Xin Jin

Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson et al. · microsoft-research

We introduce Phi-4-Mini and Phi-4-Multimodal, compact yet highly capable language and multimodal models. Phi-4-Mini is a 3.8-billion-parameter language model trained on high-quality web and synthetic data, significantly outperforming recent open-source models of similar size and matching the performance of models twice its size on math and coding tasks requiring complex reasoning. This achievement is driven by a carefully curated synthetic data recipe emphasizing high-quality math and coding datasets. Compared to its predecessor, Phi-3.5-Mini, Phi-4-Mini features an expanded vocabulary size of 200K tokens to better support multilingual applications, as well as group query attention for more efficient long-sequence generation. Phi-4-Multimodal is a multimodal model that integrates text, vision, and speech/audio input modalities into a single model. Its novel modality extension approach leverages LoRA adapters and modality-specific routers to allow multiple inference modes combining various modalities without interference. For example, it now ranks first in the OpenASR leaderboard to date, although the LoRA component of the speech/audio modality has just 460 million parameters. Phi-4-Multimodal supports scenarios involving (vision + language), (vision + speech), and (speech/audio) inputs, outperforming larger vision-language and speech-language models on a wide range of tasks. Additionally, we experiment to further train Phi-4-Mini to enhance its reasoning capabilities. Despite its compact 3.8-billion-parameter size, this experimental version achieves reasoning performance on par with or surpassing significantly larger models, including DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B.

Ruoyu Feng, Xin Jin, Zongyu Guo et al. · microsoft-research

Image Coding for Machines (ICM) aims to compress images for AI tasks analysis rather than meeting human perception. Learning a kind of feature that is both general (for AI tasks) and compact (for compression) is pivotal for its success. In this paper, we attempt to develop an ICM framework by learning universal features while also considering compression. We name such features as omnipotent features and the corresponding framework as Omni-ICM. Considering self-supervised learning (SSL) improves feature generalization, we integrate it with the compression task into the Omni-ICM framework to learn omnipotent features. However, it is non-trivial to coordinate semantics modeling in SSL and redundancy removing in compression, so we design a novel information filtering (IF) module between them by co-optimization of instance distinguishment and entropy minimization to adaptively drop information that is weakly related to AI tasks (e.g., some texture redundancy). Different from previous task-specific solutions, Omni-ICM could directly support AI tasks analysis based on the learned omnipotent features without joint training or extra transformation. Albeit simple and intuitive, Omni-ICM significantly outperforms existing traditional and learning-based codecs on multiple fundamental vision tasks.

Mingde Yao, Jie Huang, Xin Jin et al.

Lightness adaptation is vital to the success of image processing to avoid unexpected visual deterioration, which covers multiple aspects, e.g., low-light image enhancement, image retouching, and inverse tone mapping. Existing methods typically work well on their trained lightness conditions but perform poorly in unknown ones due to their limited generalization ability. To address this limitation, we propose a novel generalized lightness adaptation algorithm that extends conventional normalization techniques through a channel filtering design, dubbed Channel Selective Normalization (CSNorm). The proposed CSNorm purposely normalizes the statistics of lightness-relevant channels and keeps other channels unchanged, so as to improve feature generalization and discrimination. To optimize CSNorm, we propose an alternating training strategy that effectively identifies lightness-relevant channels. The model equipped with our CSNorm only needs to be trained on one lightness condition and can be well generalized to unknown lightness conditions. Experimental results on multiple benchmark datasets demonstrate the effectiveness of CSNorm in enhancing the generalization ability for the existing lightness adaptation methods. Code is available at https://github.com/mdyao/CSNorm.

Yinmin Zhong, Zili Zhang, Bingyang Wu et al.

We present RLHFuse, an efficient training system with stage fusion for Reinforcement Learning from Human Feedback (RLHF). Due to the intrinsic nature of RLHF training, i.e., the data skewness in the generation stage and the pipeline bubbles in the training stage, existing RLHF systems suffer from low GPU utilization. RLHFuse breaks the traditional view of RLHF workflow as a composition of individual tasks, splitting each task into finer-grained subtasks, and performing stage fusion to improve GPU utilization. RLHFuse contains two key ideas. First, for generation and inference tasks, RLHFuse splits them into sample-level subtasks, enabling efficient inter-stage fusion to overlap the execution of generation and inference stages, thus mitigating the original generation bottleneck dominated by long-tailed samples. Second, for training tasks, RLHFuse breaks them into subtasks of micro-batches and performs intra-stage fusion to concurrently execute these subtasks in the training stage with a fused pipeline schedule, effectively mitigating the pipeline bubbles. The experiments show that RLHFuse increases the training throughput by up to $3.7\times$, compared to existing systems.

