Yexin Liu

Xu Zheng, Yexin Liu, Yunfan Lu et al.

Event cameras are bio-inspired sensors that capture the per-pixel intensity changes asynchronously and produce event streams encoding the time, pixel position, and polarity (sign) of the intensity changes. Event cameras possess a myriad of advantages over canonical frame-based cameras, such as high temporal resolution, high dynamic range, low latency, etc. Being capable of capturing information in challenging visual conditions, event cameras have the potential to overcome the limitations of frame-based cameras in the computer vision and robotics community. In very recent years, deep learning (DL) has been brought to this emerging field and inspired active research endeavors in mining its potential. However, there is still a lack of taxonomies in DL techniques for event-based vision. We first scrutinize the typical event representations with quality enhancement methods as they play a pivotal role as inputs to the DL models. We then provide a comprehensive survey of existing DL-based methods by structurally grouping them into two major categories: 1) image/video reconstruction and restoration; 2) event-based scene understanding and 3D vision. We conduct benchmark experiments for the existing methods in some representative research directions, i.e., image reconstruction, deblurring, and object recognition, to identify some critical insights and problems. Finally, we have discussions regarding the challenges and provide new perspectives for inspiring more research studies.

Xu Zheng, Jinjing Zhu, Yexin Liu et al.

The ability of scene understanding has sparked active research for panoramic image semantic segmentation. However, the performance is hampered by distortion of the equirectangular projection (ERP) and a lack of pixel-wise annotations. For this reason, some works treat the ERP and pinhole images equally and transfer knowledge from the pinhole to ERP images via unsupervised domain adaptation (UDA). However, they fail to handle the domain gaps caused by: 1) the inherent differences between camera sensors and captured scenes; 2) the distinct image formats (e.g., ERP and pinhole images). In this paper, we propose a novel yet flexible dual-path UDA framework, DPPASS, taking ERP and tangent projection (TP) images as inputs. To reduce the domain gaps, we propose cross-projection and intra-projection training. The cross-projection training includes tangent-wise feature contrastive training and prediction consistency training. That is, the former formulates the features with the same projection locations as positive examples and vice versa, for the models' awareness of distortion, while the latter ensures the consistency of cross-model predictions between the ERP and TP. Moreover, adversarial intra-projection training is proposed to reduce the inherent gap, between the features of the pinhole images and those of the ERP and TP images, respectively. Importantly, the TP path can be freely removed after training, leading to no additional inference cost. Extensive experiments on two benchmarks show that our DPPASS achieves +1.06$\%$ mIoU increment than the state-of-the-art approaches.

Yexin Liu, Weiming Zhang, Guoyang Zhao et al.

The ability to scene understanding in adverse visual conditions, e.g., nighttime, has sparked active research for RGB-Thermal (RGB-T) semantic segmentation. However, it is essentially hampered by two critical problems: 1) the day-night gap of RGB images is larger than that of thermal images, and 2) the class-wise performance of RGB images at night is not consistently higher or lower than that of thermal images. we propose the first test-time adaptation (TTA) framework, dubbed Night-TTA, to address the problems for nighttime RGBT semantic segmentation without access to the source (daytime) data during adaptation. Our method enjoys three key technical parts. Firstly, as one modality (e.g., RGB) suffers from a larger domain gap than that of the other (e.g., thermal), Imaging Heterogeneity Refinement (IHR) employs an interaction branch on the basis of RGB and thermal branches to prevent cross-modal discrepancy and performance degradation. Then, Class Aware Refinement (CAR) is introduced to obtain reliable ensemble logits based on pixel-level distribution aggregation of the three branches. In addition, we also design a specific learning scheme for our TTA framework, which enables the ensemble logits and three student logits to collaboratively learn to improve the quality of predictions during the testing phase of our Night TTA. Extensive experiments show that our method achieves state-of-the-art (SoTA) performance with a 13.07% boost in mIoU.

Weiming Zhang, Yexin Liu, Xu Zheng et al.

This paper presents GoodSAM++, a novel framework utilizing the powerful zero-shot instance segmentation capability of SAM (i.e., teacher) to learn a compact panoramic semantic segmentation model, i.e., student, without requiring any labeled data. GoodSAM++ addresses two critical challenges: 1) SAM's inability to provide semantic labels and inherent distortion problems of panoramic images; 2) the significant capacity disparity between SAM and the student. The `out-of-the-box' insight of GoodSAM++ is to introduce a teacher assistant (TA) to provide semantic information for SAM, integrated with SAM to obtain reliable pseudo semantic maps to bridge both domain and capacity gaps. To make this possible, we first propose a Distortion-Aware Rectification (DARv2) module to address the domain gap. It effectively mitigates the object deformation and distortion problem in panoramic images to obtain pseudo semantic maps. We then introduce a Multi-level Knowledge Adaptation (MKA) module to efficiently transfer the semantic information from the TA and pseudo semantic maps to our compact student model, addressing the significant capacity gap. We conduct extensive experiments on both outdoor and indoor benchmark datasets, showing that our GoodSAM++ achieves a remarkable performance improvement over the state-of-the-art (SOTA) domain adaptation methods. Moreover, diverse open-world scenarios demonstrate the generalization capacity of our GoodSAM++. Last but not least, our most lightweight student model achieves comparable performance to the SOTA models with only 3.7 million parameters.

