A dexterous hand capable of generalizable grasping objects is fundamental for the development of general-purpose embodied AI. However, previous methods focus narrowly on low-level grasp stability metrics, neglecting affordance-aware positioning and human-like poses which are crucial for downstream manipulation. To address these limitations, we propose AffordDex, a novel framework with two-stage training that learns a universal grasping policy with an inherent understanding of both motion priors and object affordances. In the first stage, a trajectory imitator is pre-trained on a large corpus of human hand motions to instill a strong prior for natural movement. In the second stage, a residual module is trained to adapt these general human-like motions to specific object instances. This refinement is critically guided by two components: our Negative Affordance-aware Segmentation (NAA) module, which identifies functionally inappropriate contact regions, and a privileged teacher-student distillation process that ensures the final vision-based policy is highly successful. Extensive experiments demonstrate that AffordDex not only achieves universal dexterous grasping but also remains remarkably human-like in posture and functionally appropriate in contact location. As a result, AffordDex significantly outperforms state-of-the-art baselines across seen objects, unseen instances, and even entirely novel categories.

Graph Transformers (GTs) have emerged as a promising graph learning tool, leveraging their all-pair connected property to effectively capture global information. To address the over-smoothing problem in deep GNNs, global attention was initially introduced, eliminating the necessity for using deep GNNs. However, through empirical and theoretical analysis, we verify that the introduced global attention exhibits severe over-smoothing, causing node representations to become indistinguishable due to its inherent low-pass filtering. This effect is even stronger than that observed in GNNs. To mitigate this, we propose PageRank Transformer (ParaFormer), which features a PageRank-enhanced attention module designed to mimic the behavior of deep Transformers. We theoretically and empirically demonstrate that ParaFormer mitigates over-smoothing by functioning as an adaptive-pass filter. Experiments show that ParaFormer achieves consistent performance improvements across both node classification and graph classification tasks on 11 datasets ranging from thousands to millions of nodes, validating its efficacy. The supplementary material, including code and appendix, can be found in https://github.com/chaohaoyuan/ParaFormer.

Guizhen Chen, Weiwen Xu, Hao Zhang, Hou Pong Chan, Chaoqun Liu, Lidong Bing, Deli Zhao, Anh Tuan Luu, Yu Rong. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

As LLMs continuously evolve, there is an urgent need for a reliable evaluation method that delivers trustworthy results promptly.Currently, static benchmarks suffer from inflexibility and unreliability, leading users to prefer human voting platforms like Chatbot Arena.However, human evaluations require significant manual effort.Therefore, we propose Auto-Arena, an innovative framework that automates the entire evaluation process using LLM-powered agents.Firstly, an LLM examiner generates questions.Then, two LLM candidates engage in a multi-round peer battle based on the questions, aiming at revealing their true performance differences.Finally, a committee of LLM judges collaboratively discusses and decides the winner, reducing bias and enhancing fairness.During the peer battles, we observe intriguing scenarios where the LLM candidates display competitive behaviors and learn from the opponents.In our extensive experiments involving 15 recent LLMs, Auto-Arena shows a 92.14% correlation with human preferences, surpassing all previous expert-annotated benchmarks without any manual efforts.Auto-Arena offers a promising alternative to current human evaluation platforms for evaluating LLMs automatically. 1

Contrastive loss is a powerful approach for representation learning, where larger batch sizes enhance performance by providing more negative samples to better distinguish between similar and dissimilar data. However, the full instantiation of the similarity matrix demands substantial GPU memory, making large batch training highly resource-intensive. To address this, we propose a tile-based computation strategy that partitions the contrastive loss calculation into small blocks, avoiding full materialization of the similarity matrix. Additionally, we introduce a multi-level tiling implementation to leverage the hierarchical structure of distributed systems, using ring-based communication at the GPU level to optimize synchronization and fused kernels at the CUDA core level to reduce I/O overhead. Experimental results show that the proposed method significantly reduces GPU memory usage in contrastive loss. For instance, it enables contrastive training of a CLIP-ViT-L/14 model with a batch size of 4M using only 8 A800 80GB GPUs, without sacrificing accuracy. Compared to state-of-the-art memory-efficient solutions, it achieves a two-order-of-magnitude reduction in memory while maintaining comparable speed. The code will be made publicly available.1