Baao Xie, Qiuyu Chen, Yunnan Wang et al.

Disentangled representation learning (DRL) aims to identify and decompose underlying factors behind observations, thus facilitating data perception and generation. However, current DRL approaches often rely on the unrealistic assumption that semantic factors are statistically independent. In reality, these factors may exhibit correlations, which off-the-shelf solutions have yet to properly address. To tackle this challenge, we introduce a bidirectional weighted graph-based framework, to learn factorized attributes and their interrelations within complex data. Specifically, we propose a $β$-VAE based module to extract factors as the initial nodes of the graph, and leverage the multimodal large language model (MLLM) to discover and rank latent correlations, thereby updating the weighted edges. By integrating these complementary modules, our model successfully achieves fine-grained, practical and unsupervised disentanglement. Experiments demonstrate our method's superior performance in disentanglement and reconstruction. Furthermore, the model inherits enhanced interpretability and generalizability from MLLMs.

Wei Zhao, Yijun Wang, Tianyu He et al.

The creation of lifelike speech-driven 3D facial animation requires a natural and precise synchronization between audio input and facial expressions. However, existing works still fail to render shapes with flexible head poses and natural facial details (e.g., wrinkles). This limitation is mainly due to two aspects: 1) Collecting training set with detailed 3D facial shapes is highly expensive. This scarcity of detailed shape annotations hinders the training of models with expressive facial animation. 2) Compared to mouth movement, the head pose is much less correlated to speech content. Consequently, concurrent modeling of both mouth movement and head pose yields the lack of facial movement controllability. To address these challenges, we introduce VividTalker, a new framework designed to facilitate speech-driven 3D facial animation characterized by flexible head pose and natural facial details. Specifically, we explicitly disentangle facial animation into head pose and mouth movement and encode them separately into discrete latent spaces. Then, these attributes are generated through an autoregressive process leveraging a window-based Transformer architecture. To augment the richness of 3D facial animation, we construct a new 3D dataset with detailed shapes and learn to synthesize facial details in line with speech content. Extensive quantitative and qualitative experiments demonstrate that VividTalker outperforms state-of-the-art methods, resulting in vivid and realistic speech-driven 3D facial animation.

Ruoyu Feng, Tao Yu, Xin Jin et al.

In recent studies on domain adaptation, significant emphasis has been placed on the advancement of learning shared knowledge from a source domain to a target domain. Recently, the large vision-language pre-trained model, i.e., CLIP has shown strong ability on zero-shot recognition, and parameter efficient tuning can further improve its performance on specific tasks. This work demonstrates that a simple domain prior boosts CLIP's zero-shot recognition in a specific domain. Besides, CLIP's adaptation relies less on source domain data due to its diverse pre-training dataset. Furthermore, we create a benchmark for zero-shot adaptation and pseudo-labeling based self-training with CLIP. Last but not least, we propose to improve the task generalization ability of CLIP from multiple unlabeled domains, which is a more practical and unique scenario. We believe our findings motivate a rethinking of domain adaptation benchmarks and the associated role of related algorithms in the era of CLIP.

Yunhui Zeng, Hongkun Cao, Xin Jin

Metasurfaces, capable of manipulating light at subwavelength scales, hold great potential for advancing optoelectronic applications. Generative models, particularly Generative Adversarial Networks (GANs), offer a promising approach for metasurface inverse design by efficiently navigating complex design spaces and capturing underlying data patterns. However, existing generative models struggle to achieve high electromagnetic fidelity and structural diversity. These challenges arise from the lack of explicit electromagnetic constraints during training, which hinders accurate structure-to-electromagnetic response mapping, and the absence of mechanisms to handle one-to-many mappings dilemma, resulting in insufficient structural diversity. To address these issues, we propose the Anchor-controlled Generative Adversarial Network (AcGAN), a novel framework that improves both electromagnetic fidelity and structural diversity. To achieve high electromagnetic fidelity, AcGAN proposes the Spectral Overlap Coefficient (SOC) for precise spectral fidelity assessment and develops AnchorNet, which provides real-time feedback on electromagnetic performance to refine the structure-to-electromagnetic mapping. To enhance structural diversity, AcGAN incorporates a cluster-guided controller that refines input processing and ensures multi-level spectral integration, guiding the generation process to explore multiple configurations for the same spectral target. Additionally, a dynamic loss function progressively shifts the focus from data-driven learning to optimizing both spectral fidelity and structural diversity. Empirical analysis shows that AcGAN reduces the Mean Squared Error (MSE) by 73% compared to current state-of-the-art GANs methods and significantly expands the design space to generate diverse metasurface architectures that meet precise spectral demands.