Wen-Jie Shu, Xuerui Qiu, Rui-Jie Zhu et al.

Recent advances in visual reasoning have leveraged vision transformers to tackle the ARC-AGI benchmark. However, we argue that the feed-forward architecture, where computational depth is strictly bound to parameter size, falls short of capturing the iterative, algorithmic nature of human induction. In this work, we propose a recursive architecture called Loop-ViT, which decouples reasoning depth from model capacity through weight-tied recurrence. Loop-ViT iterates a weight-tied Hybrid Block, combining local convolutions and global attention, to form a latent chain of thought. Crucially, we introduce a parameter-free Dynamic Exit mechanism based on predictive entropy: the model halts inference when its internal state ``crystallizes" into a low-uncertainty attractor. Empirical results on the ARC-AGI-1 benchmark validate this perspective: our 18M model achieves 65.8% accuracy, outperforming massive 73M-parameter ensembles. These findings demonstrate that adaptive iterative computation offers a far more efficient scaling axis for visual reasoning than simply increasing network width. The code is available at https://github.com/WenjieShu/LoopViT.

Muyang He, Yexin Liu, Boya Wu et al.

Multimodal Large Language Models (MLLMs) have demonstrated notable capabilities in general visual understanding and reasoning tasks. However, their deployment is hindered by substantial computational costs in both training and inference, limiting accessibility to the broader research and user communities. A straightforward solution is to leverage smaller pre-trained vision and language models, which inevitably cause significant performance drops. In this paper, we demonstrate the possibility of training a smaller but better MLLM with high-quality training data. Specifically, we introduce Bunny, a family of lightweight MLLMs with flexible vision and language backbones for efficient multimodal learning from selected training data. Experiments show that our Bunny-4B/8B outperforms the state-of-the-art large MLLMs on multiple benchmarks. We expect that this work can provide the community with a clean and flexible open-source tool for further research and development. The code, models, and data can be found in https://github.com/BAAI-DCAI/Bunny.

Chenyuan Wu, Pengfei Zheng, Ruiran Yan et al.

In this work, we introduce OmniGen2, a versatile and open-source generative model designed to provide a unified solution for diverse generation tasks, including text-to-image, image editing, and in-context generation. Unlike OmniGen v1, OmniGen2 features two distinct decoding pathways for text and image modalities, utilizing unshared parameters and a decoupled image tokenizer. This design enables OmniGen2 to build upon existing multimodal understanding models without the need to re-adapt VAE inputs, thereby preserving the original text generation capabilities. To facilitate the training of OmniGen2, we developed comprehensive data construction pipelines, encompassing image editing and in-context generation data. Additionally, we introduce a reflection mechanism tailored for image generation tasks and curate a dedicated reflection dataset based on OmniGen2. Despite its relatively modest parameter size, OmniGen2 achieves competitive results on multiple task benchmarks, including text-to-image and image editing. To further evaluate in-context generation, also referred to as subject-driven tasks, we introduce a new benchmark named OmniContext. OmniGen2 achieves state-of-the-art performance among open-source models in terms of consistency. We will release our models, training code, datasets, and data construction pipeline to support future research in this field. Project Page: https://vectorspacelab.github.io/OmniGen2; GitHub Link: https://github.com/VectorSpaceLab/OmniGen2

Yizhang Jin, Jian Li, Yexin Liu et al.

In the past year, Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance in tasks such as visual question answering, visual understanding and reasoning. However, the extensive model size and high training and inference costs have hindered the widespread application of MLLMs in academia and industry. Thus, studying efficient and lightweight MLLMs has enormous potential, especially in edge computing scenarios. In this survey, we provide a comprehensive and systematic review of the current state of efficient MLLMs. Specifically, we summarize the timeline of representative efficient MLLMs, research state of efficient structures and strategies, and the applications. Finally, we discuss the limitations of current efficient MLLM research and promising future directions. Please refer to our GitHub repository for more details: https://github.com/lijiannuist/Efficient-Multimodal-LLMs-Survey.

Xianfeng Wu, Yajing Bai, Haoze Zheng et al.