Video Large Language Models (Video LLMs) have recently exhibited remarkable capabilities in general video understanding. However, they mainly focus on holistic comprehension and struggle with capturing fine-grained spatial and temporal details. Besides, the lack of high-quality object-level video instruction data and a comprehensive benchmark further hinders their advancements. To tackle these challenges, we introduce the VideoRefer Suite to empower Video LLM for finer-level spatial-temporal video understanding, i.e., enabling perception and reasoning on any objects throughout the video. Specially, we thoroughly develop VideoRefer Suite across three essential aspects: dataset, model, and benchmark. Firstly, we introduce a multi-agent data engine to meticulously curate a largescale, high-quality object-level video instruction dataset, termed VideoRefer-700K. Next, we present the VideoRefer model, which equips a versatile spatial-temporal object encoder to capture precise regional and sequential representations. Finally, we meticulously create a VideoRefer-Bench to comprehensively assess the spatial-temporal understanding capability of a Video LLM, evaluating it across various aspects. Extensive experiments and analyses demonstrate that our VideoRefer model not only achieves promising performance on video referring benchmarks but also facilitates general video understanding capabilities.

Yu Sun, Xingyu Qian, Weiwen Xu, Hao Zhang, Chenghao Xiao, Long Li, Deli Zhao, Wenbing Huang, Tingyang Xu, Qifeng Bai, Yu Rong. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Spatial intelligence is emerging as a transformative frontier in AI, yet it remains constrained by the scarcity of largescale 3D datasets. Unlike the abundant 2D imagery, acquiring 3D data typically requires specialized sensors and laborious annotation. In this work, we present a scalable pipeline that converts single-view images into comprehensive, scale- and appearance-realistic 3D representations - including point clouds, camera poses, depth maps, and pseudo-RGBD - via integrated depth estimation, camera calibration, and scale calibration. Our method bridges the gap between the vast repository of imagery and the increasing demand for spatial scene understanding. By automatically generating authentic, scale-aware 3D data from images, we significantly reduce data collection costs and open new avenues for advancing spatial intelligence. We release two generated spatial datasets, i.e., COCO-3D and Objects365-v2-3D, and demonstrate through extensive experiments that our generated data can benefit various 3D tasks, ranging from fundamental perception to MLLMbased reasoning. These results validate our pipeline as an effective solution for developing AI systems capable of perceiving, understanding, and interacting with physical environments.

Compared to image-text pair data, interleaved corpora enable Vision-Language Models (VLMs) to understand the world more naturally like humans. However, such existing datasets are crawled from webpage, facing challenges like low knowledge density, loose image-text relations, and poor logical coherence between images. On the other hand, the internet hosts vast instructional videos (e.g., online geometry courses) that are widely used by humans to learn foundational subjects, yet these valuable resources remain underexplored in VLM training. In this paper, we introduce a high-quality \textbf{multimodal textbook} corpus with richer foundational knowledge for VLM pretraining. It collects over 2.5 years of instructional videos, totaling 22,000 class hours. We first use an LLM-proposed taxonomy to systematically gather instructional videos. Then we progressively extract and refine visual (keyframes), audio (ASR), and textual knowledge (OCR) from the videos, and organize as an image-text interleaved corpus based on temporal order. Compared to its counterparts, our video-centric textbook offers more coherent context, richer knowledge, and better image-text alignment. Experiments demonstrate its superb pretraining performance, particularly in knowledge- and reasoning-intensive tasks like ScienceQA and MathVista. Moreover, VLMs pre-trained on our textbook exhibit outstanding interleaved context awareness, leveraging visual and textual cues in their few-shot context for task solving. Our code are available at https://github.com/DAMO-NLP-SG/multimodal_textbook.

Establishing the relationship between 3D structures and the energy states of molecular systems has proven to be a promising approach for learning 3D molecular representations. However, existing methods are limited to modeling the molecular energy states from classical mechanics. This limitation results in a significant oversight of quantum mechanical effects, such as quantized (discrete) energy level structures, which offer a more accurate estimation of molecular energy and can be experimentally measured through energy spectra. In this paper, we propose to utilize the energy spectra to enhance the pre-training of 3D molecular representations (MolSpectra), thereby infusing the knowledge of quantum mechanics into the molecular representations. Specifically, we propose SpecFormer, a multi-spectrum encoder for encoding molecular spectra via masked patch reconstruction. By further aligning outputs from the 3D encoder and spectrum encoder using a contrastive objective, we enhance the 3D encoder's understanding of molecules. Evaluations on public benchmarks reveal that our pre-trained representations surpass existing methods in predicting molecular properties and modeling dynamics.

