Haoyu Ma

Ajay Jaiswal, Haoyu Ma, Tianlong Chen et al. · meta-ai

Pruning large neural networks to create high-quality, independently trainable sparse masks, which can maintain similar performance to their dense counterparts, is very desirable due to the reduced space and time complexity. As research effort is focused on increasingly sophisticated pruning methods that leads to sparse subnetworks trainable from the scratch, we argue for an orthogonal, under-explored theme: improving training techniques for pruned sub-networks, i.e. sparse training. Apart from the popular belief that only the quality of sparse masks matters for sparse training, in this paper we demonstrate an alternative opportunity: one can carefully customize the sparse training techniques to deviate from the default dense network training protocols, consisting of introducing ``ghost" neurons and skip connections at the early stage of training, and strategically modifying the initialization as well as labels. Our new sparse training recipe is generally applicable to improving training from scratch with various sparse masks. By adopting our newly curated techniques, we demonstrate significant performance gains across various popular datasets (CIFAR-10, CIFAR-100, TinyImageNet), architectures (ResNet-18/32/104, Vgg16, MobileNet), and sparse mask options (lottery ticket, SNIP/GRASP, SynFlow, or even randomly pruning), compared to the default training protocols, especially at high sparsity levels. Code is at https://github.com/VITA-Group/ToST

Ren Yang, Radu Timofte, Meisong Zheng et al. · tencent-ai

This paper reviews the NTIRE 2022 Challenge on Super-Resolution and Quality Enhancement of Compressed Video. In this challenge, we proposed the LDV 2.0 dataset, which includes the LDV dataset (240 videos) and 95 additional videos. This challenge includes three tracks. Track 1 aims at enhancing the videos compressed by HEVC at a fixed QP. Track 2 and Track 3 target both the super-resolution and quality enhancement of HEVC compressed video. They require x2 and x4 super-resolution, respectively. The three tracks totally attract more than 600 registrations. In the test phase, 8 teams, 8 teams and 12 teams submitted the final results to Tracks 1, 2 and 3, respectively. The proposed methods and solutions gauge the state-of-the-art of super-resolution and quality enhancement of compressed video. The proposed LDV 2.0 dataset is available at https://github.com/RenYang-home/LDV_dataset. The homepage of this challenge (including open-sourced codes) is at https://github.com/RenYang-home/NTIRE22_VEnh_SR.

Haoyu Ma, Jialong Wu, Ningya Feng et al. · tsinghua

Model-based reinforcement learning (MBRL) holds the promise of sample-efficient learning by utilizing a world model, which models how the environment works and typically encompasses components for two tasks: observation modeling and reward modeling. In this paper, through a dedicated empirical investigation, we gain a deeper understanding of the role each task plays in world models and uncover the overlooked potential of sample-efficient MBRL by mitigating the domination of either observation or reward modeling. Our key insight is that while prevalent approaches of explicit MBRL attempt to restore abundant details of the environment via observation models, it is difficult due to the environment's complexity and limited model capacity. On the other hand, reward models, while dominating implicit MBRL and adept at learning compact task-centric dynamics, are inadequate for sample-efficient learning without richer learning signals. Motivated by these insights and discoveries, we propose a simple yet effective approach, HarmonyDream, which automatically adjusts loss coefficients to maintain task harmonization, i.e. a dynamic equilibrium between the two tasks in world model learning. Our experiments show that the base MBRL method equipped with HarmonyDream gains 10%-69% absolute performance boosts on visual robotic tasks and sets a new state-of-the-art result on the Atari 100K benchmark. Code is available at https://github.com/thuml/HarmonyDream.

Zhengang Li, Mengshu Sun, Alec Lu et al. · meta-ai

Vision transformers (ViTs) are emerging with significantly improved accuracy in computer vision tasks. However, their complex architecture and enormous computation/storage demand impose urgent needs for new hardware accelerator design methodology. This work proposes an FPGA-aware automatic ViT acceleration framework based on the proposed mixed-scheme quantization. To the best of our knowledge, this is the first FPGA-based ViT acceleration framework exploring model quantization. Compared with state-of-the-art ViT quantization work (algorithmic approach only without hardware acceleration), our quantization achieves 0.47% to 1.36% higher Top-1 accuracy under the same bit-width. Compared with the 32-bit floating-point baseline FPGA accelerator, our accelerator achieves around 5.6x improvement on the frame rate (i.e., 56.8 FPS vs. 10.0 FPS) with 0.71% accuracy drop on ImageNet dataset for DeiT-base.

Zhenglun Kong, Haoyu Ma, Geng Yuan et al. · harvard, meta-ai

Vision transformers (ViTs) have recently obtained success in many applications, but their intensive computation and heavy memory usage at both training and inference time limit their generalization. Previous compression algorithms usually start from the pre-trained dense models and only focus on efficient inference, while time-consuming training is still unavoidable. In contrast, this paper points out that the million-scale training data is redundant, which is the fundamental reason for the tedious training. To address the issue, this paper aims to introduce sparsity into data and proposes an end-to-end efficient training framework from three sparse perspectives, dubbed Tri-Level E-ViT. Specifically, we leverage a hierarchical data redundancy reduction scheme, by exploring the sparsity under three levels: number of training examples in the dataset, number of patches (tokens) in each example, and number of connections between tokens that lie in attention weights. With extensive experiments, we demonstrate that our proposed technique can noticeably accelerate training for various ViT architectures while maintaining accuracy. Remarkably, under certain ratios, we are able to improve the ViT accuracy rather than compromising it. For example, we can achieve 15.2% speedup with 72.6% (+0.4) Top-1 accuracy on Deit-T, and 15.7% speedup with 79.9% (+0.1) Top-1 accuracy on Deit-S. This proves the existence of data redundancy in ViT.

Deying Kong, Linguang Zhang, Liangjian Chen et al. · meta-ai

Reconstructing 3D hand meshes from monocular RGB images has attracted increasing amount of attention due to its enormous potential applications in the field of AR/VR. Most state-of-the-art methods attempt to tackle this task in an anonymous manner. Specifically, the identity of the subject is ignored even though it is practically available in real applications where the user is unchanged in a continuous recording session. In this paper, we propose an identity-aware hand mesh estimation model, which can incorporate the identity information represented by the intrinsic shape parameters of the subject. We demonstrate the importance of the identity information by comparing the proposed identity-aware model to a baseline which treats subject anonymously. Furthermore, to handle the use case where the test subject is unseen, we propose a novel personalization pipeline to calibrate the intrinsic shape parameters using only a few unlabeled RGB images of the subject. Experiments on two large scale public datasets validate the state-of-the-art performance of our proposed method.