Recent advances in text-to-image generation have primarily relied on extensive datasets and parameter-heavy architectures. These requirements severely limit accessibility for researchers and practitioners who lack substantial computational resources. In this paper, we introduce \model, an efficient training paradigm for image generation models that uses knowledge distillation (KD) and Direct Preference Optimization (DPO). Drawing inspiration from the success of data KD techniques widely adopted in Multi-Modal Large Language Models (MLLMs), LightGen distills knowledge from state-of-the-art (SOTA) text-to-image models into a compact Masked Autoregressive (MAR) architecture with only $0.7B$ parameters. Using a compact synthetic dataset of just $2M$ high-quality images generated from varied captions, we demonstrate that data diversity significantly outweighs data volume in determining model performance. This strategy dramatically reduces computational demands and reduces pre-training time from potentially thousands of GPU-days to merely 88 GPU-days. Furthermore, to address the inherent shortcomings of synthetic data, particularly poor high-frequency details and spatial inaccuracies, we integrate the DPO technique that refines image fidelity and positional accuracy. Comprehensive experiments confirm that LightGen achieves image generation quality comparable to SOTA models while significantly reducing computational resources and expanding accessibility for resource-constrained environments. Code is available at https://github.com/XianfengWu01/LightGen

Yexin Liu, Wen-Jie Shu, Zile Huang et al.

Text-guided image-to-video (TI2V) generation has recently achieved remarkable progress, particularly in maintaining subject consistency and temporal coherence. However, existing methods still struggle to adhere to fine-grained prompt semantics, especially when prompts entail substantial transformations of the input image (e.g., object addition, deletion, or modification), a shortcoming we term semantic negligence. In a pilot study, we find that applying a Gaussian blur to the input image improves semantic adherence. Analyzing attention maps, we observe clearer foreground-background separation. From an energy perspective, this corresponds to a lower-entropy cross-attention distribution. Motivated by this, we introduce AlignVid, a training-free framework with two components: (i) Attention Scaling Modulation (ASM), which directly reweights attention via lightweight Q or K scaling, and (ii) Guidance Scheduling (GS), which applies ASM selectively across transformer blocks and denoising steps to reduce visual quality degradation. This minimal intervention improves prompt adherence while limiting aesthetic degradation. In addition, we introduce OmitI2V to evaluate semantic negligence in TI2V generation, comprising 367 human-annotated samples that span addition, deletion, and modification scenarios. Extensive experiments demonstrate that AlignVid can enhance semantic fidelity.

Shunian Chen, Hejin Huang, Yexin Liu et al.

Audio-driven talking head synthesis has achieved remarkable photorealism, yet state-of-the-art (SOTA) models exhibit a critical failure: they lack generalization to the full spectrum of human diversity in ethnicity, language, and age groups. We argue that this generalization gap is a direct symptom of limitations in existing training data, which lack the necessary scale, quality, and diversity. To address this challenge, we introduce TalkVid, a new large-scale, high-quality, and diverse dataset containing 1244 hours of video from 7729 unique speakers. TalkVid is curated through a principled, multi-stage automated pipeline that rigorously filters for motion stability, aesthetic quality, and facial detail, and is validated against human judgments to ensure its reliability. Furthermore, we construct and release TalkVid-Bench, a stratified evaluation set of 500 clips meticulously balanced across key demographic and linguistic axes. Our experiments demonstrate that a model trained on TalkVid outperforms counterparts trained on previous datasets, exhibiting superior cross-dataset generalization. Crucially, our analysis on TalkVid-Bench reveals performance disparities across subgroups that are obscured by traditional aggregate metrics, underscoring its necessity for future research. Code and data can be found in https://github.com/FreedomIntelligence/TalkVid

Haoze Zheng, Zihao Wang, Xianfeng Wu et al.

Single-image relighting is highly under-constrained: small illumination changes can produce large, nonlinear variations in shading, shadows, and specularities, while geometry and materials remain unobserved. Existing diffusion-based approaches either rely on intrinsic or G-buffer pipelines that require dense and fragile supervision, or operate purely in latent space without physical grounding, making fine-grained control of direction, intensity, and color unreliable. We observe that a full intrinsic decomposition is unnecessary and redundant for accurate relighting. Instead, sparse but physically meaningful cues, indicating where illumination should change and how materials should respond, are sufficient to guide a diffusion model. Based on this insight, we introduce LightCtrl that integrates physical priors at two levels: a few-shot latent proxy encoder that extracts compact material-geometry cues from limited PBR supervision, and a lighting-aware mask that identifies sensitive illumination regions and steers the denoiser toward shading relevant pixels. To compensate for scarce PBR data, we refine the proxy branch using a DPO-based objective that enforces physical consistency in the predicted cues. We also present ScaLight, a large-scale object-level dataset with systematically varied illumination and complete camera-light metadata, enabling physically consistent and controllable training. Across object and scene level benchmarks, our method achieves photometrically faithful relighting with accurate continuous control, surpassing prior diffusion and intrinsic-based baselines, including gains of up to +2.4 dB PSNR and 35% lower RMSE under controlled lighting shifts.

Yexin Liu, Manyuan Zhang, Yueze Wang et al.

