-Bench-2 comprises 24K multiple-choice questions with accurate human annotations, which spans 27 dimensions, including the evaluation of both text and image generation. Multiple-choice questions with groundtruth options derived from human annotation enables an objective and efficient assessment of model performance, eliminating the need for human or GPT intervention during evaluation. We further evaluate the performance of 23 prominent open-source MLLMs and summarize valuable observations. By revealing the limitations of existing MLLMs through extensive evaluations, we aim for SEED-Bench-2 to provide insights that will motivate future research towards the goal of General Artificial Intelligence. Dataset and evaluation code are available at \href{https://github.com/AILab-CVC/SEED-Bench}

Bohao Li, Rui Wang, Guangzhi Wang et al. · tencent-ai

Based on powerful Large Language Models (LLMs), recent generative Multimodal Large Language Models (MLLMs) have gained prominence as a pivotal research area, exhibiting remarkable capability for both comprehension and generation. In this work, we address the evaluation of generative comprehension in MLLMs as a preliminary step towards a comprehensive assessment of generative models, by introducing a benchmark named SEED-Bench. SEED-Bench consists of 19K multiple choice questions with accurate human annotations (x 6 larger than existing benchmarks), which spans 12 evaluation dimensions including the comprehension of both the image and video modality. We develop an advanced pipeline for generating multiple-choice questions that target specific evaluation dimensions, integrating both automatic filtering and manual verification processes. Multiple-choice questions with groundtruth options derived from human annotation enables an objective and efficient assessment of model performance, eliminating the need for human or GPT intervention during evaluation. We further evaluate the performance of 18 models across all 12 dimensions, covering both the spatial and temporal understanding. By revealing the limitations of existing MLLMs through evaluation results, we aim for SEED-Bench to provide insights for motivating future research. We will launch and consistently maintain a leaderboard to provide a platform for the community to assess and investigate model capability.

Guangzhi Wang, Hehe Fan, Mohan Kankanhalli

Automatically localizing a position based on a few natural language instructions is essential for future robots to communicate and collaborate with humans. To approach this goal, we focus on the text-to-point-cloud cross-modal localization problem. Given a textual query, it aims to identify the described location from city-scale point clouds. The task involves two challenges. 1) In city-scale point clouds, similar ambient instances may exist in several locations. Searching each location in a huge point cloud with only instances as guidance may lead to less discriminative signals and incorrect results. 2) In textual descriptions, the hints are provided separately. In this case, the relations among those hints are not explicitly described, leading to difficulties of learning relations. To overcome these two challenges, we propose a unified Relation-Enhanced Transformer (RET) to improve representation discriminability for both point cloud and natural language queries. The core of the proposed RET is a novel Relation-enhanced Self-Attention (RSA) mechanism, which explicitly encodes instance (hint)-wise relations for the two modalities. Moreover, we propose a fine-grained cross-modal matching method to further refine the location predictions in a subsequent instance-hint matching stage. Experimental results on the KITTI360Pose dataset demonstrate that our approach surpasses the previous state-of-the-art method by large margin.

Guangzhi Wang, Yangyang Guo, Yongkang Wong et al.

Human-Object Interaction (HOI) detection has received considerable attention in the context of scene understanding. Despite the growing progress on benchmarks, we realize that existing methods often perform unsatisfactorily on distant interactions, where the leading causes are two-fold: 1) Distant interactions are by nature more difficult to recognize than close ones. A natural scene often involves multiple humans and objects with intricate spatial relations, making the interaction recognition for distant human-object largely affected by complex visual context. 2) Insufficient number of distant interactions in benchmark datasets results in under-fitting on these instances. To address these problems, in this paper, we propose a novel two-stage method for better handling distant interactions in HOI detection. One essential component in our method is a novel Far Near Distance Attention module. It enables information propagation between humans and objects, whereby the spatial distance is skillfully taken into consideration. Besides, we devise a novel Distance-Aware loss function which leads the model to focus more on distant yet rare interactions. We conduct extensive experiments on two challenging datasets - HICO-DET and V-COCO. The results demonstrate that the proposed method can surpass existing approaches by a large margin, resulting in new state-of-the-art performance.