Zecheng He, Bo Sun, Felix Juefei-Xu et al. · meta-ai

Diffusion models have demonstrated remarkable efficacy across various image-to-image tasks. In this research, we introduce Imagine yourself, a state-of-the-art model designed for personalized image generation. Unlike conventional tuning-based personalization techniques, Imagine yourself operates as a tuning-free model, enabling all users to leverage a shared framework without individualized adjustments. Moreover, previous work met challenges balancing identity preservation, following complex prompts and preserving good visual quality, resulting in models having strong copy-paste effect of the reference images. Thus, they can hardly generate images following prompts that require significant changes to the reference image, \eg, changing facial expression, head and body poses, and the diversity of the generated images is low. To address these limitations, our proposed method introduces 1) a new synthetic paired data generation mechanism to encourage image diversity, 2) a fully parallel attention architecture with three text encoders and a fully trainable vision encoder to improve the text faithfulness, and 3) a novel coarse-to-fine multi-stage finetuning methodology that gradually pushes the boundary of visual quality. Our study demonstrates that Imagine yourself surpasses the state-of-the-art personalization model, exhibiting superior capabilities in identity preservation, visual quality, and text alignment. This model establishes a robust foundation for various personalization applications. Human evaluation results validate the model's SOTA superiority across all aspects (identity preservation, text faithfulness, and visual appeal) compared to the previous personalization models.

Haoyu Ma, Zhe Wang, Yifei Chen et al. · meta-ai

Recently, the vision transformer and its variants have played an increasingly important role in both monocular and multi-view human pose estimation. Considering image patches as tokens, transformers can model the global dependencies within the entire image or across images from other views. However, global attention is computationally expensive. As a consequence, it is difficult to scale up these transformer-based methods to high-resolution features and many views. In this paper, we propose the token-Pruned Pose Transformer (PPT) for 2D human pose estimation, which can locate a rough human mask and performs self-attention only within selected tokens. Furthermore, we extend our PPT to multi-view human pose estimation. Built upon PPT, we propose a new cross-view fusion strategy, called human area fusion, which considers all human foreground pixels as corresponding candidates. Experimental results on COCO and MPII demonstrate that our PPT can match the accuracy of previous pose transformer methods while reducing the computation. Moreover, experiments on Human 3.6M and Ski-Pose demonstrate that our Multi-view PPT can efficiently fuse cues from multiple views and achieve new state-of-the-art results.

Xiangyi Yan, Junayed Naushad, Chenyu You et al. · meta-ai

Recent advancements in self-supervised learning have demonstrated that effective visual representations can be learned from unlabeled images. This has led to increased interest in applying self-supervised learning to the medical domain, where unlabeled images are abundant and labeled images are difficult to obtain. However, most self-supervised learning approaches are modeled as image level discriminative or generative proxy tasks, which may not capture the finer level representations necessary for dense prediction tasks like multi-organ segmentation. In this paper, we propose a novel contrastive learning framework that integrates Localized Region Contrast (LRC) to enhance existing self-supervised pre-training methods for medical image segmentation. Our approach involves identifying Super-pixels by Felzenszwalb's algorithm and performing local contrastive learning using a novel contrastive sampling loss. Through extensive experiments on three multi-organ segmentation datasets, we demonstrate that integrating LRC to an existing self-supervised method in a limited annotation setting significantly improves segmentation performance. Moreover, we show that LRC can also be applied to fully-supervised pre-training methods to further boost performance.

Shanlin Sun, Thanh-Tung Le, Chenyu You et al. · meta-ai

We present Hybrid-CSR, a geometric deep-learning model that combines explicit and implicit shape representations for cortical surface reconstruction. Specifically, Hybrid-CSR begins with explicit deformations of template meshes to obtain coarsely reconstructed cortical surfaces, based on which the oriented point clouds are estimated for the subsequent differentiable poisson surface reconstruction. By doing so, our method unifies explicit (oriented point clouds) and implicit (indicator function) cortical surface reconstruction. Compared to explicit representation-based methods, our hybrid approach is more friendly to capture detailed structures, and when compared with implicit representation-based methods, our method can be topology aware because of end-to-end training with a mesh-based deformation module. In order to address topology defects, we propose a new topology correction pipeline that relies on optimization-based diffeomorphic surface registration. Experimental results on three brain datasets show that our approach surpasses existing implicit and explicit cortical surface reconstruction methods in numeric metrics in terms of accuracy, regularity, and consistency.

Yifeng Xiong, Haoyu Ma, Shanlin Sun et al. · meta-ai

Single-view novel view synthesis (NVS), the task of generating images from new viewpoints based on a single reference image, is important but challenging in computer vision. Recent advancements in NVS have leveraged Denoising Diffusion Probabilistic Models (DDPMs) for their exceptional ability to produce high-fidelity images. However, current diffusion-based methods typically utilize camera pose matrices to globally and implicitly enforce 3D constraints, which can lead to inconsistencies in images generated from varying viewpoints, particularly in regions with complex textures and structures. To address these limitations, we present Light Field Diffusion (LFD), a novel conditional diffusion-based approach that transcends the conventional reliance on camera pose matrices. Starting from the camera pose matrices, LFD transforms them into light field encoding, with the same shape as the reference image, to describe the direction of each ray. By integrating light field encoding with the reference image, our method imposes local pixel-wise constraints within the diffusion process, fostering enhanced view consistency. Our approach not only involves training image LFD on the ShapeNet Car dataset but also includes fine-tuning a pre-trained latent diffusion model on the Objaverse dataset. This enables our latent LFD model to exhibit remarkable zero-shot generalization capabilities across out-of-distribution datasets like RTMV as well as in-the-wild images. Experiments demonstrate that LFD not only produces high-fidelity images but also achieves superior 3D consistency in complex regions, outperforming existing novel view synthesis methods.