Accurate Subseasonal-to-Seasonal (S2S) climate forecasting is pivotal for decision-making including agriculture planning and disaster preparedness but is known to be challenging due to its chaotic nature. Although recent data-driven models have shown promising results, their performance is limited by inadequate consideration of geometric inductive biases. Usually, they treat the spherical weather data as planar images, resulting in an inaccurate representation of locations and spatial relations. In this work, we propose the geometric-inspired Circular Transformer (CirT) to model the cyclic characteristic of the graticule, consisting of two key designs: (1) Decomposing the weather data by latitude into circular patches that serve as input tokens to the Transformer; (2) Leveraging Fourier transform in self-attention to capture the global information and model the spatial periodicity. Extensive experiments on the Earth Reanalysis 5 (ERA5) reanalysis dataset demonstrate our model yields a significant improvement over the advanced data-driven models, including PanguWeather and GraphCast, as well as skillful ECMWF systems. Additionally, we empirically show the effectiveness of our model designs and high-quality prediction over spatial and temporal dimensions.

Accurately predicting 3D structures and dynamics of physical systems is crucial in scientific applications. Existing approaches that rely on geometric Graph Neural Networks (GNNs) effectively enforce $\mathrm{E}(3)$-equivariance, but they often fail in leveraging extensive broader information. While direct application of Large Language Models (LLMs) can incorporate external knowledge, they lack the capability for spatial reasoning with guaranteed equivariance. In this paper, we propose EquiLLM, a novel framework for representing 3D physical systems that seamlessly integrates $\mathrm{E}(3)$-equivariance with LLM capabilities. Specifically, EquiLLM comprises four key components: geometry-aware prompting, an equivariant encoder, an LLM, and an equivariant adapter. Essentially, the LLM guided by the instructive prompt serves as a sophisticated invariant feature processor, while 3D directional information is exclusively handled by the equivariant encoder and adapter modules. Experimental results demonstrate that EquiLLM delivers significant improvements over previous methods across molecular dynamics simulation, human motion simulation, and antibody design, highlighting its promising generalizability.

Accurate weather forecasting on Subseasonal-to-Seasonal (S2S) timescale is critical to human society such as agriculture planning and extreme weather preparation. Although data-driven models have become alternatives to computationally intensive Numerical Weather Prediction (NWP) systems, existing Transformer-based approaches suffer from biases due to planar projections distorting the spherical geometry and inadequate handling of vector-scalar variable interactions (e.g., wind velocity vs. temperature). To address these limitations, we propose a graph-based Equivariant and Invariant Message Passing (EIMP) framework that directly processes spherical grid data. It maintains SO(3) equivariant embeddings for vector data and SO(3) invariant embeddings for scalar data, which are interacted by a shared invariant message embedding. Guaranteed equivariant and invariant message aggregation functions are proposed to update embeddings under strict symmetry constraints. Extensive experiments on the Earth Reanalysis 5 (ERA5) reanalysis dataset of 41 years demonstrate the proposed model achieves significant improvement over advanced data-driven models and skillful numerical ECMWF systems. Additionally, we empirically show that EIMP demonstrates geometrically superior predictions and conduct ablation studies to validate the efficacy of its design.

Equivariant Graph Neural Networks (GNNs) have demonstrated significant success across various applications. To achieve completeness---that is, the universal approximation property over the space of equivariant functions---the network must effectively capture the intricate multi-body interactions among different nodes. Prior methods attain this via deeper architectures, augmented body orders, or increased degrees of steerable features, often at high computational cost and without polynomial-time solutions. In this work, we present a theoretically grounded framework for constructing complete equivariant GNNs that is both efficient and practical. We prove that a complete equivariant GNN can be achieved through two key components: 1) a complete scalar function, referred to as the canonical form of the geometric graph; and 2) a full-rank steerable basis set. Leveraging this finding, we propose an efficient algorithm for constructing complete equivariant GNNs based on two common models: EGNN and TFN. Empirical results demonstrate that our model demonstrates superior completeness and excellent performance with only a few layers, thereby significantly reducing computational overhead while maintaining strong practical efficacy.

Recent advancements in large multimodal models (LMMs) have significantly enhanced performance across diverse tasks, with ongoing efforts to further integrate additional modalities such as video and audio. However, most existing LMMs remain vulnerable to hallucinations, the discrepancy between the factual multimodal input and the generated textual output, which has limited their applicability in various real-world scenarios. This paper presents the first systematic investigation of hallucinations in LMMs involving the three most common modalities: language, visual, and audio. Our study reveals two key contributors to hallucinations: overreliance on unimodal priors and spurious inter-modality correlations. To address these challenges, we introduce the benchmark The Curse of Multi-Modalities (CMM), which comprehensively evaluates hallucinations in LMMs, providing a detailed analysis of their underlying issues. Our findings highlight key vulnerabilities, including imbalances in modality integration and biases from training data, underscoring the need for balanced cross-modal learning and enhanced hallucination mitigation strategies. Based on our observations and findings, we suggest potential research directions that could enhance the reliability of LMMs.