Ruoyu Feng, Yunpeng Qi, Jinming Liu et al.

Image compression methods are usually optimized isolatedly for human perception or machine analysis tasks. We reveal fundamental commonalities between these objectives: preserving accurate semantic information is paramount, as it directly dictates the integrity of critical information for intelligent tasks and aids human understanding. Concurrently, enhanced perceptual quality not only improves visual appeal but also, by ensuring realistic image distributions, benefits semantic feature extraction for machine tasks. Based on this insight, we propose Diff-ICMH, a generative image compression framework aiming for harmonizing machine and human vision in image compression. It ensures perceptual realism by leveraging generative priors and simultaneously guarantees semantic fidelity through the incorporation of Semantic Consistency loss (SC loss) during training. Additionally, we introduce the Tag Guidance Module (TGM) that leverages highly semantic image-level tags to stimulate the pre-trained diffusion model's generative capabilities, requiring minimal additional bit rates. Consequently, Diff-ICMH supports multiple intelligent tasks through a single codec and bitstream without any task-specific adaptation, while preserving high-quality visual experience for human perception. Extensive experimental results demonstrate Diff-ICMH's superiority and generalizability across diverse tasks, while maintaining visual appeal for human perception. Code is available at: https://github.com/RuoyuFeng/Diff-ICMH.

Xin Jin, Pengyi Jiao, Zheng-Peng Duan et al.

Volumetric rendering based methods, like NeRF, excel in HDR view synthesis from RAWimages, especially for nighttime scenes. While, they suffer from long training times and cannot perform real-time rendering due to dense sampling requirements. The advent of 3D Gaussian Splatting (3DGS) enables real-time rendering and faster training. However, implementing RAW image-based view synthesis directly using 3DGS is challenging due to its inherent drawbacks: 1) in nighttime scenes, extremely low SNR leads to poor structure-from-motion (SfM) estimation in distant views; 2) the limited representation capacity of spherical harmonics (SH) function is unsuitable for RAW linear color space; and 3) inaccurate scene structure hampers downstream tasks such as refocusing. To address these issues, we propose LE3D (Lighting Every darkness with 3DGS). Our method proposes Cone Scatter Initialization to enrich the estimation of SfM, and replaces SH with a Color MLP to represent the RAW linear color space. Additionally, we introduce depth distortion and near-far regularizations to improve the accuracy of scene structure for downstream tasks. These designs enable LE3D to perform real-time novel view synthesis, HDR rendering, refocusing, and tone-mapping changes. Compared to previous volumetric rendering based methods, LE3D reduces training time to 1% and improves rendering speed by up to 4,000 times for 2K resolution images in terms of FPS. Code and viewer can be found in https://github.com/Srameo/LE3D .

Kecheng Zheng, Yifei Zhang, Wei Wu et al.

Language-image pre-training largely relies on how precisely and thoroughly a text describes its paired image. In practice, however, the contents of an image can be so rich that well describing them requires lengthy captions (e.g., with 10 sentences), which are usually missing in existing datasets. Consequently, there are currently no clear evidences on whether and how language-image pre-training could benefit from long captions. To figure this out, we first re-caption 30M images with detailed descriptions using a pre-trained Multi-modality Large Language Model (MLLM), and then study the usage of the resulting captions under a contrastive learning framework. We observe that, each sentence within a long caption is very likely to describe the image partially (e.g., an object). Motivated by this, we propose to dynamically sample sub-captions from the text label to construct multiple positive pairs, and introduce a grouping loss to match the embeddings of each sub-caption with its corresponding local image patches in a self-supervised manner. Experimental results on a wide rage of downstream tasks demonstrate the consistent superiority of our method, termed DreamLIP, over previous alternatives, highlighting its fine-grained representational capacity. It is noteworthy that, on the tasks of image-text retrieval and semantic segmentation, our model trained with 30M image-text pairs achieves on par or even better performance than CLIP trained with 400M pairs. Project page is available at https://zyf0619sjtu.github.io/dream-lip.