Haoyu Ma, Tong Zhang, Shanlin Sun et al. · meta-ai

Reconstructing personalized animatable head avatars has significant implications in the fields of AR/VR. Existing methods for achieving explicit face control of 3D Morphable Models (3DMM) typically rely on multi-view images or videos of a single subject, making the reconstruction process complex. Additionally, the traditional rendering pipeline is time-consuming, limiting real-time animation possibilities. In this paper, we introduce CVTHead, a novel approach that generates controllable neural head avatars from a single reference image using point-based neural rendering. CVTHead considers the sparse vertices of mesh as the point set and employs the proposed Vertex-feature Transformer to learn local feature descriptors for each vertex. This enables the modeling of long-range dependencies among all the vertices. Experimental results on the VoxCeleb dataset demonstrate that CVTHead achieves comparable performance to state-of-the-art graphics-based methods. Moreover, it enables efficient rendering of novel human heads with various expressions, head poses, and camera views. These attributes can be explicitly controlled using the coefficients of 3DMMs, facilitating versatile and realistic animation in real-time scenarios.

Haoyu Ma, Xiangru Lin, Yizhou Yu

Unsupervised domain adaptation (UDA) for semantic segmentation is a promising task freeing people from heavy annotation work. However, domain discrepancies in low-level image statistics and high-level contexts compromise the segmentation performance over the target domain. A key idea to tackle this problem is to perform both image-level and feature-level adaptation jointly. Unfortunately, there is a lack of such unified approaches for UDA tasks in the existing literature. This paper proposes a novel UDA pipeline for semantic segmentation that unifies image-level and feature-level adaptation. Concretely, for image-level domain shifts, we propose a global photometric alignment module and a global texture alignment module that align images in the source and target domains in terms of image-level properties. For feature-level domain shifts, we perform global manifold alignment by projecting pixel features from both domains onto the feature manifold of the source domain; and we further regularize category centers in the source domain through a category-oriented triplet loss and perform target domain consistency regularization over augmented target domain images. Experimental results demonstrate that our pipeline significantly outperforms previous methods. In the commonly tested GTA5$\rightarrow$Cityscapes task, our proposed method using Deeplab V3+ as the backbone surpasses previous SOTA by 8%, achieving 58.2% in mIoU.

Xiaopeng Sun, Weiqi Li, Zhenyu Zhang et al.

360° omnidirectional images have gained research attention due to their immersive and interactive experience, particularly in AR/VR applications. However, they suffer from lower angular resolution due to being captured by fisheye lenses with the same sensor size for capturing planar images. To solve the above issues, we propose a two-stage framework for 360° omnidirectional image superresolution. The first stage employs two branches: model A, which incorporates omnidirectional position-aware deformable blocks (OPDB) and Fourier upsampling, and model B, which adds a spatial frequency fusion module (SFF) to model A. Model A aims to enhance the feature extraction ability of 360° image positional information, while Model B further focuses on the high-frequency information of 360° images. The second stage performs same-resolution enhancement based on the structure of model A with a pixel unshuffle operation. In addition, we collected data from YouTube to improve the fitting ability of the transformer, and created pseudo low-resolution images using a degradation network. Our proposed method achieves superior performance and wins the NTIRE 2023 challenge of 360° omnidirectional image super-resolution.

Shanlin Sun, Kun Han, Chenyu You et al.

Image registration is an essential step in many medical image analysis tasks. Traditional methods for image registration are primarily optimization-driven, finding the optimal deformations that maximize the similarity between two images. Recent learning-based methods, trained to directly predict transformations between two images, run much faster, but suffer from performance deficiencies due to model generalization and the inefficiency in handling individual image specific deformations. Here we present a new neural net based image registration framework, called NIR (Neural Image Registration), which is based on optimization but utilizes deep neural nets to model deformations between image pairs. NIR represents the transformation between two images with a continuous function implemented via neural fields, receiving a 3D coordinate as input and outputting the corresponding deformation vector. NIR provides two ways of generating deformation field: directly output a displacement vector field for general deformable registration, or output a velocity vector field and integrate the velocity field to derive the deformation field for diffeomorphic image registration. The optimal registration is discovered by updating the parameters of the neural field via stochastic gradient descent. We describe several design choices that facilitate model optimization, including coordinate encoding, sinusoidal activation, coordinate sampling, and intensity sampling. Experiments on two 3D MR brain scan datasets demonstrate that NIR yields state-of-the-art performance in terms of both registration accuracy and regularity, while running significantly faster than traditional optimization-based methods.

Leon Liangyu Chen, Haoyu Ma, Zhipeng Fan et al.

Unified models can handle both multimodal understanding and generation within a single architecture, yet they typically operate in a single pass without iteratively refining their outputs. Many multimodal tasks, especially those involving complex spatial compositions, multiple interacting objects, or evolving instructions, require decomposing instructions, verifying intermediate results, and making iterative corrections. While test-time scaling (TTS) has demonstrated that allocating additional inference compute for iterative reasoning substantially improves language model performance, extending this paradigm to unified multimodal models remains an open challenge. We introduce UniT, a framework for multimodal chain-of-thought test-time scaling that enables a single unified model to reason, verify, and refine across multiple rounds. UniT combines agentic data synthesis, unified model training, and flexible test-time inference to elicit cognitive behaviors including verification, subgoal decomposition, and content memory. Our key findings are: (1) unified models trained on short reasoning trajectories generalize to longer inference chains at test time; (2) sequential chain-of-thought reasoning provides a more scalable and compute-efficient TTS strategy than parallel sampling; (3) training on generation and editing trajectories improves out-of-distribution visual reasoning. These results establish multimodal test-time scaling as an effective paradigm for advancing both generation and understanding in unified models.

Boxiang Yang, Ning Chen, Xia Yue et al.

Recently, Hyperspectral Image (HSI) classification has attracted increasing attention in remote sensing. However, HSI data are inherently high-dimensional but low-rank, with discriminative information concentrated on a low-dimensional latent manifold. In real-world remote sensing scenarios, the superposition of multiple degradation factors disrupts this intrinsic manifold structure, driving samples away from their original low-dimensional distribution and introducing substantial redundant and non-discriminative variations. To better handle this challenge, this paper proposes a manifold-space diffusion framework (MSDiff) for robust hyperspectral classification under complex degradation conditions. Specifically, the proposed method first maps high-dimensional, degradation-affected HSI data into a compact low-dimensional manifold through a discriminative spectral-spatial reconstruction task, preserving class semantics and reducing redundant variations. A diffusion-based generative model is then applied to regularize the spectral-spatial distribution within the manifold, enabling progressive refinement and stabilization of latent features against residual degradations. The key advantage of the proposed framework lies in performing diffusion-based distribution modeling directly on the low-dimensional manifold, effectively decoupling degradation-induced disturbances from intrinsic discriminative structures and enhancing representation stability under complex degradations. Experimental results on multiple hyperspectral benchmarks demonstrate consistent performance improvements over state-of-the-art methods under diverse composite degradation settings. The code will be available at https://github.com/yangboxiang1207/MSDiff

Kunpeng Song, Tingbo Hou, Zecheng He et al.