The emergence of multimodal large language models (MLLMs) has driven breakthroughs in egocentric vision applications. These applications necessitate persistent, context-aware understanding of objects, as users interact with tools in dynamic and cluttered environments. However, existing embodied benchmarks primarily focus on static scene exploration, emphasizing object's appearance and spatial attributes while neglecting the assessment of dynamic changes arising from users' interactions.capabilities in object-level spatiotemporal reasoning required for real-world interactions.To address this gap, we introduce EOC-Bench, an innovative benchmark designed to systematically evaluate object-centric embodied cognition in dynamic egocentric scenarios.Specially, EOC-Bench features 3,277 meticulously annotated QA pairs categorized into three temporal categories: Past, Present, and Future, covering 11 fine-grained evaluation dimensions and 3 visual object referencing types.To ensure thorough assessment, we develop a mixed-format human-in-the-loop annotation frameworkBased on EOC-Bench, we conduct comprehensive evaluations of various proprietary, open-source, and object-level MLLMs. EOC-Bench serves as a crucial tool for advancing the embodied object cognitive capabilities of MLLMs, establishing a robust foundation for developing reliable core models for embodied systems.

To enhance the generalization ability of graph neural networks (GNNs) in learning and simulation physical dynamics, a series of equivariant GNNs have been developed to incorporate the symmetric inductive bias. However, the existing methods do not take into account the non-stationarity nature of physical dynamics, where the joint distribution changes over time. Moreover, previous approaches for modeling non-stationary time series typically involve normalizing the data, which disrupts the symmetric assumption inherent in physical dynamics. To model the non-stationary physical dynamics while preserving the symmetric inductive bias, we introduce a Non-Stationary Equivariant Graph Neural Network (NS-EGNN) to capture the non-stationarity in physical dynamics while preserving the symmetric property of the model. Specifically, NS-EGNN employs Fourier Transform on segments of physical dynamics to extract time-varying frequency information from the trajectories. It then uses the first and second-order differences to mitigate non-stationarity, followed by pooling for future predictions. Through capturing varying frequency characteristics and alleviate the linear and quadric trend in the raw physical dynamics, NS-EGNN better models the temporal dependencies in the physical dynamics. NS-EGNN has been applied on various types of physical dynamics, including molecular, motion and protein dynamics. In various scenario, NS-EGNN consistently surpasses the performance of existing state-of-the-art algorithms, underscoring its effectiveness. The implementation of NS-EGNN is available at https://github.com/MaojiWEN/NS-EGNN.

Abstract Geometric graphs are a special kind of graph with geometric features, which are vital to model many scientific problems. Unlike generic graphs, geometric graphs often exhibit physical symmetries of translations, rotations, and reflections, making them ineffectively processed by current Graph Neural Networks (GNNs). To address this issue, researchers proposed a variety of geometric GNNs equipped with invariant/equivariant properties to better characterize the geometry and topology of geometric graphs. Given the current progress in this field, it is imperative to conduct a comprehensive survey of data structures, models, and applications related to geometric GNNs. In this paper, based on the necessary but concise mathematical preliminaries, we formalize geometric graph as the data structure, on top of which we provide a unified view of existing models from the geometric message passing perspective. Additionally, we summarize the applications as well as the related datasets to facilitate later research for methodology development and experimental evaluation. We also discuss the challenges and future potential directions of geometric GNNs at the end of this survey.

Query optimization is a complex planning and decision-making problem within the exponentially growing plan space in database management systems (DBMS). Traditional optimization techniques have been extensively studied over decades, leaving limited room for further improvement along this track. Recent developments of Large Language Models (LLMs) have demonstrated their potential in solving complex planning and decision-making problems, such as arithmetic and programmatic tasks. In this paper, we try to explore the potential of LLMs in handling query optimization and propose a tentative LLM-based query optimizer dubbed LLM-QO, established on PostgreSQL's execution engine. In LLM-QO, we formulate query optimization in an autoregressive fashion which directly generates the execution plan without explicit plan enumeration. To investigate the essential input of LLM-QO, we design a customized data recipe named QInstruct to collect the training data from various optimizers and serialize the database's meta data, queries and corresponding plans into a textual format. Based on QInstruct, we implement a two-stage fine-tuning pipeline, Query Instruction Tuning (QIT) and Query Direct Preference Optimization (QDPO), to empower the capability of general-purpose LLMs in handling query optimization. In our experiments, LLM-QO can generate valid and high-quality plans and consistently outperforms both traditional and learned optimizers on three query workloads. Our findings verify that LLMs can be derived as query optimizers where generalization, efficiency and adaptivity deserve further research efforts.