Despite the promising progress in subject-driven image generation, current models often deviate from the reference identities and struggle in complex scenes with multiple subjects. To address this challenge, we introduce OpenSubject, a video-derived large-scale corpus with 2.5M samples and 4.35M images for subject-driven generation and manipulation. The dataset is built with a four-stage pipeline that exploits cross-frame identity priors. (i) Video Curation. We apply resolution and aesthetic filtering to obtain high-quality clips. (ii) Cross-Frame Subject Mining and Pairing. We utilize vision-language model (VLM)-based category consensus, local grounding, and diversity-aware pairing to select image pairs. (iii) Identity-Preserving Reference Image Synthesis. We introduce segmentation map-guided outpainting to synthesize the input images for subject-driven generation and box-guided inpainting to generate input images for subject-driven manipulation, together with geometry-aware augmentations and irregular boundary erosion. (iv) Verification and Captioning. We utilize a VLM to validate synthesized samples, re-synthesize failed samples based on stage (iii), and then construct short and long captions. In addition, we introduce a benchmark covering subject-driven generation and manipulation, and then evaluate identity fidelity, prompt adherence, manipulation consistency, and background consistency with a VLM judge. Extensive experiments show that training with OpenSubject improves generation and manipulation performance, particularly in complex scenes.

Ruibin Min, Yexin Liu, Aimin Pan et al.

Representation alignment has recently emerged as an effective paradigm for accelerating Diffusion Transformer training. Despite their success, existing alignment methods typically impose a fixed supervision target or a fixed alignment granularity throughout the entire denoising trajectory, whether the guidance is provided by external vision encoders, internal self-representations, or VAE-derived features. We argue that such timestep-agnostic alignment is suboptimal because the useful granularity of representation supervision changes systematically with the signal-to-noise ratio. In high-noise regimes, diffusion models benefit more from coarse semantic and layout-level anchoring, whereas in low-noise regimes, the training signal should emphasize spatially detailed and structurally faithful refinement. This non-stationary alignment behavior creates a representational mismatch for static single-level supervisors. To address this issue, we propose Adaptive Hierarchical Prior Alignment (AHPA), a lightweight alignment framework that exploits the hierarchical representations naturally embedded in the frozen VAE encoder. Instead of using only a single compressed latent as the alignment target, AHPA extracts multi-level VAE features that provide complementary priors ranging from local geometry and spatial topology to coarse semantic layout. A timestep-conditioned Dynamic Router adaptively selects and weights these hierarchical priors along the denoising trajectory, thereby synchronizing the alignment granularity with the model's evolving training needs. Extensive experiments show that AHPA improves convergence and generation quality over baselines and incurs no additional inference cost while avoiding external encoder supervision during training.

Danli Shi, Weiyi Zhang, Xiaolan Chen et al.

Artificial intelligence (AI) is vital in ophthalmology, tackling tasks like diagnosis, classification, and visual question answering (VQA). However, existing AI models in this domain often require extensive annotation and are task-specific, limiting their clinical utility. While recent developments have brought about foundation models for ophthalmology, they are limited by the need to train separate weights for each imaging modality, preventing a comprehensive representation of multi-modal features. This highlights the need for versatile foundation models capable of handling various tasks and modalities in ophthalmology. To address this gap, we present EyeFound, a multimodal foundation model for ophthalmic images. Unlike existing models, EyeFound learns generalizable representations from unlabeled multimodal retinal images, enabling efficient model adaptation across multiple applications. Trained on 2.78 million images from 227 hospitals across 11 ophthalmic modalities, EyeFound facilitates generalist representations and diverse multimodal downstream tasks, even for detecting challenging rare diseases. It outperforms previous work RETFound in diagnosing eye diseases, predicting systemic disease incidents, and zero-shot multimodal VQA. EyeFound provides a generalizable solution to improve model performance and lessen the annotation burden on experts, facilitating widespread clinical AI applications for retinal imaging.

Weiming Zhang, Yexin Liu, Xu Zheng et al.

This paper tackles a novel yet challenging problem: how to transfer knowledge from the emerging Segment Anything Model (SAM) -- which reveals impressive zero-shot instance segmentation capacity -- to learn a compact panoramic semantic segmentation model, i.e., student, without requiring any labeled data. This poses considerable challenges due to SAM's inability to provide semantic labels and the large capacity gap between SAM and the student. To this end, we propose a novel framework, called GoodSAM, that introduces a teacher assistant (TA) to provide semantic information, integrated with SAM to generate ensemble logits to achieve knowledge transfer. Specifically, we propose a Distortion-Aware Rectification (DAR) module that first addresses the distortion problem of panoramic images by imposing prediction-level consistency and boundary enhancement. This subtly enhances TA's prediction capacity on panoramic images. DAR then incorporates a cross-task complementary fusion block to adaptively merge the predictions of SAM and TA to obtain more reliable ensemble logits. Moreover, we introduce a Multi-level Knowledge Adaptation (MKA) module to efficiently transfer the multi-level feature knowledge from TA and ensemble logits to learn a compact student model. Extensive experiments on two benchmarks show that our GoodSAM achieves a remarkable +3.75\% mIoU improvement over the state-of-the-art (SOTA) domain adaptation methods. Also, our most lightweight model achieves comparable performance to the SOTA methods with only 3.7M parameters.