In this paper, we introduce DirectorLLM, a novel video generation model that employs a large language model (LLM) to orchestrate human poses within videos. As foundational text-to-video models rapidly evolve, the demand for high-quality human motion and interaction grows. To address this need and enhance the authenticity of human motions, we extend the LLM from a text generator to a video director and human motion simulator. Utilizing open-source resources from Llama 3, we train the DirectorLLM to generate detailed instructional signals, such as human poses, to guide video generation. This approach offloads the simulation of human motion from the video generator to the LLM, effectively creating informative outlines for human-centric scenes. These signals are used as conditions by the video renderer, facilitating more realistic and prompt-following video generation. As an independent LLM module, it can be applied to different video renderers, including UNet and DiT, with minimal effort. Experiments on automatic evaluation benchmarks and human evaluations show that our model outperforms existing ones in generating videos with higher human motion fidelity, improved prompt faithfulness, and enhanced rendered subject naturalness.

Adam Polyak, Amit Zohar, Andrew Brown et al. · meta-ai

We present Movie Gen, a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. We also show additional capabilities such as precise instruction-based video editing and generation of personalized videos based on a user's image. Our models set a new state-of-the-art on multiple tasks: text-to-video synthesis, video personalization, video editing, video-to-audio generation, and text-to-audio generation. Our largest video generation model is a 30B parameter transformer trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second. We show multiple technical innovations and simplifications on the architecture, latent spaces, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations that allow us to reap the benefits of scaling pre-training data, model size, and training compute for training large scale media generation models. We hope this paper helps the research community to accelerate progress and innovation in media generation models. All videos from this paper are available at https://go.fb.me/MovieGenResearchVideos.

Jialong Wu, Haoyu Ma, Chaoyi Deng et al.

Unsupervised pre-training methods utilizing large and diverse datasets have achieved tremendous success across a range of domains. Recent work has investigated such unsupervised pre-training methods for model-based reinforcement learning (MBRL) but is limited to domain-specific or simulated data. In this paper, we study the problem of pre-training world models with abundant in-the-wild videos for efficient learning of downstream visual control tasks. However, in-the-wild videos are complicated with various contextual factors, such as intricate backgrounds and textured appearance, which precludes a world model from extracting shared world knowledge to generalize better. To tackle this issue, we introduce Contextualized World Models (ContextWM) that explicitly separate context and dynamics modeling to overcome the complexity and diversity of in-the-wild videos and facilitate knowledge transfer between distinct scenes. Specifically, a contextualized extension of the latent dynamics model is elaborately realized by incorporating a context encoder to retain contextual information and empower the image decoder, which encourages the latent dynamics model to concentrate on essential temporal variations. Our experiments show that in-the-wild video pre-training equipped with ContextWM can significantly improve the sample efficiency of MBRL in various domains, including robotic manipulation, locomotion, and autonomous driving. Code is available at this repository: https://github.com/thuml/ContextWM.

Tianlong Chen, Zhenyu Zhang, Pengjun Wang et al.

Recent studies demonstrate that deep networks, even robustified by the state-of-the-art adversarial training (AT), still suffer from large robust generalization gaps, in addition to the much more expensive training costs than standard training. In this paper, we investigate this intriguing problem from a new perspective, i.e., injecting appropriate forms of sparsity during adversarial training. We introduce two alternatives for sparse adversarial training: (i) static sparsity, by leveraging recent results from the lottery ticket hypothesis to identify critical sparse subnetworks arising from the early training; (ii) dynamic sparsity, by allowing the sparse subnetwork to adaptively adjust its connectivity pattern (while sticking to the same sparsity ratio) throughout training. We find both static and dynamic sparse methods to yield win-win: substantially shrinking the robust generalization gap and alleviating the robust overfitting, meanwhile significantly saving training and inference FLOPs. Extensive experiments validate our proposals with multiple network architectures on diverse datasets, including CIFAR-10/100 and Tiny-ImageNet. For example, our methods reduce robust generalization gap and overfitting by 34.44% and 4.02%, with comparable robust/standard accuracy boosts and 87.83%/87.82% training/inference FLOPs savings on CIFAR-100 with ResNet-18. Besides, our approaches can be organically combined with existing regularizers, establishing new state-of-the-art results in AT. Codes are available in https://github.com/VITA-Group/Sparsity-Win-Robust-Generalization.

Deying Kong, Yifei Chen, Haoyu Ma et al.

In this paper, we propose a new architecture called Adaptive Graphical Model Network (AGMN) to tackle the task of 2D hand pose estimation from a monocular RGB image. The AGMN consists of two branches of deep convolutional neural networks for calculating unary and pairwise potential functions, followed by a graphical model inference module for integrating unary and pairwise potentials. Unlike existing architectures proposed to combine DCNNs with graphical models, our AGMN is novel in that the parameters of its graphical model are conditioned on and fully adaptive to individual input images. Experiments show that our approach outperforms the state-of-the-art method used in 2D hand keypoints estimation by a notable margin on two public datasets. Code can be found at https://github.com/deyingk/agmn.

Xu Ma, Peize Sun, Haoyu Ma et al.

Autoregressive (AR) models, long dominant in language generation, are increasingly applied to image synthesis but are often considered less competitive than Diffusion-based models. A primary limitation is the substantial number of image tokens required for AR models, which constrains both training and inference efficiency, as well as image resolution. To address this, we present Token-Shuffle, a novel yet simple method that reduces the number of image tokens in Transformer. Our key insight is the dimensional redundancy of visual vocabularies in Multimodal Large Language Models (MLLMs), where low-dimensional visual codes from visual encoder are directly mapped to high-dimensional language vocabularies. Leveraging this, we consider two key operations: token-shuffle, which merges spatially local tokens along channel dimension to decrease the input token number, and token-unshuffle, which untangles the inferred tokens after Transformer blocks to restore the spatial arrangement for output. Jointly training with textual prompts, our strategy requires no additional pretrained text-encoder and enables MLLMs to support extremely high-resolution image synthesis in a unified next-token prediction way while maintaining efficient training and inference. For the first time, we push the boundary of AR text-to-image generation to a resolution of 2048x2048 with gratifying generation performance. In GenAI-benchmark, our 2.7B model achieves 0.77 overall score on hard prompts, outperforming AR models LlamaGen by 0.18 and diffusion models LDM by 0.15. Exhaustive large-scale human evaluations also demonstrate our prominent image generation ability in terms of text-alignment, visual flaw, and visual appearance. We hope that Token-Shuffle can serve as a foundational design for efficient high-resolution image generation within MLLMs.