This work presents AnyDoor, a diffusion-based image generator with the power to teleport target objects to new scenes at user-specified locations with desired shapes. Instead of tuning parameters for each object, our model is trained only once and effortlessly generalizes to diverse object-scene combinations at the inference stage. Such a challenging zero-shot setting requires an adequate characterization of a certain object. To this end, we leverage the powerful self-supervised image encoder (i.e., DINOv2) to extract the discriminative dentity feature of the target object. Besides, we complement the identity feature with detail features, which are carefully designed to maintain appearance details yet allow versatile local variations (e.g., lighting, orientation, posture, etc.), supporting the object in favorably blending with different surroundings. We further propose to borrow knowledge from video datasets, where we can observe various forms (i.e., along the time axis) of a single object, leading to stronger model generalizability and robustness. Starting from the task of object insertion, we further extend the framework of AnyDoor to a general solution with region-to-region image reference. With the different definitions of the source region and target region, the tasks of object insertion, object removal, and image variation could be integrated into one model without introducing extra parameters. In addition, we investigate incorporating other conditions like the mask, pose skeleton, and depth map as additional guidance to achieve more controllable generation.

This paper strives for image editing via generative models. Flow Matching is an emerging generative modeling technique that offers the advantage of simple and efficient training. Simultaneously, a new transformer-based U-ViT has recently been proposed to replace the commonly used UNet for better scalability and performance in generative modeling. Hence, Flow Matching with a transformer backbone offers the potential for scalable and high-quality generative modeling, but their latent structure and editing ability are as of yet unknown. Hence, we adopt this setting and explore how to edit images through latent space manipulation. We introduce an editing space, which we call u-space, that can be manipulated in a controllable, accumulative, and composable manner. Additionally, we propose a tailored sampling solution to enable sampling with the more efficient adaptive step-size ODE solvers. Lastly, we put forth a straightforward yet powerful method for achieving fine-grained and nuanced editing using text prompts. Our framework is simple and efficient, all while being highly effective at editing images while preserving the essence of the original content. Our code will be publicly available at https://taohu.me/lfm/

This work presents AnyDoor, a diffusion-based image generator with the power to teleport target objects to new scenes at user-specified locations with desired shapes. Instead of tuning parameters for each object, our model is trained only once and effortlessly generalizes to diverse object-scene combinations at the inference stage. Such a challenging zero-shot setting requires an adequate characterization of a certain object. To this end, we complement the commonly used identity feature with detail features, which are carefully designed to maintain appearance details yet allow versatile local variations (e.g., lighting, orientation, posture, etc.), supporting the object in favorably blending with different surroundings. We further propose to borrow knowledge from video datasets, where we can observe various forms (i.e., along the time axis) of a single object, leading to stronger model generalizability and robustness. Extensive experiments demonstrate the superiority of our approach over existing alternatives as well as its great potential in real-world applications, such as virtual try-on, shape editing, and object swapping. Code is released at github.com/ali-vilab/AnyDoor.

A diffusion model, which is formulated to produce an image using thousands of denoising steps, usually suffers from a slow inference speed. Existing acceleration algorithms simplify the sampling by skipping most steps yet exhibit considerable performance degradation. By viewing the generation of diffusion models as a discretized integral process, we argue that the quality drop is partly caused by applying an inaccurate integral direction to a timestep interval. To rectify this issue, we propose a timestep tuner that helps find a more accurate integral direction for a particular interval at the minimum cost. Specifically, at each denoising step, we replace the original parameterization by conditioning the network on a new timestep, enforcing the sampling distribution towards the real one. Extensive experiments show that our plug-in design can be trained efficiently and boost the inference performance of various state-of-the-art acceleration methods, especially when there are few denoising steps. For example, when using 10 denoising steps on LSUN Bedroom dataset, we improve the FID of DDIM from 9.65 to 6.07, simply by adopting our method for a more appropriate set of timesteps. Code is available at https://github.com/THU-LYJ-Lab/time-tuner.