Guangzhi Wang, Yangyang Guo, Yongkang Wong et al.

Detecting Human-Object Interaction (HOI) in images is an important step towards high-level visual comprehension. Existing work often shed light on improving either human and object detection, or interaction recognition. However, due to the limitation of datasets, these methods tend to fit well on frequent interactions conditioned on the detected objects, yet largely ignoring the rare ones, which is referred to as the object bias problem in this paper. In this work, we for the first time, uncover the problem from two aspects: unbalanced interaction distribution and biased model learning. To overcome the object bias problem, we propose a novel plug-and-play Object-wise Debiasing Memory (ODM) method for re-balancing the distribution of interactions under detected objects. Equipped with carefully designed read and write strategies, the proposed ODM allows rare interaction instances to be more frequently sampled for training, thereby alleviating the object bias induced by the unbalanced interaction distribution. We apply this method to three advanced baselines and conduct experiments on the HICO-DET and HOI-COCO datasets. To quantitatively study the object bias problem, we advocate a new protocol for evaluating model performance. As demonstrated in the experimental results, our method brings consistent and significant improvements over baselines, especially on rare interactions under each object. In addition, when evaluating under the conventional standard setting, our method achieves new state-of-the-art on the two benchmarks.

Yangyang Guo, Guangzhi Wang, Mohan Kankanhalli

Applying a pre-trained large model to downstream tasks is prohibitive under resource-constrained conditions. Recent dominant approaches for addressing efficiency issues involve adding a few learnable parameters to the fixed backbone model. This strategy, however, leads to more challenges in loading large models for downstream fine-tuning with limited resources. In this paper, we propose a novel method for increasing the parameter efficiency of pre-trained models by introducing an intermediate pre-training stage. To this end, we first employ low-rank approximation to compress the original large model and then devise a feature distillation module and a weight perturbation regularization module. These modules are specifically designed to enhance the low-rank model. In particular, we update only the low-rank model while freezing the backbone parameters during pre-training. This allows for direct and efficient utilization of the low-rank model for downstream fine-tuning tasks. The proposed method achieves both efficiencies in terms of required parameters and computation time while maintaining comparable results with minimal modifications to the backbone architecture. Specifically, when applied to three vision-only and one vision-language Transformer models, our approach often demonstrates a merely $\sim$0.6 point decrease in performance while reducing the original parameter size by 1/3 to 2/3.

Guangzhi Wang, Yangyang Guo, Mohan Kankanhalli

Human-Object Interaction Detection is a crucial aspect of human-centric scene understanding, with important applications in various domains. Despite recent progress in this field, recognizing subtle and detailed interactions remains challenging. Existing methods try to use human-related clues to alleviate the difficulty, but rely heavily on external annotations or knowledge, limiting their practical applicability in real-world scenarios. In this work, we propose a novel Part Semantic Network (PSN) to solve this problem. The core of PSN is a Conditional Part Attention (CPA) mechanism, where human features are taken as keys and values, and the object feature is used as query for the computation in a cross-attention mechanism. In this way, our model learns to automatically focus on the most informative human parts conditioned on the involved object, generating more semantically meaningful features for interaction recognition. Additionally, we propose an Occluded Part Extrapolation (OPE) strategy to facilitate interaction recognition under occluded scenarios, which teaches the model to extrapolate detailed features from partially occluded ones. Our method consistently outperforms prior approaches on the V-COCO and HICO-DET datasets, without external data or extra annotations. Additional ablation studies validate the effectiveness of each component of our proposed method.

Haowei Li, Wenqing Yan, Du Liu et al.