Zekun Qi, Wenyao Zhang, Yufei Ding et al. · pku, stanford

While spatial reasoning has made progress in object localization relationships, it often overlooks object orientation-a key factor in 6-DoF fine-grained manipulation. Traditional pose representations rely on pre-defined frames or templates, limiting generalization and semantic grounding. In this paper, we introduce the concept of semantic orientation, which defines object orientations using natural language in a reference-frame-free manner (e.g., the "plug-in" direction of a USB or the "handle" direction of a cup). To support this, we construct OrienText300K, a large-scale dataset of 3D objects annotated with semantic orientations, and develop PointSO, a general model for zero-shot semantic orientation prediction. By integrating semantic orientation into VLM agents, our SoFar framework enables 6-DoF spatial reasoning and generates robotic actions. Extensive experiments demonstrated the effectiveness and generalization of our SoFar, e.g., zero-shot 48.7% successful rate on Open6DOR and zero-shot 74.9% successful rate on SIMPLER-Env.

Wenyao Zhang, Hongsi Liu, Zekun Qi et al.

Recent advances in vision-language-action (VLA) models have shown promise in integrating image generation with action prediction to improve generalization and reasoning in robot manipulation. However, existing methods are limited to challenging image-based forecasting, which suffers from redundant information and lacks comprehensive and critical world knowledge, including dynamic, spatial and semantic information. To address these limitations, we propose DreamVLA, a novel VLA framework that integrates comprehensive world knowledge forecasting to enable inverse dynamics modeling, thereby establishing a perception-prediction-action loop for manipulation tasks. Specifically, DreamVLA introduces a dynamic-region-guided world knowledge prediction, integrated with the spatial and semantic cues, which provide compact yet comprehensive representations for action planning. This design aligns with how humans interact with the world by first forming abstract multimodal reasoning chains before acting. To mitigate interference among the dynamic, spatial and semantic information during training, we adopt a block-wise structured attention mechanism that masks their mutual attention, preventing information leakage and keeping each representation clean and disentangled. Moreover, to model the conditional distribution over future actions, we employ a diffusion-based transformer that disentangles action representations from shared latent features. Extensive experiments on both real-world and simulation environments demonstrate that DreamVLA achieves 76.7% success rate on real robot tasks and 4.44 average length on the CALVIN ABC-D benchmarks.

Sangmin Lee, Eunpil Park, Angel Canelo et al.

This paper reviews the NTIRE 2025 Efficient Burst HDR and Restoration Challenge, which aims to advance efficient multi-frame high dynamic range (HDR) and restoration techniques. The challenge is based on a novel RAW multi-frame fusion dataset, comprising nine noisy and misaligned RAW frames with various exposure levels per scene. Participants were tasked with developing solutions capable of effectively fusing these frames while adhering to strict efficiency constraints: fewer than 30 million model parameters and a computational budget under 4.0 trillion FLOPs. A total of 217 participants registered, with six teams finally submitting valid solutions. The top-performing approach achieved a PSNR of 43.22 dB, showcasing the potential of novel methods in this domain. This paper provides a comprehensive overview of the challenge, compares the proposed solutions, and serves as a valuable reference for researchers and practitioners in efficient burst HDR and restoration.

Qi Wang, Zhipeng Zhang, Baao Xie et al.

Training visual reinforcement learning (RL) in practical scenarios presents a significant challenge, $\textit{i.e.,}$ RL agents suffer from low sample efficiency in environments with variations. While various approaches have attempted to alleviate this issue by disentangled representation learning, these methods usually start learning from scratch without prior knowledge of the world. This paper, in contrast, tries to learn and understand underlying semantic variations from distracting videos via offline-to-online latent distillation and flexible disentanglement constraints. To enable effective cross-domain semantic knowledge transfer, we introduce an interpretable model-based RL framework, dubbed Disentangled World Models (DisWM). Specifically, we pretrain the action-free video prediction model offline with disentanglement regularization to extract semantic knowledge from distracting videos. The disentanglement capability of the pretrained model is then transferred to the world model through latent distillation. For finetuning in the online environment, we exploit the knowledge from the pretrained model and introduce a disentanglement constraint to the world model. During the adaptation phase, the incorporation of actions and rewards from online environment interactions enriches the diversity of the data, which in turn strengthens the disentangled representation learning. Experimental results validate the superiority of our approach on various benchmarks.