Cong Wei, Bo Sun, Haoyu Ma et al.

Recent advancements in video generation have achieved impressive motion realism, yet they often overlook character-driven storytelling, a crucial task for automated film, animation generation. We introduce Talking Characters, a more realistic task to generate talking character animations directly from speech and text. Unlike talking head, Talking Characters aims at generating the full portrait of one or more characters beyond the facial region. In this paper, we propose MoCha, the first of its kind to generate talking characters. To ensure precise synchronization between video and speech, we propose a speech-video window attention mechanism that effectively aligns speech and video tokens. To address the scarcity of large-scale speech-labeled video datasets, we introduce a joint training strategy that leverages both speech-labeled and text-labeled video data, significantly improving generalization across diverse character actions. We also design structured prompt templates with character tags, enabling, for the first time, multi-character conversation with turn-based dialogue-allowing AI-generated characters to engage in context-aware conversations with cinematic coherence. Extensive qualitative and quantitative evaluations, including human preference studies and benchmark comparisons, demonstrate that MoCha sets a new standard for AI-generated cinematic storytelling, achieving superior realism, expressiveness, controllability and generalization.

Yunhan Zhao, Haoyu Ma, Shu Kong et al. · meta-ai

Egocentric sensors such as AR/VR devices capture human-object interactions and offer the potential to provide task-assistance by recalling 3D locations of objects of interest in the surrounding environment. This capability requires instance tracking in real-world 3D scenes from egocentric videos (IT3DEgo). We explore this problem by first introducing a new benchmark dataset, consisting of RGB and depth videos, per-frame camera pose, and instance-level annotations in both 2D camera and 3D world coordinates. We present an evaluation protocol which evaluates tracking performance in 3D coordinates with two settings for enrolling instances to track: (1) single-view online enrollment where an instance is specified on-the-fly based on the human wearer's interactions. and (2) multi-view pre-enrollment where images of an instance to be tracked are stored in memory ahead of time. To address IT3DEgo, we first re-purpose methods from relevant areas, e.g., single object tracking (SOT) -- running SOT methods to track instances in 2D frames and lifting them to 3D using camera pose and depth. We also present a simple method that leverages pretrained segmentation and detection models to generate proposals from RGB frames and match proposals with enrolled instance images. Our experiments show that our method (with no finetuning) significantly outperforms SOT-based approaches in the egocentric setting. We conclude by arguing that the problem of egocentric instance tracking is made easier by leveraging camera pose and using a 3D allocentric (world) coordinate representation.

Feng Liang, Haoyu Ma, Zecheng He et al.

Video personalization, which generates customized videos using reference images, has gained significant attention. However, prior methods typically focus on single-concept personalization, limiting broader applications that require multi-concept integration. Attempts to extend these models to multiple concepts often lead to identity blending, which results in composite characters with fused attributes from multiple sources. This challenge arises due to the lack of a mechanism to link each concept with its specific reference image. We address this with anchored prompts, which embed image anchors as unique tokens within text prompts, guiding accurate referencing during generation. Additionally, we introduce concept embeddings to encode the order of reference images. Our approach, Movie Weaver, seamlessly weaves multiple concepts-including face, body, and animal images-into one video, allowing flexible combinations in a single model. The evaluation shows that Movie Weaver outperforms existing methods for multi-concept video personalization in identity preservation and overall quality.

Haoyu Ma, Yushu Chen, Wenlai Zhao et al. · tsinghua

Time series foundation models have demonstrated strong performance in zero-shot learning, making them well-suited for predicting rapidly evolving patterns in real-world applications where relevant training data are scarce. However, most of these models rely on the Transformer architecture, which incurs quadratic complexity as input length increases. To address this, we introduce TSMamba, a linear-complexity foundation model for time series forecasting built on the Mamba architecture. The model captures temporal dependencies through both forward and backward Mamba encoders, achieving high prediction accuracy. To reduce reliance on large datasets and lower training costs, TSMamba employs a two-stage transfer learning process that leverages pretrained Mamba LLMs, allowing effective time series modeling with a moderate training set. In the first stage, the forward and backward backbones are optimized via patch-wise autoregressive prediction; in the second stage, the model trains a prediction head and refines other components for long-term forecasting. While the backbone assumes channel independence to manage varying channel numbers across datasets, a channel-wise compressed attention module is introduced to capture cross-channel dependencies during fine-tuning on specific multivariate datasets. Experiments show that TSMamba's zero-shot performance is comparable to state-of-the-art time series foundation models, despite using significantly less training data. It also achieves competitive or superior full-shot performance compared to task-specific prediction models. The code will be made publicly available.

Mingzhen Huang, Fu-Jen Chu, Bugra Tekin et al.

We introduce HOIGPT, a token-based generative method that unifies 3D hand-object interactions (HOI) perception and generation, offering the first comprehensive solution for captioning and generating high-quality 3D HOI sequences from a diverse range of conditional signals (\eg text, objects, partial sequences). At its core, HOIGPT utilizes a large language model to predict the bidrectional transformation between HOI sequences and natural language descriptions. Given text inputs, HOIGPT generates a sequence of hand and object meshes; given (partial) HOI sequences, HOIGPT generates text descriptions and completes the sequences. To facilitate HOI understanding with a large language model, this paper introduces two key innovations: (1) a novel physically grounded HOI tokenizer, the hand-object decomposed VQ-VAE, for discretizing HOI sequences, and (2) a motion-aware language model trained to process and generate both text and HOI tokens. Extensive experiments demonstrate that HOIGPT sets new state-of-the-art performance on both text generation (+2.01% R Precision) and HOI generation (-2.56 FID) across multiple tasks and benchmarks.