Haochen Xue, Feilong Tang, Ming Hu et al.

Recent multimodal large language models (MLLMs) have demonstrated significant potential in open-ended conversation, generating more accurate and personalized responses. However, their abilities to memorize, recall, and reason in sustained interactions within real-world scenarios remain underexplored. This paper introduces MMRC, a Multi-Modal Real-world Conversation benchmark for evaluating six core open-ended abilities of MLLMs: information extraction, multi-turn reasoning, information update, image management, memory recall, and answer refusal. With data collected from real-world scenarios, MMRC comprises 5,120 conversations and 28,720 corresponding manually labeled questions, posing a significant challenge to existing MLLMs. Evaluations on 20 MLLMs in MMRC indicate an accuracy drop during open-ended interactions. We identify four common failure patterns: long-term memory degradation, inadequacies in updating factual knowledge, accumulated assumption of error propagation, and reluctance to say no. To mitigate these issues, we propose a simple yet effective NOTE-TAKING strategy, which can record key information from the conversation and remind the model during its responses, enhancing conversational capabilities. Experiments across six MLLMs demonstrate significant performance improvements.

Xiaolan Chen, Jiayang Xiang, Shanfu Lu et al.

Large language models (LLMs) have emerged as powerful tools with transformative potential across numerous domains, including healthcare and medicine. In the medical domain, LLMs hold promise for tasks ranging from clinical decision support to patient education. However, evaluating the performance of LLMs in medical contexts presents unique challenges due to the complex and critical nature of medical information. This paper provides a comprehensive overview of the landscape of medical LLM evaluation, synthesizing insights from existing studies and highlighting evaluation data sources, task scenarios, and evaluation methods. Additionally, it identifies key challenges and opportunities in medical LLM evaluation, emphasizing the need for continued research and innovation to ensure the responsible integration of LLMs into clinical practice.

Mingzhe Zheng, Yongqi Xu, Haojian Huang et al.

Current video generation models excel at short clips but fail to produce cohesive multi-shot narratives due to disjointed visual dynamics and fractured storylines. Existing solutions either rely on extensive manual scripting/editing or prioritize single-shot fidelity over cross-scene continuity, limiting their practicality for movie-like content. We introduce VideoGen-of-Thought (VGoT), a step-by-step framework that automates multi-shot video synthesis from a single sentence by systematically addressing three core challenges: (1) Narrative Fragmentation: Existing methods lack structured storytelling. We propose dynamic storyline modeling, which first converts the user prompt into concise shot descriptions, then elaborates them into detailed, cinematic specifications across five domains (character dynamics, background continuity, relationship evolution, camera movements, HDR lighting), ensuring logical narrative progression with self-validation. (2) Visual Inconsistency: Existing approaches struggle with maintaining visual consistency across shots. Our identity-aware cross-shot propagation generates identity-preserving portrait (IPP) tokens that maintain character fidelity while allowing trait variations (expressions, aging) dictated by the storyline. (3) Transition Artifacts: Abrupt shot changes disrupt immersion. Our adjacent latent transition mechanisms implement boundary-aware reset strategies that process adjacent shots' features at transition points, enabling seamless visual flow while preserving narrative continuity. VGoT generates multi-shot videos that outperform state-of-the-art baselines by 20.4% in within-shot face consistency and 17.4% in style consistency, while achieving over 100% better cross-shot consistency and 10x fewer manual adjustments than alternatives.

Mingzhe Zheng, Yongqi Xu, Haojian Huang et al.

Current video generation models excel at short clips but fail to produce cohesive multi-shot narratives due to disjointed visual dynamics and fractured storylines. Existing solutions either rely on extensive manual scripting/editing or prioritize single-shot fidelity over cross-scene continuity, limiting their practicality for movie-like content. We introduce VideoGen-of-Thought (VGoT), a step-by-step framework that automates multi-shot video synthesis from a single sentence by systematically addressing three core challenges: (1) Narrative fragmentation: Existing methods lack structured storytelling. We propose dynamic storyline modeling, which turns the user prompt into concise shot drafts and then expands them into detailed specifications across five domains (character dynamics, background continuity, relationship evolution, camera movements, and HDR lighting) with self-validation to ensure logical progress. (2) Visual inconsistency: previous approaches struggle to maintain consistent appearance across shots. Our identity-aware cross-shot propagation builds identity-preserving portrait (IPP) tokens that keep character identity while allowing controlled trait changes (expressions, aging) required by the story. (3) Transition artifacts: Abrupt shot changes disrupt immersion. Our adjacent latent transition mechanisms implement boundary-aware reset strategies that process adjacent shots' features at transition points, enabling seamless visual flow while preserving narrative continuity. Combined in a training-free pipeline, VGoT surpasses strong baselines by 20.4\% in within-shot face consistency and 17.4\% in style consistency, while requiring 10x fewer manual adjustments. VGoT bridges the gap between raw visual synthesis and director-level storytelling for automated multi-shot video generation.