Crystals are the foundation of numerous scientific and industrial applications. While various learning-based approaches have been proposed for crystal generation, existing methods neglect the spacegroup constraint which is crucial in describing the geometry of crystals and closely relevant to many desirable properties. However, considering spacegroup constraint is challenging owing to its diverse and nontrivial forms. In this paper, we reduce the spacegroup constraint into an equivalent formulation that is more tractable to be handcrafted into the generation process. In particular, we translate the spacegroup constraint into two cases: the basis constraint of the invariant exponential space of the lattice matrix and the Wyckoff position constraint of the fractional coordinates. Upon the derived constraints, we then propose DiffCSP++, a novel diffusion model that has enhanced a previous work DiffCSP by further taking spacegroup constraint into account. Experiments on several popular datasets verify the benefit of the involvement of the spacegroup constraint, and show that our DiffCSP++ achieves the best or comparable performance on crystal structure prediction and ab initio crystal generation.

Diffusion models, which employ stochastic differential equations to sample images through integrals, have emerged as a dominant class of generative models. However, the rationality of the diffusion process itself receives limited attention, leaving the question of whether the problem is well-posed and well-conditioned. In this paper, we uncover a vexing propensity of diffusion models: they frequently exhibit the infinite Lipschitz near the zero point of timesteps. We provide theoretical proofs to illustrate the presence of infinite Lipschitz constants and empirical results to confirm it. The Lipschitz singularities pose a threat to the stability and accuracy during both the training and inference processes of diffusion models. Therefore, the mitigation of Lipschitz singularities holds great potential for enhancing the performance of diffusion models. To address this challenge, we propose a novel approach, dubbed E-TSDM, which alleviates the Lipschitz singularities of the diffusion model near the zero point. Remarkably, our technique yields a substantial improvement in performance. Moreover, as a byproduct of our method, we achieve a dramatic reduction in the Fréchet Inception Distance of acceleration methods relying on network Lipschitz, including DDIM and DPM-Solver, by over 33\%. Extensive experiments on diverse datasets validate our theory and method. Our work may advance the understanding of the general diffusion process, and also provide insights for the design of diffusion models.

This work presents a unified knowledge protocol, called UKnow, which facilitates knowledge-based studies from the perspective of data. Particularly focusing on visual and linguistic modalities, we categorize data knowledge into five unit types, namely, in-image, in-text, cross-image, cross-text, and image-text, and set up an efficient pipeline to help construct the multimodal knowledge graph from any data collection. Thanks to the logical information naturally contained in knowledge graph, organizing datasets under UKnow format opens up more possibilities of data usage compared to the commonly used image-text pairs. Following UKnow protocol, we collect, from public international news, a large-scale multimodal knowledge graph dataset that consists of 1,388,568 nodes (with 571,791 vision-related ones) and 3,673,817 triplets. The dataset is also annotated with rich event tags, including 11 coarse labels and 9,185 fine labels. Experiments on four benchmarks demonstrate the potential of UKnow in supporting common-sense reasoning and boosting vision-language pre-training with a single dataset, benefiting from its unified form of knowledge organization. Code, dataset, and models will be made publicly available. See Appendix to download the dataset.

Generative Adversarial Networks (GANs) have significantly advanced image synthesis through mapping randomly sampled latent codes to high-fidelity synthesized images. However, applying well-trained GANs to real image editing remains challenging. A common solution is to find an approximate latent code that can adequately recover the input image to edit, which is also known as GAN inversion. To invert a GAN model, prior works typically focus on reconstructing the target image at the pixel level, yet few studies are conducted on whether the inverted result can well support manipulation at the semantic level. This work fills in this gap by proposing in-domain GAN inversion, which consists of a domain-guided encoder and a domain-regularized optimizer, to regularize the inverted code in the native latent space of the pre-trained GAN model. In this way, we manage to sufficiently reuse the knowledge learned by GANs for image reconstruction, facilitating a wide range of editing applications without any retraining. We further make comprehensive analyses on the effects of the encoder structure, the starting inversion point, as well as the inversion parameter space, and observe the trade-off between the reconstruction quality and the editing property. Such a trade-off sheds light on how a GAN model represents an image with various semantics encoded in the learned latent distribution.

This work presents two astonishing findings on neural networks learned for large-scale image classification. 1) Given a well-trained model, the logits predicted for some category can be directly obtained by linearly combining the predictions of a few other categories, which we call neural dependency. 2) Neural dependencies exist not only within a single model, but even between two independently learned models, regardless of their architectures. Towards a theoretical analysis of such phenomena, we demonstrate that identifying neural dependencies is equivalent to solving the Covariance Lasso (CovLasso) regression problem proposed in this paper. Through investigating the properties of the problem solution, we confirm that neural dependency is guaranteed by a redundant logit covariance matrix, which condition is easily met given massive categories, and that neural dependency is highly sparse, implying that one category correlates to only a few others. We further empirically show the potential of neural dependencies in understanding internal data correlations, generalizing models to unseen categories, and improving model robustness with a dependency-derived regularizer. Code to reproduce the results in this paper is available at https://github.com/RuiLiFengiNeural-Dependencies.