Augmented Reality (AR) has been used to facilitate surgical guidance during External Ventricular Drain (EVD) surgery, reducing the risks of misplacement in manual operations. During this procedure, the key challenge is accurately estimating the spatial relationship between pre-operative images and actual patient anatomy in AR environment. This research proposes a novel framework utilizing Time of Flight (ToF) depth sensors integrated in commercially available AR Head Mounted Devices (HMD) for precise EVD surgical guidance. As previous studies have proven depth errors for ToF sensors, we first assessed their properties on AR-HMDs. Subsequently, a depth error model and patient-specific parameter identification method are introduced for accurate surface information. A tracking pipeline combining retro-reflective markers and point clouds is then proposed for accurate head tracking. The head surface is reconstructed using depth data for spatial registration, avoiding fixing tracking targets rigidly on the patient's skull. Firstly, $7.580\pm 1.488 mm$ depth value error was revealed on human skin, indicating the significance of depth correction. Our results showed that the error was reduced by over $85\%$ using proposed depth correction method on head phantoms in different materials. Meanwhile, the head surface reconstructed with corrected depth data achieved sub-millimetre accuracy. An experiment on sheep head revealed $0.79 mm$ reconstruction error. Furthermore, a user study was conducted for the performance in simulated EVD surgery, where five surgeons performed nine k-wire injections on a head phantom with virtual guidance. Results of this study revealed $2.09 \pm 0.16 mm$ translational accuracy and $2.97\pm 0.91$ degree orientational accuracy.

Lingen Li, Guangzhi Wang, Xiaoyu Li et al.

Generating high-quality 360° panoramic videos from perspective input is one of the crucial applications for virtual reality (VR), whereby high-resolution videos are especially important for immersive experience. Existing methods are constrained by computational limitations of vanilla diffusion models, only supporting $\leq$ 1K resolution native generation and relying on suboptimal post super-resolution to increase resolution. We introduce CubeComposer, a novel spatio-temporal autoregressive diffusion model that natively generates 4K-resolution 360° videos. By decomposing videos into cubemap representations with six faces, CubeComposer autoregressively synthesizes content in a well-planned spatio-temporal order, reducing memory demands while enabling high-resolution output. Specifically, to address challenges in multi-dimensional autoregression, we propose: (1) a spatio-temporal autoregressive strategy that orchestrates 360° video generation across cube faces and time windows for coherent synthesis; (2) a cube face context management mechanism, equipped with a sparse context attention design to improve efficiency; and (3) continuity-aware techniques, including cube-aware positional encoding, padding, and blending to eliminate boundary seams. Extensive experiments on benchmark datasets demonstrate that CubeComposer outperforms state-of-the-art methods in native resolution and visual quality, supporting practical VR application scenarios. Project page: https://lg-li.github.io/project/cubecomposer

Sicheng Zhao, Guangzhi Wang, Shanghang Zhang et al.

Deep neural networks suffer from performance decay when there is domain shift between the labeled source domain and unlabeled target domain, which motivates the research on domain adaptation (DA). Conventional DA methods usually assume that the labeled data is sampled from a single source distribution. However, in practice, labeled data may be collected from multiple sources, while naive application of the single-source DA algorithms may lead to suboptimal solutions. In this paper, we propose a novel multi-source distilling domain adaptation (MDDA) network, which not only considers the different distances among multiple sources and the target, but also investigates the different similarities of the source samples to the target ones. Specifically, the proposed MDDA includes four stages: (1) pre-train the source classifiers separately using the training data from each source; (2) adversarially map the target into the feature space of each source respectively by minimizing the empirical Wasserstein distance between source and target; (3) select the source training samples that are closer to the target to fine-tune the source classifiers; and (4) classify each encoded target feature by corresponding source classifier, and aggregate different predictions using respective domain weight, which corresponds to the discrepancy between each source and target. Extensive experiments are conducted on public DA benchmarks, and the results demonstrate that the proposed MDDA significantly outperforms the state-of-the-art approaches. Our source code is released at: https://github.com/daoyuan98/MDDA.

Guangzhi Wang, Xiaohui Yang, Kai Li et al.