Bin Cao, Jianhao Yuan, Yexin Liu et al.

In the rapidly evolving area of image synthesis, a serious challenge is the presence of complex artifacts that compromise perceptual realism of synthetic images. To alleviate artifacts and improve quality of synthetic images, we fine-tune Vision-Language Model (VLM) as artifact classifier to automatically identify and classify a wide range of artifacts and provide supervision for further optimizing generative models. Specifically, we develop a comprehensive artifact taxonomy and construct a dataset of synthetic images with artifact annotations for fine-tuning VLM, named SynArtifact-1K. The fine-tuned VLM exhibits superior ability of identifying artifacts and outperforms the baseline by 25.66%. To our knowledge, this is the first time such end-to-end artifact classification task and solution have been proposed. Finally, we leverage the output of VLM as feedback to refine the generative model for alleviating artifacts. Visualization results and user study demonstrate that the quality of images synthesized by the refined diffusion model has been obviously improved.

Harold Haodong Chen, Haojian Huang, Xianfeng Wu et al.

Temporal quality is a critical aspect of video generation, as it ensures consistent motion and realistic dynamics across frames. However, achieving high temporal coherence and diversity remains challenging. In this work, we explore temporal augmentation in video generation for the first time, and introduce FluxFlow for initial investigation, a strategy designed to enhance temporal quality. Operating at the data level, FluxFlow applies controlled temporal perturbations without requiring architectural modifications. Extensive experiments on UCF-101 and VBench benchmarks demonstrate that FluxFlow significantly improves temporal coherence and diversity across various video generation models, including U-Net, DiT, and AR-based architectures, while preserving spatial fidelity. These findings highlight the potential of temporal augmentation as a simple yet effective approach to advancing video generation quality.

Haolin Yang, Feilong Tang, Ming Hu et al.

Video diffusion models (VDMs) facilitate the generation of high-quality videos, with current research predominantly concentrated on scaling efforts during training through improvements in data quality, computational resources, and model complexity. However, inference-time scaling has received less attention, with most approaches restricting models to a single generation attempt. Recent studies have uncovered the existence of "golden noises" that can enhance video quality during generation. Building on this, we find that guiding the scaling inference-time search of VDMs to identify better noise candidates not only evaluates the quality of the frames generated in the current step but also preserves the high-level object features by referencing the anchor frame from previous multi-chunks, thereby delivering long-term value. Our analysis reveals that diffusion models inherently possess flexible adjustments of computation by varying denoising steps, and even a one-step denoising approach, when guided by a reward signal, yields significant long-term benefits. Based on the observation, we proposeScalingNoise, a plug-and-play inference-time search strategy that identifies golden initial noises for the diffusion sampling process to improve global content consistency and visual diversity. Specifically, we perform one-step denoising to convert initial noises into a clip and subsequently evaluate its long-term value, leveraging a reward model anchored by previously generated content. Moreover, to preserve diversity, we sample candidates from a tilted noise distribution that up-weights promising noises. In this way, ScalingNoise significantly reduces noise-induced errors, ensuring more coherent and spatiotemporally consistent video generation. Extensive experiments on benchmark datasets demonstrate that the proposed ScalingNoise effectively improves long video generation.

Xuran Ma, Yexin Liu, Yaofu Liu et al.

Recent advances in diffusion models have demonstrated remarkable capabilities in video generation. However, the computational intensity remains a significant challenge for practical applications. While feature caching has been proposed to reduce the computational burden of diffusion models, existing methods typically overlook the heterogeneous significance of individual blocks, resulting in suboptimal reuse and degraded output quality. To this end, we address this gap by introducing ProfilingDiT, a novel adaptive caching strategy that explicitly disentangles foreground and background-focused blocks. Through a systematic analysis of attention distributions in diffusion models, we reveal a key observation: 1) Most layers exhibit a consistent preference for either foreground or background regions. 2) Predicted noise shows low inter-step similarity initially, which stabilizes as denoising progresses. This finding inspires us to formulate a selective caching strategy that preserves full computation for dynamic foreground elements while efficiently caching static background features. Our approach substantially reduces computational overhead while preserving visual fidelity. Extensive experiments demonstrate that our framework achieves significant acceleration (e.g., 2.01 times speedup for Wan2.1) while maintaining visual fidelity across comprehensive quality metrics, establishing a viable method for efficient video generation.

Yan Shu, Hangui Lin, Yexin Liu et al.