Haoyu Ma, Shahin Mahdizadehaghdam, Bichen Wu et al. · meta-ai

Recent advances in generative AI have significantly enhanced image and video editing, particularly in the context of text prompt control. State-of-the-art approaches predominantly rely on diffusion models to accomplish these tasks. However, the computational demands of diffusion-based methods are substantial, often necessitating large-scale paired datasets for training, and therefore challenging the deployment in real applications. To address these issues, this paper breaks down the text-based video editing task into two stages. First, we leverage an pre-trained text-to-image diffusion model to simultaneously edit few keyframes in an zero-shot way. Second, we introduce an efficient model called MaskINT, which is built on non-autoregressive masked generative transformers and specializes in frame interpolation between the edited keyframes, using the structural guidance from intermediate frames. Experimental results suggest that our MaskINT achieves comparable performance with diffusion-based methodologies, while significantly improve the inference time. This research offers a practical solution for text-based video editing and showcases the potential of non-autoregressive masked generative transformers in this domain.

Jingyu Song, Haoyu Ma, Onur Bagoren et al.

Underwater simulators offer support for building robust underwater perception solutions. Significant work has recently been done to develop new simulators and to advance the performance of existing underwater simulators. Still, there remains room for improvement on physics-based underwater sensor modeling and rendering efficiency. In this paper, we propose OceanSim, a high-fidelity GPU-accelerated underwater simulator to address this research gap. We propose advanced physics-based rendering techniques to reduce the sim-to-real gap for underwater image simulation. We develop OceanSim to fully leverage the computing advantages of GPUs and achieve real-time imaging sonar rendering and fast synthetic data generation. We evaluate the capabilities and realism of OceanSim using real-world data to provide qualitative and quantitative results. The code and detailed documentation are made available on the project website to support the marine robotics community: https://umfieldrobotics.github.io/OceanSim.

Haochen Zhang, Animesh Sinha, Felix Juefei-Xu et al.

Conversational image generation requires a model to follow user instructions across multiple rounds of interaction, grounded in interleaved text and images that accumulate as chat history. While recent multimodal large language models (MLLMs) can generate and edit images, most existing multi-turn benchmarks and training recipes are effectively Markov: the next output depends primarily on the most recent image, enabling shortcut solutions that ignore long-range history. In this work we formalize and target the more challenging non-Markov setting, where a user may refer back to earlier states, undo changes, or reference entities introduced several rounds ago. We present (i) non-Markov multi-round data construction strategies, including rollback-style editing that forces retrieval of earlier visual states and name-based multi-round personalization that binds names to appearances across rounds; (ii) a history-conditioned training and inference framework with token-level caching to prevent multi-round identity drift; and (iii) enabling improvements for high-fidelity image reconstruction and editable personalization, including a reconstruction-based DiT detokenizer and a multi-stage fine-tuning curriculum. We demonstrate that explicitly training for non-Markov interactions yields substantial improvements in multi-round consistency and instruction compliance, while maintaining strong single-round editing and personalization.

Mengyi Shan, Zecheng He, Haoyu Ma et al.

Can a video generation model be repurposed as an interactive world simulator? We explore the affordance perception potential of text-to-video models by teaching them to predict human-environment interaction. Given a scene image and a prompt describing human actions, we fine-tune the model to insert a person into the scene, while ensuring coherent behavior, appearance, harmonization, and scene affordance. Unlike prior work, we infer human affordance for video generation (i.e., where to insert a person and how they should behave) from a single scene image, without explicit conditions like bounding boxes or body poses. An in-depth study of cross-attention heatmaps demonstrates that we can uncover the inherent affordance perception of a pre-trained video model without labeled affordance datasets.

Ming Cheng, Haoyu Ma, Qiufang Ma et al.

Multi-stage strategies are frequently employed in image restoration tasks. While transformer-based methods have exhibited high efficiency in single-image super-resolution tasks, they have not yet shown significant advantages over CNN-based methods in stereo super-resolution tasks. This can be attributed to two key factors: first, current single-image super-resolution transformers are unable to leverage the complementary stereo information during the process; second, the performance of transformers is typically reliant on sufficient data, which is absent in common stereo-image super-resolution algorithms. To address these issues, we propose a Hybrid Transformer and CNN Attention Network (HTCAN), which utilizes a transformer-based network for single-image enhancement and a CNN-based network for stereo information fusion. Furthermore, we employ a multi-patch training strategy and larger window sizes to activate more input pixels for super-resolution. We also revisit other advanced techniques, such as data augmentation, data ensemble, and model ensemble to reduce overfitting and data bias. Finally, our approach achieved a score of 23.90dB and emerged as the winner in Track 1 of the NTIRE 2023 Stereo Image Super-Resolution Challenge.

Kun Han, Shanlin sun, Xiangyi Yan et al.

Diffeomorphic image registration, offering smooth transformation and topology preservation, is required in many medical image analysis tasks.Traditional methods impose certain modeling constraints on the space of admissible transformations and use optimization to find the optimal transformation between two images. Specifying the right space of admissible transformations is challenging: the registration quality can be poor if the space is too restrictive, while the optimization can be hard to solve if the space is too general. Recent learning-based methods, utilizing deep neural networks to learn the transformation directly, achieve fast inference, but face challenges in accuracy due to the difficulties in capturing the small local deformations and generalization ability. Here we propose a new optimization-based method named DNVF (Diffeomorphic Image Registration with Neural Velocity Field) which utilizes deep neural network to model the space of admissible transformations. A multilayer perceptron (MLP) with sinusoidal activation function is used to represent the continuous velocity field and assigns a velocity vector to every point in space, providing the flexibility of modeling complex deformations as well as the convenience of optimization. Moreover, we propose a cascaded image registration framework (Cas-DNVF) by combining the benefits of both optimization and learning based methods, where a fully convolutional neural network (FCN) is trained to predict the initial deformation, followed by DNVF for further refinement. Experiments on two large-scale 3D MR brain scan datasets demonstrate that our proposed methods significantly outperform the state-of-the-art registration methods.

Mengshu Sun, Haoyu Ma, Guoliang Kang et al.

The transformer architectures with attention mechanisms have obtained success in Nature Language Processing (NLP), and Vision Transformers (ViTs) have recently extended the application domains to various vision tasks. While achieving high performance, ViTs suffer from large model size and high computation complexity that hinders the deployment of them on edge devices. To achieve high throughput on hardware and preserve the model accuracy simultaneously, we propose VAQF, a framework that builds inference accelerators on FPGA platforms for quantized ViTs with binary weights and low-precision activations. Given the model structure and the desired frame rate, VAQF will automatically output the required quantization precision for activations as well as the optimized parameter settings of the accelerator that fulfill the hardware requirements. The implementations are developed with Vivado High-Level Synthesis (HLS) on the Xilinx ZCU102 FPGA board, and the evaluation results with the DeiT-base model indicate that a frame rate requirement of 24 frames per second (FPS) is satisfied with 8-bit activation quantization, and a target of 30 FPS is met with 6-bit activation quantization. To the best of our knowledge, this is the first time quantization has been incorporated into ViT acceleration on FPGAs with the help of a fully automatic framework to guide the quantization strategy on the software side and the accelerator implementations on the hardware side given the target frame rate. Very small compilation time cost is incurred compared with quantization training, and the generated accelerators show the capability of achieving real-time execution for state-of-the-art ViT models on FPGAs.