Jie Luo, Yuxuan Jiang, Xin Jin et al.

Semantic segmentation serves as a cornerstone of scene understanding in autonomous driving but continues to face significant challenges under complex conditions such as occlusion. Light field and LiDAR modalities provide complementary visual and spatial cues that are beneficial for robust perception; however, their effective integration is hindered by limited viewpoint diversity and inherent modality discrepancies. To address these challenges, the first multimodal semantic segmentation dataset integrating light field data and point cloud data is proposed. Based on this dataset, we proposed a multi-modal light field point-cloud fusion segmentation network(Mlpfseg), incorporating feature completion and depth perception to segment both camera images and LiDAR point clouds simultaneously. The feature completion module addresses the density mismatch between point clouds and image pixels by performing differential reconstruction of point-cloud feature maps, enhancing the fusion of these modalities. The depth perception module improves the segmentation of occluded objects by reinforcing attention scores for better occlusion awareness. Our method outperforms image-only segmentation by 1.71 Mean Intersection over Union(mIoU) and point cloud-only segmentation by 2.38 mIoU, demonstrating its effectiveness.

Jinming Liu, Zhaoyang Jia, Jiahao Li et al.

The increasing deployment of powerful Multimodal Large Language Models (MLLMs), typically hosted on cloud platforms, urgently requires effective compression techniques to efficiently transmit signal inputs (e.g., images, videos) from edge devices with minimal bandwidth usage. However, conventional image codecs are optimized for fidelity to serve the Human Visual System (HVS) and ill-suited for MLLMs, in which diverse downstream tasks are jointly considered. In this paper, we first systematically analyze the impact of compression artifacts on several mainstream MLLMs. We find that: Compression distortion unevenly impacts different-level image features, leading to varying effects on MLLMs' downstream tasks depending on their feature-level reliance. Motivated by this discovery, we propose an image Codec TAilored to MLLMs (CoTAM) designed to adaptively protect multi-level features and suit different demands of downstream tasks. The encoder leverages CLIP's shallow-layer attention to generate an importance map for bit allocation, preserving critical semantic regions. Concurrently, the decoder integrates a lightweight adapter with a multi-level loss function to ensure the faithful reconstruction both of low-level details and high-level semantic context for robust synthesis of cross-level features. Extensive experiments validate that our method achieves up to 35.99\% bitrate saving while maintaining the same performance on the MLLM tasks, outperforming previous SOTA neural codecs.

Haiqiang Wang, Ioannis Katsavounidis, Jiantong Zhou et al.

A new methodology to measure coded image/video quality using the just-noticeable-difference (JND) idea was proposed. Several small JND-based image/video quality datasets were released by the Media Communications Lab at the University of Southern California. In this work, we present an effort to build a large-scale JND-based coded video quality dataset. The dataset consists of 220 5-second sequences in four resolutions (i.e., $1920 \times 1080$, $1280 \times 720$, $960 \times 540$ and $640 \times 360$). For each of the 880 video clips, we encode it using the H.264 codec with $QP=1, \cdots, 51$ and measure the first three JND points with 30+ subjects. The dataset is called the "VideoSet", which is an acronym for "Video Subject Evaluation Test (SET)". This work describes the subjective test procedure, detection and removal of outlying measured data, and the properties of collected JND data. Finally, the significance and implications of the VideoSet to future video coding research and standardization efforts are pointed out. All source/coded video clips as well as measured JND data included in the VideoSet are available to the public in the IEEE DataPort.

Xin Jin, Xiaoyang Tan

Over the last two decades, face alignment or localizing fiducial facial points has received increasing attention owing to its comprehensive applications in automatic face analysis. However, such a task has proven extremely challenging in unconstrained environments due to many confounding factors, such as pose, occlusions, expression and illumination. While numerous techniques have been developed to address these challenges, this problem is still far away from being solved. In this survey, we present an up-to-date critical review of the existing literatures on face alignment, focusing on those methods addressing overall difficulties and challenges of this topic under uncontrolled conditions. Specifically, we categorize existing face alignment techniques, present detailed descriptions of the prominent algorithms within each category, and discuss their advantages and disadvantages. Furthermore, we organize special discussions on the practical aspects of face alignment in-the-wild, towards the development of a robust face alignment system. In addition, we show performance statistics of the state of the art, and conclude this paper with several promising directions for future research.