Large Multimodal Models (LMMs) have achieved impressive progress in visual perception and reasoning. However, when confronted with visually ambiguous or non-semantic scene text, they often struggle to accurately spot and understand the content, frequently generating semantically plausible yet visually incorrect answers, which we refer to as semantic hallucination. In this work, we investigate the underlying causes of semantic hallucination and identify a key finding: Transformer layers in LLM with stronger attention focus on scene text regions are less prone to producing semantic hallucinations. Thus, we propose a training-free semantic hallucination mitigation framework comprising two key components: (1) ZoomText, a coarse-to-fine strategy that identifies potential text regions without external detectors; and (2) Grounded Layer Correction, which adaptively leverages the internal representations from layers less prone to hallucination to guide decoding, correcting hallucinated outputs for non-semantic samples while preserving the semantics of meaningful ones. To enable rigorous evaluation, we introduce TextHalu-Bench, a benchmark of 1,740 samples spanning both semantic and non-semantic cases, with manually curated question answer pairs designed to probe model hallucinations. Extensive experiments demonstrate that our method not only effectively mitigates semantic hallucination but also achieves strong performance on public benchmarks for scene text spotting and understanding.

Xianghong Fang, Litao Guo, Hengchao Chen et al.

The success of autoregressive models largely depends on the effectiveness of vector quantization, a technique that discretizes continuous features by mapping them to the nearest code vectors within a learnable codebook. Two critical issues in existing vector quantization methods are training instability and codebook collapse. Training instability arises from the gradient discrepancy introduced by the straight-through estimator, especially in the presence of significant quantization errors, while codebook collapse occurs when only a small subset of code vectors are utilized during training. A closer examination of these issues reveals that they are primarily driven by a mismatch between the distributions of the features and code vectors, leading to unrepresentative code vectors and significant data information loss during compression. To address this, we employ the Wasserstein distance to align these two distributions, achieving near 100\% codebook utilization and significantly reducing the quantization error. Both empirical and theoretical analyses validate the effectiveness of the proposed approach.

Weiming Zhang, Dingwen Xiao, Aobotao Dai et al.

360 video captures the complete surrounding scenes with the ultra-large field of view of 360X180. This makes 360 scene understanding tasks, eg, segmentation and tracking, crucial for appications, such as autonomous driving, robotics. With the recent emergence of foundation models, the community is, however, impeded by the lack of large-scale, labelled real-world datasets. This is caused by the inherent spherical properties, eg, severe distortion in polar regions, and content discontinuities, rendering the annotation costly yet complex. This paper introduces Leader360V, the first large-scale, labeled real-world 360 video datasets for instance segmentation and tracking. Our datasets enjoy high scene diversity, ranging from indoor and urban settings to natural and dynamic outdoor scenes. To automate annotation, we design an automatic labeling pipeline, which subtly coordinates pre-trained 2D segmentors and large language models to facilitate the labeling. The pipeline operates in three novel stages. Specifically, in the Initial Annotation Phase, we introduce a Semantic- and Distortion-aware Refinement module, which combines object mask proposals from multiple 2D segmentors with LLM-verified semantic labels. These are then converted into mask prompts to guide SAM2 in generating distortion-aware masks for subsequent frames. In the Auto-Refine Annotation Phase, missing or incomplete regions are corrected either by applying the SDR again or resolving the discontinuities near the horizontal borders. The Manual Revision Phase finally incorporates LLMs and human annotators to further refine and validate the annotations. Extensive user studies and evaluations demonstrate the effectiveness of our labeling pipeline. Meanwhile, experiments confirm that Leader360V significantly enhances model performance for 360 video segmentation and tracking, paving the way for more scalable 360 scene understanding.

Zidong Cao, Zhan Wang, Yexin Liu et al.

Viewing omnidirectional images (ODIs) in virtual reality (VR) represents a novel form of media that provides immersive experiences for users to navigate and interact with digital content. Nonetheless, this sense of immersion can be greatly compromised by a blur effect that masks details and hampers the user's ability to engage with objects of interest. In this paper, we present a novel system, called OmniVR, designed to enhance visual clarity during VR navigation. Our system enables users to effortlessly locate and zoom in on the objects of interest in VR. It captures user commands for navigation and zoom, converting these inputs into parameters for the Mobius transformation matrix. Leveraging these parameters, the ODI is refined using a learning-based algorithm. The resultant ODI is presented within the VR media, effectively reducing blur and increasing user engagement. To verify the effectiveness of our system, we first evaluate our algorithm with state-of-the-art methods on public datasets, which achieves the best performance. Furthermore, we undertake a comprehensive user study to evaluate viewer experiences across diverse scenarios and to gather their qualitative feedback from multiple perspectives. The outcomes reveal that our system enhances user engagement by improving the viewers' recognition, reducing discomfort, and improving the overall immersive experience. Our system makes the navigation and zoom more user-friendly.

Yexin Liu, Weiming Zhang, Athanasios V. Vasilakos et al.