Current state-of-the-art approaches for few-shot action recognition achieve promising performance by conducting frame-level matching on learned visual features. However, they generally suffer from two limitations: i) the matching procedure between local frames tends to be inaccurate due to the lack of guidance to force long-range temporal perception; ii) explicit motion learning is usually ignored, leading to partial information loss. To address these issues, we develop a Motion-augmented Long-short Contrastive Learning (MoLo) method that contains two crucial components, including a long-short contrastive objective and a motion autodecoder. Specifically, the long-short contrastive objective is to endow local frame features with long-form temporal awareness by maximizing their agreement with the global token of videos belonging to the same class. The motion autodecoder is a lightweight architecture to reconstruct pixel motions from the differential features, which explicitly embeds the network with motion dynamics. By this means, MoLo can simultaneously learn long-range temporal context and motion cues for comprehensive few-shot matching. To demonstrate the effectiveness, we evaluate MoLo on five standard benchmarks, and the results show that MoLo favorably outperforms recent advanced methods. The source code is available at https://github.com/alibaba-mmai-research/MoLo.

Generating a talking face video from the input audio sequence is a practical yet challenging task. Most existing methods either fail to capture fine facial details or need to train a specific model for each identity. We argue that a codebook pre-learned on high-quality face images can serve as a useful prior that facilitates high-fidelity and generalizable talking head synthesis. Thanks to the strong capability of the codebook in representing face textures, we simplify the talking face generation task as finding proper lip-codes to characterize the variation of lips during portrait talking. To this end, we propose LipFormer, a Transformer-based framework to model the audio-visual coherence and predict the lip-codes sequence based on input audio features. We further introduce an adaptive face warping module, which helps warp the reference face to the target pose in the feature space, to alleviate the difficulty of lip-code prediction under different poses. By this means, LipFormer can make better use of prelearned priors in images and is robust to posture change. Extensive experiments show that LipFormer can produce more realistic talking face videos compared to previous methods and faithfully generalize to unseen identities.

Contrastive Language-Image Pre-training (CLIP) is attracting increasing attention for its impressive zero-shot recognition performance on different down-stream tasks. However, training CLIP is data-hungry and requires lots of image-text pairs to memorize various semantic concepts. In this paper, we propose a novel and efficient framework: Retrieval Augmented Contrastive Language-Image Pre-training (RA-CLIP) to augment embeddings by online retrieval. Specifically, we sample part of image-text data as a hold-out reference set. Given an input image, relevant image-text pairs are retrieved from the reference set to enrich the representation of input image. This process can be considered as an open-book exam: with the reference set as a cheat sheet, the proposed method doesn't need to memorize all visual concepts in the training data. It explores how to recognize visual concepts by exploiting correspondence between images and texts in the cheat sheet. The proposed RA-CLIP implements this idea and comprehensive experiments are conducted to show how RA-CLIP works. Performances on 10 image classification datasets and 2 object detection datasets show that RA-CLIP outperforms vanilla CLIP baseline by a large margin on zero-shot image classification task (+12.7%), linear probe image classification task (+6.9%) and zero-shot ROI classification task (+2.8%).

Diffusion models, which learn to reverse a signal destruction process to generate new data, typically require the signal at each step to have the same dimension. We argue that, considering the spatial redundancy in image signals, there is no need to maintain a high dimensionality in the evolution process, especially in the early generation phase. To this end, we make a theoretical generalization of the forward diffusion process via signal decomposition. Concretely, we manage to decompose an image into multiple orthogonal components and control the attenuation of each component when perturbing the image. That way, along with the noise strength increasing, we are able to diminish those inconsequential components and thus use a lower-dimensional signal to represent the source, barely losing information. Such a reformulation allows to vary dimensions in both training and inference of diffusion models. Extensive experiments on a range of datasets suggest that our approach substantially reduces the computational cost and achieves on-par or even better synthesis performance compared to baseline methods. We also show that our strategy facilitates high-resolution image synthesis and improves FID of diffusion model trained on FFHQ at$1024\times 1024$resolution from 52.40 to 10.46. Code is available at https://github.com/damo-vilab/dvdp.