Xiangyi Yan, Hao Tang, Shanlin Sun et al.

Recent advances in transformer-based models have drawn attention to exploring these techniques in medical image segmentation, especially in conjunction with the U-Net model (or its variants), which has shown great success in medical image segmentation, under both 2D and 3D settings. Current 2D based methods either directly replace convolutional layers with pure transformers or consider a transformer as an additional intermediate encoder between the encoder and decoder of U-Net. However, these approaches only consider the attention encoding within one single slice and do not utilize the axial-axis information naturally provided by a 3D volume. In the 3D setting, convolution on volumetric data and transformers both consume large GPU memory. One has to either downsample the image or use cropped local patches to reduce GPU memory usage, which limits its performance. In this paper, we propose Axial Fusion Transformer UNet (AFTer-UNet), which takes both advantages of convolutional layers' capability of extracting detailed features and transformers' strength on long sequence modeling. It considers both intra-slice and inter-slice long-range cues to guide the segmentation. Meanwhile, it has fewer parameters and takes less GPU memory to train than the previous transformer-based models. Extensive experiments on three multi-organ segmentation datasets demonstrate that our method outperforms current state-of-the-art methods.

Haoyu Ma, Liangjian Chen, Deying Kong et al.

Estimating the 2D human poses in each view is typically the first step in calibrated multi-view 3D pose estimation. But the performance of 2D pose detectors suffers from challenging situations such as occlusions and oblique viewing angles. To address these challenges, previous works derive point-to-point correspondences between different views from epipolar geometry and utilize the correspondences to merge prediction heatmaps or feature representations. Instead of post-prediction merge/calibration, here we introduce a transformer framework for multi-view 3D pose estimation, aiming at directly improving individual 2D predictors by integrating information from different views. Inspired by previous multi-modal transformers, we design a unified transformer architecture, named TransFusion, to fuse cues from both current views and neighboring views. Moreover, we propose the concept of epipolar field to encode 3D positional information into the transformer model. The 3D position encoding guided by the epipolar field provides an efficient way of encoding correspondences between pixels of different views. Experiments on Human 3.6M and Ski-Pose show that our method is more efficient and has consistent improvements compared to other fusion methods. Specifically, we achieve 25.8 mm MPJPE on Human 3.6M with only 5M parameters on 256 x 256 resolution.

Rui Su, Wenjing Huang, Haoyu Ma et al.

Recently, deep learning based video object detection has attracted more and more attention. Compared with object detection of static images, video object detection is more challenging due to the motion of objects, while providing rich temporal information. The RNN-based algorithm is an effective way to enhance detection performance in videos with temporal information. However, most studies in this area only focus on accuracy while ignoring the calculation cost and the number of parameters. In this paper, we propose an efficient method that combines channel-reduced convolutional GRU (Squeezed GRU), and Information Entropy map for video object detection (SGE-Net). The experimental results validate the accuracy improvement, computational savings of the Squeezed GRU, and superiority of the information entropy attention mechanism on the classification performance. The mAP has increased by 3.7 contrasted with the baseline, and the number of parameters has decreased from 6.33 million to 0.67 million compared with the standard GRU.

Haoyu Ma, Tianlong Chen, Ting-Kuei Hu et al.

Knowledge Distillation (KD) is a widely used technique to transfer knowledge from pre-trained teacher models to (usually more lightweight) student models. However, in certain situations, this technique is more of a curse than a blessing. For instance, KD poses a potential risk of exposing intellectual properties (IPs): even if a trained machine learning model is released in 'black boxes' (e.g., as executable software or APIs without open-sourcing code), it can still be replicated by KD through imitating input-output behaviors. To prevent this unwanted effect of KD, this paper introduces and investigates a concept called Nasty Teacher: a specially trained teacher network that yields nearly the same performance as a normal one, but would significantly degrade the performance of student models learned by imitating it. We propose a simple yet effective algorithm to build the nasty teacher, called self-undermining knowledge distillation. Specifically, we aim to maximize the difference between the output of the nasty teacher and a normal pre-trained network. Extensive experiments on several datasets demonstrate that our method is effective on both standard KD and data-free KD, providing the desirable KD-immunity to model owners for the first time. We hope our preliminary study can draw more awareness and interest in this new practical problem of both social and legal importance.

Haoyu Ma, Xiangru Lin, Zifeng Wu et al.

Unsupervised domain adaptation (UDA) in semantic segmentation is a fundamental yet promising task relieving the need for laborious annotation works. However, the domain shifts/discrepancies problem in this task compromise the final segmentation performance. Based on our observation, the main causes of the domain shifts are differences in imaging conditions, called image-level domain shifts, and differences in object category configurations called category-level domain shifts. In this paper, we propose a novel UDA pipeline that unifies image-level alignment and category-level feature distribution regularization in a coarse-to-fine manner. Specifically, on the coarse side, we propose a photometric alignment module that aligns an image in the source domain with a reference image from the target domain using a set of image-level operators; on the fine side, we propose a category-oriented triplet loss that imposes a soft constraint to regularize category centers in the source domain and a self-supervised consistency regularization method in the target domain. Experimental results show that our proposed pipeline improves the generalization capability of the final segmentation model and significantly outperforms all previous state-of-the-arts.

Ajay Kumar Jaiswal, Haoyu Ma, Tianlong Chen et al.

Lottery Ticket Hypothesis (LTH) suggests that a dense neural network contains a sparse sub-network that can match the performance of the original dense network when trained in isolation from scratch. Most works retrain the sparse sub-network with the same training protocols as its dense network, such as initialization, architecture blocks, and training recipes. However, till now it is unclear that whether these training protocols are optimal for sparse networks. In this paper, we demonstrate that it is unnecessary for spare retraining to strictly inherit those properties from the dense network. Instead, by plugging in purposeful "tweaks" of the sparse subnetwork architecture or its training recipe, its retraining can be significantly improved than the default, especially at high sparsity levels. Combining all our proposed "tweaks" can yield the new state-of-the-art performance of LTH, and these modifications can be easily adapted to other sparse training algorithms in general. Specifically, we have achieved a significant and consistent performance gain of1.05% - 4.93% for ResNet18 on CIFAR-100 over vanilla-LTH. Moreover, our methods are shown to generalize across datasets (CIFAR10, CIFAR100, TinyImageNet) and architectures (Vgg16, ResNet-18/ResNet-34, MobileNet). All codes will be publicly available.