As retrieval models converge on generic benchmarks, the pressing question is no longer "who scores higher" but rather "where do systems fail, and why?" Person-job matching is a domain that urgently demands such diagnostic capability -- it requires systems not only to verify explicit constraints but also to perform skill-transfer inference and job-competency reasoning, yet existing benchmarks provide no systematic diagnostic support for this task. We introduce PJB (Person-Job Benchmark), a reasoning-aware retrieval evaluation dataset that uses complete job descriptions as queries and complete resumes as documents, defines relevance through job-competency judgment, is grounded in real-world recruitment data spanning six industry domains and nearly 200,000 resumes, and upgrades evaluation from "who scores higher" to "where do systems differ, and why" through domain-family and reasoning-type diagnostic labels. Diagnostic experiments using dense retrieval reveal that performance heterogeneity across industry domains far exceeds the gains from module upgrades for the same model, indicating that aggregate scores alone can severely mislead optimization decisions. At the module level, reranking yields stable improvements while query understanding not only fails to help but actually degrades overall performance when combined with reranking -- the two modules face fundamentally different improvement bottlenecks. The value of PJB lies not in yet another leaderboard of average scores, but in providing recruitment retrieval systems with a capability map that pinpoints where to invest.

Guangzhi Wang, Yinghao Jiao, Zhi Liu

The central challenge of reasoning-intensive retrieval lies in identifying implicitreasoning relationships between queries and documents, rather than superficial se-mantic or lexical similarity. The contrastive learning paradigm is fundamentallya static representation consolidation technique: during training, it encodes hier-archical relevance concepts into fixed geometric structures in the vector space,and at inference time it cannot dynamically adjust relevance judgments accord-ing to the specific reasoning demands of each query. Consequently, performancedegrades noticeably when vocabulary mismatch exists between queries and doc-uments or when implicit reasoning is required to establish relevance. This pa-per proposes Thought 1 (T1), a generative retrieval model that shifts relevancemodeling from static alignment to dynamic reasoning. On the query side, T1 dy-namically generates intermediate reasoning trajectories for each query to bridgeimplicit reasoning relationships and uses <embtoken> as a semantic aggregationpoint for the reasoning output. On the document side, it employs an instruction+ text + <embtoken> encoding format to support high-throughput indexing. Tointernalize dynamic reasoning capabilities into vector representations, we adopt athree-stage training curriculum and introduce GRPO in the third stage, enablingthe model to learn optimal derivation strategies for different queries through trial-and-error reinforcement learning. On the BRIGHT benchmark, T1-4B exhibitsstrong performance under the original query setting, outperforming larger modelstrained with contrastive learning overall, and achieving performance comparableto multi-stage retrieval pipelines. The results demonstrate that replacing static rep-resentation alignment with dynamic reasoning generation can effectively improvereasoning-intensive retrieval performance.

Haowei Li, Wenqing Yan, Jiasheng Zhao et al.

Purpose: Biopsies play a crucial role in determining the classification and staging of tumors. Ultrasound is frequently used in this procedure to provide real-time anatomical information. Using augmented reality (AR), surgeons can visualize ultrasound data and spatial navigation information seamlessly integrated with real tissues. This innovation facilitates faster and more precise biopsy operations. Methods: We developed an AR biopsy navigation system with low display latency and high accuracy. Ultrasound data is initially read by an image capture card and streamed to Unity via net communication. In Unity, navigation information is rendered and transmitted to the HoloLens 2 device using holographic remoting. Retro-reflective tool tracking is implemented on the HoloLens 2, enabling simultaneous tracking of the ultrasound probe and biopsy needle. Distinct navigation information is provided during in-plane and out-of-plane punctuation. To evaluate the effectiveness of our system, we conducted a study involving ten participants, for puncture accuracy and biopsy time, comparing to traditional methods. Results: Our proposed framework enables ultrasound visualization in AR with only $16.22\pm11.45ms$ additional latency. Navigation accuracy reached $1.23\pm 0.68mm$ in the image plane and $0.95\pm 0.70mm$ outside the image plane. Remarkably, the utilization of our system led to $98\%$ and $95\%$ success rate in out-of-plane and in-plane biopsy. Conclusion: To sum up, this paper introduces an AR-based ultrasound biopsy navigation system characterized by high navigation accuracy and minimal latency. The system provides distinct visualization contents during in-plane and out-of-plane operations according to their different characteristics. Use case study in this paper proved that our system can help young surgeons perform biopsy faster and more accurately.