Unsupervised visible-infrared person re-identification (UVI-ReID) has recently gained great attention due to its potential for enhancing human detection in diverse environments without labeling. Previous methods utilize intra-modality clustering and cross-modality feature matching to achieve UVI-ReID. However, there exist two challenges: 1) noisy pseudo labels might be generated in the clustering process, and 2) the cross-modality feature alignment via matching the marginal distribution of visible and infrared modalities may misalign the different identities from two modalities. In this paper, we first conduct a theoretic analysis where an interpretable generalization upper bound is introduced. Based on the analysis, we then propose a novel unsupervised cross-modality person re-identification framework (PRAISE). Specifically, to address the first challenge, we propose a pseudo-label correction strategy that utilizes a Beta Mixture Model to predict the probability of mis-clustering based network's memory effect and rectifies the correspondence by adding a perceptual term to contrastive learning. Next, we introduce a modality-level alignment strategy that generates paired visible-infrared latent features and reduces the modality gap by aligning the labeling function of visible and infrared features to learn identity discriminative and modality-invariant features. Experimental results on two benchmark datasets demonstrate that our method achieves state-of-the-art performance than the unsupervised visible-ReID methods.

Hongyu Li, Manyuan Zhang, Dian Zheng et al.

Instruction-based image editing has emerged as a prominent research area, which, benefiting from image generation foundation models, have achieved high aesthetic quality, making instruction-following capability the primary challenge. Existing approaches improve instruction adherence via supervised or reinforcement learning, yet single-turn success rates remain limited due to inherent stochasticity and a lack of deliberation. In this work, we propose a deliberative editing framework to 'think' while they edit, which simulates the human cognitive loop by iteratively executing a Think-while-Edit cycle: Critiquing results and Refining instructions , followed by Repeating the generation until satisfactory. Specifically, we train a single MLLM, EditThinker, to act as the reasoning engine of this framework, which jointly produce the critique score, reasoning process, and refined instructions. We employ reinforcement learning to align the EditThinker's thinking with its editing, thereby generating more targeted instruction improvements. Extensive experiments on four benchmarks demonstrate that our approach significantly improves the instruction-following capability of any image editing model by a large margin. We will release our data construction framework, datasets, and models to benefit the community.

Dian Zheng, Manyuan Zhang, Hongyu Li et al.

Unified multimodal models for image generation and understanding represent a significant step toward AGI and have attracted widespread attention from researchers. The main challenge of this task lies in the difficulty in establishing an optimal training paradigm due to inherent conflicting targets in understanding and generation tasks. To alleviate these conflicts and pursue higher performance, many researchers adopt varying degrees of model decoupling (e.g., Double image encoders, MOE/MOT architecture, or frozen MLLM). However, excessive model decoupling can lead to the loss of interleave generation ability, undermining the original intent of unified models. In this work, we aim to explore how to mitigate task conflicts without resorting to model decoupling. Firstly, we analyze why decoupling alleviates conflicts by studying the cross-modal attention behavior of models. We observe that model decoupling essentially drives models toward task-specific multimodal interaction patterns, as seen in Qwen-VL and HunyuanImage, and that the more thorough the decoupling, the more consistent the behavior becomes. Motivated by this observation, we propose Attention Interaction Alignment (AIA) loss, which explicitly learns Task-Specific multimodal interaction patterns during training. To demonstrate the generalizability of our AIA loss, we apply it to Emu3 and Janus-Pro during SFT and post-training stage respectively. Without bells and whistles, AIA not only refines cross-modal attention patterns, but also boosts both generation and understanding performance.

Yexin Liu, Zhengyang Liang, Yueze Wang et al.

Multimodal Large Language Models (MLLMs) have displayed remarkable performance in multi-modal tasks, particularly in visual comprehension. However, we reveal that MLLMs often generate incorrect answers even when they understand the visual content. To this end, we manually construct a benchmark with 12 categories and design evaluation metrics that assess the degree of error in MLLM responses even when the visual content is seemingly understood. Based on this benchmark, we test 15 leading MLLMs and analyze the distribution of attention maps and logits of some MLLMs. Our investigation identifies two primary issues: 1) most instruction tuning datasets predominantly feature questions that 'directly' relate to the visual content, leading to a bias in MLLMs' responses to other indirect questions, and 2) MLLMs' attention to visual tokens is notably lower than to system and question tokens. We further observe that attention scores between questions and visual tokens as well as the model's confidence in the answers are lower in response to misleading questions than to straightforward ones. To address the first challenge, we introduce a paired positive and negative data construction pipeline to diversify the dataset. For the second challenge, we propose to enhance the model's focus on visual content during decoding by refining the text and visual prompt. For the text prompt, we propose a content guided refinement strategy that performs preliminary visual content analysis to generate structured information before answering the question. Additionally, we employ a visual attention refinement strategy that highlights question-relevant visual tokens to increase the model's attention to visual content that aligns with the question. Extensive experiments demonstrate that these challenges can be significantly mitigated with our proposed dataset and techniques.