Deying Kong, Haoyu Ma, Xiaohui Xie

Graph Neural Networks (GNNs) generalize neural networks from applications on regular structures to applications on arbitrary graphs, and have shown success in many application domains such as computer vision, social networks and chemistry. In this paper, we extend GNNs along two directions: a) allowing features at each node to be represented by 2D spatial confidence maps instead of 1D vectors; and b) proposing an efficient operation to integrate information from neighboring nodes through 2D convolutions with different learnable kernels at each edge. The proposed SIA-GCN can efficiently extract spatial information from 2D maps at each node and propagate them through graph convolution. By associating each edge with a designated convolution kernel, the SIA-GCN could capture different spatial relationships for different pairs of neighboring nodes. We demonstrate the utility of SIA-GCN on the task of estimating hand keypoints from single-frame images, where the nodes represent the 2D coordinate heatmaps of keypoints and the edges denote the kinetic relationships between keypoints. Experiments on multiple datasets show that SIA-GCN provides a flexible and yet powerful framework to account for structural constraints between keypoints, and can achieve state-of-the-art performance on the task of hand pose estimation.

Juncheng Zhang, Qingmin Liao, Shaojun Liu et al.

Multi-focus image fusion, a technique to generate an all-in-focus image from two or more partially-focused source images, can benefit many computer vision tasks. However, currently there is no large and realistic dataset to perform convincing evaluation and comparison of algorithms in multi-focus image fusion. Moreover, it is difficult to train a deep neural network for multi-focus image fusion without a suitable dataset. In this letter, we introduce a large and realistic multi-focus dataset called Real-MFF, which contains 710 pairs of source images with corresponding ground truth images. The dataset is generated by light field images, and both the source images and the ground truth images are realistic. To serve as both a well-established benchmark for existing multi-focus image fusion algorithms and an appropriate training dataset for future development of deep-learning-based methods, the dataset contains a variety of scenes, including buildings, plants, humans, shopping malls, squares and so on. We also evaluate 10 typical multi-focus algorithms on this dataset for the purpose of illustration.

Deying Kong, Haoyu Ma, Yifei Chen et al.

In this paper, we propose a new architecture named Rotation-invariant Mixed Graphical Model Network (R-MGMN) to solve the problem of 2D hand pose estimation from a monocular RGB image. By integrating a rotation net, the R-MGMN is invariant to rotations of the hand in the image. It also has a pool of graphical models, from which a combination of graphical models could be selected, conditioning on the input image. Belief propagation is performed on each graphical model separately, generating a set of marginal distributions, which are taken as the confidence maps of hand keypoint positions. Final confidence maps are obtained by aggregating these confidence maps together. We evaluate the R-MGMN on two public hand pose datasets. Experiment results show our model outperforms the state-of-the-art algorithm which is widely used in 2D hand pose estimation by a noticeable margin.

Yifei Chen, Haoyu Ma, Deying Kong et al.

Hand pose estimation is more challenging than body pose estimation due to severe articulation, self-occlusion and high dexterity of the hand. Current approaches often rely on a popular body pose algorithm, such as the Convolutional Pose Machine (CPM), to learn 2D keypoint features. These algorithms cannot adequately address the unique challenges of hand pose estimation, because they are trained solely based on keypoint positions without seeking to explicitly model structural relationship between them. We propose a novel Nonparametric Structure Regularization Machine (NSRM) for 2D hand pose estimation, adopting a cascade multi-task architecture to learn hand structure and keypoint representations jointly. The structure learning is guided by synthetic hand mask representations, which are directly computed from keypoint positions, and is further strengthened by a novel probabilistic representation of hand limbs and an anatomically inspired composition strategy of mask synthesis. We conduct extensive studies on two public datasets - OneHand 10k and CMU Panoptic Hand. Experimental results demonstrate that explicitly enforcing structure learning consistently improves pose estimation accuracy of CPM baseline models, by 1.17% on the first dataset and 4.01% on the second one. The implementation and experiment code is freely available online. Our proposal of incorporating structural learning to hand pose estimation requires no additional training information, and can be a generic add-on module to other pose estimation models.

Haoyu Ma, Qingmin Liao, Juncheng Zhang et al.

Capturing an all-in-focus image with a single camera is difficult since the depth of field of the camera is usually limited. An alternative method to obtain the all-in-focus image is to fuse several images focusing at different depths. However, existing multi-focus image fusion methods cannot obtain clear results for areas near the focused/defocused boundary (FDB). In this paper, a novel α-matte boundary defocus model is proposed to generate realistic training data with the defocus spread effect precisely modeled, especially for areas near the FDB. Based on this α-matte defocus model and the generated data, a cascaded boundary aware convolutional network termed MMF-Net is proposed and trained, aiming to achieve clearer fusion results around the FDB. More specifically, the MMF-Net consists of two cascaded sub-nets for initial fusion and boundary fusion, respectively; these two sub-nets are designed to first obtain a guidance map of FDB and then refine the fusion near the FDB. Experiments demonstrate that with the help of the new α-matte boundary defocus model, the proposed MMF-Net outperforms the state-of-the-art methods both qualitatively and quantitatively.

Haoyu Ma, Juncheng Zhang, Shaojun Liu et al.

Since it is usually difficult to capture an all-in-focus image of a 3D scene directly, various multi-focus image fusion methods are employed to generate it from several images focusing at different depths. However, the performance of existing methods is barely satisfactory and often degrades for areas near the focused/defocused boundary (FDB). In this paper, a boundary aware method using deep neural network is proposed to overcome this problem. (1) Aiming to acquire improved fusion images, a 2-channel deep network is proposed to better extract the relative defocus information of the two source images. (2) After analyzing the different situations for patches far away from and near the FDB, we use two networks to handle them respectively. (3) To simulate the reality more precisely, a new approach of dataset generation is designed. Experiments demonstrate that the proposed method outperforms the state-of-the-art methods, both qualitatively and quantitatively.