Shuzhou Yang, Xiaoyu Li, Xiaodong Cun et al.

Video compositing combines live-action footage to create video production, serving as a crucial technique in video creation and film production. Traditional pipelines require intensive labor efforts and expert collaboration, resulting in lengthy production cycles and high manpower costs. To address this issue, we automate this process with generative models, called generative video compositing. This new task strives to adaptively inject identity and motion information of foreground video to the target video in an interactive manner, allowing users to customize the size, motion trajectory, and other attributes of the dynamic elements added in final video. Specifically, we designed a novel Diffusion Transformer (DiT) pipeline based on its intrinsic properties. To maintain consistency of the target video before and after editing, we revised a light-weight DiT-based background preservation branch with masked token injection. As to inherit dynamic elements from other sources, a DiT fusion block is proposed using full self-attention, along with a simple yet effective foreground augmentation for training. Besides, for fusing background and foreground videos with different layouts based on user control, we developed a novel position embedding, named Extended Rotary Position Embedding (ERoPE). Finally, we curated a dataset comprising 61K sets of videos for our new task, called VideoComp. This data includes complete dynamic elements and high-quality target videos. Experiments demonstrate that our method effectively realizes generative video compositing, outperforming existing possible solutions in fidelity and consistency.

Yaowei Li, Lingen Li, Zhaoyang Zhang et al.

As user expectations for image editing continue to rise, the demand for flexible, fine-grained manipulation of specific visual elements presents a challenge for current diffusion-based methods. In this work, we present BlobCtrl, a framework for element-level image editing based on a probabilistic blob-based representation. Treating blobs as visual primitives, BlobCtrl disentangles layout from appearance, affording fine-grained, controllable object-level manipulation. Our key contributions are twofold: (1) an in-context dual-branch diffusion model that separates foreground and background processing, incorporating blob representations to explicitly decouple layout and appearance, and (2) a self-supervised disentangle-then-reconstruct training paradigm with an identity-preserving loss function, along with tailored strategies to efficiently leverage blob-image pairs. To foster further research, we introduce BlobData for large-scale training and BlobBench, a benchmark for systematic evaluation. Experimental results demonstrate that BlobCtrl achieves state-of-the-art performance in a variety of element-level editing tasks, such as object addition, removal, scaling, and replacement, while maintaining computational efficiency. Project Webpage: https://liyaowei-stu.github.io/project/BlobCtrl/

Guangzhi Wang, Tianyi Chen, Kamran Ghasedi et al.

Face attribute editing plays a pivotal role in various applications. However, existing methods encounter challenges in achieving high-quality results while preserving identity, editing faithfulness, and temporal consistency. These challenges are rooted in issues related to the training pipeline, including limited supervision, architecture design, and optimization strategy. In this work, we introduce S3Editor, a Sparse Semantic-disentangled Self-training framework for face video editing. S3Editor is a generic solution that comprehensively addresses these challenges with three key contributions. Firstly, S3Editor adopts a self-training paradigm to enhance the training process through semi-supervision. Secondly, we propose a semantic disentangled architecture with a dynamic routing mechanism that accommodates diverse editing requirements. Thirdly, we present a structured sparse optimization schema that identifies and deactivates malicious neurons to further disentangle impacts from untarget attributes. S3Editor is model-agnostic and compatible with various editing approaches. Our extensive qualitative and quantitative results affirm that our approach significantly enhances identity preservation, editing fidelity, as well as temporal consistency.

Guangzhi Wang, Kai Li, Yinghao Jiao et al.

We propose RT (Refine Thought), a method that can enhance the semantic rea-soning ability of text embedding models. The method obtains the final semanticrepresentation by running multiple forward passes of the text embedding model.Experiments show that RT achieves significant improvements on semantic reason-ing tasks in BRIGHT and the person job matching benchmark PJBenchmark1, while maintaining consistent performance on general-purpose semantic under-standing tasks such as C-MTEB. Our results indicate that RT is effective becauseit further activates the semantic reasoning ability learned during pretraining bydecoder-only text embedding models(e.g., Qwen3-Embedding-8B). RT canbe seen as a test-time inference method.

Lingen Li, Guangzhi Wang, Zhaoyang Zhang et al.

Traditional cartoon and anime production involves keyframing, inbetweening, and colorization stages, which require intensive manual effort. Despite recent advances in AI, existing methods often handle these stages separately, leading to error accumulation and artifacts. For instance, inbetweening approaches struggle with large motions, while colorization methods require dense per-frame sketches. To address this, we introduce ToonComposer, a generative model that unifies inbetweening and colorization into a single post-keyframing stage. ToonComposer employs a sparse sketch injection mechanism to provide precise control using keyframe sketches. Additionally, it uses a cartoon adaptation method with the spatial low-rank adapter to tailor a modern video foundation model to the cartoon domain while keeping its temporal prior intact. Requiring as few as a single sketch and a colored reference frame, ToonComposer excels with sparse inputs, while also supporting multiple sketches at any temporal location for more precise motion control. This dual capability reduces manual workload and improves flexibility, empowering artists in real-world scenarios. To evaluate our model, we further created PKBench, a benchmark featuring human-drawn sketches that simulate real-world use cases. Our evaluation demonstrates that ToonComposer outperforms existing methods in visual quality, motion consistency, and production efficiency, offering a superior and more flexible solution for AI-assisted cartoon production.

Guangzhi Wang, Yixiao Ge, Xiaohan Ding et al.

We empirically investigate proper pre-training methods to build good visual tokenizers, making Large Language Models (LLMs) powerful Multimodal Large Language Models (MLLMs). In our benchmark, which is curated to evaluate MLLMs visual semantic understanding and fine-grained perception capabilities, we discussed different visual tokenizers pre-trained with dominant methods (i.e., DeiT, CLIP, MAE, DINO), and observe that: i) Fully/weakly supervised models capture more semantics than self-supervised models, but the gap is narrowed by scaling up the pre-training dataset. ii) Self-supervised models are better at fine-grained perception, where patch-level supervision is particularly effective. iii) Tuning the visual tokenizer leads to the loss of semantics obtained from large-scale pretraining, which is unfavorable with relatively small-scale instruction-tuning dataset. Given the findings, we reviewed methods that attempted to unify semantics and fine-grained visual understanding, e.g., patch-level feature distillation with semantically-rich targets. We obtain an intriguing insight mask-based strategies that were once all the rage may not be applicable for obtaining good visual tokenizers. Based on this critical observation, we obtain a new MLLM equipped with a tailored Good Visual Tokenizer (GVT), which exhibits strong visual comprehension capability at multiple scales. In particular, without introducing extra parameters and task-specific fine-tuning, GVT achieves superior performance on visual question answering, image captioning, and other fine-grained visual understanding tasks such as object counting and multi-class identification.

Ziwei Xu, Guangzhi Wang, Yongkang Wong et al.

This paper proposes a novel model for recognizing images with composite attribute-object concepts, notably for composite concepts that are unseen during model training. We aim to explore the three key properties required by the task --- relation-aware, consistent, and decoupled --- to learn rich and robust features for primitive concepts that compose attribute-object pairs. To this end, we propose the Blocked Message Passing Network (BMP-Net). The model consists of two modules. The concept module generates semantically meaningful features for primitive concepts, whereas the visual module extracts visual features for attributes and objects from input images. A message passing mechanism is used in the concept module to capture the relations between primitive concepts. Furthermore, to prevent the model from being biased towards seen composite concepts and reduce the entanglement between attributes and objects, we propose a blocking mechanism that equalizes the information available to the model for both seen and unseen concepts. Extensive experiments and ablation studies on two benchmarks show the efficacy of the proposed model.