Non-exemplar class-incremental learning aims to recognize both the old and new classes without access to old class samples. The conflict between old and new class optimization is exacerbated since the shared neural pathways can only be differentiated by the incremental samples. To address this problem, we propose a novel self-organizing pathway expansion scheme. Our scheme consists of a class-specific pathway organization strategy that reduces the coupling of optimization pathway among different classes to enhance the independence of the feature representation, and a pathway-guided feature optimization mechanism to mitigate the update interference between the old and new classes. Extensive experiments on four datasets demonstrate significant performance gains, outperforming the state-of-the-art methods by a margin of 1%, 3%, 2% and 2%, respectively.

Vector-quantized image modeling has shown great potential in synthesizing high-quality images. However, generating high-resolution images remains a challenging task due to the quadratic computational overhead of the self-attention process. In this study, we seek to explore a more efficient two-stage framework for high-resolution image generation with improvements in the following three aspects. (1) Based on the observation that the first quantization stage has solid local property, we employ a local attention-based quantization model instead of the global attention mechanism used in previous methods, leading to better efficiency and reconstruction quality. (2) We emphasize the importance of multi-grained feature interaction during image generation and introduce an efficient attention mechanism that combines global attention (long-range semantic consistency within the whole image) and local attention (fined-grained details). This approach results in faster generation speed, higher generation fidelity, and improved resolution. (3) We propose a new generation pipeline incorporating autoencoding training and autoregressive generation strategy, demonstrating a better paradigm for image synthesis. Extensive experiments demonstrate the superiority of our approach in high-quality and high-resolution image reconstruction and generation.

Foundation models are pre-trained on massive data and transferred to downstream tasks via fine-tuning. This work presents Vision Middleware (ViM), a new learning paradigm that targets unified transferring from a single foundation model to a variety of downstream tasks. ViM consists of a zoo of lightweight plug-in modules, each of which is independently learned on a midstream dataset with a shared frozen backbone. Downstream tasks can then benefit from an adequate aggregation of the module zoo thanks to the rich knowledge inherited from midstream tasks. There are three major advantages of such a design. From the efficiency aspect, the upstream backbone can be trained only once and reused for all downstream tasks without tuning. From the scalability aspect, we can easily append additional modules to ViM with no influence on existing modules. From the performance aspect, ViM can include as many midstream tasks as possible, narrowing the task gap between upstream and downstream. Considering these benefits, we believe that ViM, which the community could maintain and develop together, would serve as a powerful tool to assist foundation models.

The performance of text-video retrieval has been significantly improved by vision-language cross-modal learning schemes. The typical solution is to directly align the global video-level and sentence-level features during learning, which would ignore the intrinsic video-text relations, i.e., a text description only corresponds to a spatio-temporal part of videos. Hence, the matching process should consider both fine-grained spatial content and various temporal semantic events. To this end, we propose a text-video learning framework with progressive spatio-temporal prototype matching. Specifically, the matching process is decomposed into two complementary phases: object-phrase prototype matching and event-sentence prototype matching. In the object-phrase prototype matching phase, the spatial prototype generation mechanism predicts key patches or words, which are aggregated into object or phrase prototypes. Importantly, optimizing the local alignment between object-phrase prototypes helps the model perceive spatial details. In the event-sentence prototype matching phase, we design a temporal prototype generation mechanism to associate intra-frame objects and interact inter-frame temporal relations. Such progressively generated event prototypes can reveal semantic diversity in videos for dynamic matching. Validated by comprehensive experiments, our method consistently outperforms the state-of-the-art methods on four video retrieval benchmark.1

Video temporal grounding aims to pinpoint a video segment that matches the query description. Despite the recent advance in short-form videos (e.g., in minutes), temporal grounding in long videos (e.g., in hours) is still at its early stage. To address this challenge, a common practice is to employ a sliding window, yet can be inefficient and inflexible due to the limited number of frames within the window. In this work, we propose an end-to-end framework for fast temporal grounding, which is able to model an hours-long video with one-time network execution. Our pipeline is formulated in a coarse-to-fine manner, where we first extract context knowledge from non-overlapped video clips (i.e., anchors), and then supplement the anchors that highly response to the query with detailed content knowledge. Besides the remarkably high pipeline efficiency, another advantage of our approach is the capability of capturing long-range temporal correlation, thanks to modeling the entire video as a whole, and hence facilitates more accurate grounding. Experimental results suggest that, on the long-form video datasets MAD and Ego4d, our method significantly outperforms state-of-the-arts, and achieves 14.6× / 102.8× higher efficiency respectively. Project can be found at https://github.com/afcedf/SOONet.git.

Recently, prompt-based learning has made impressive progress on image class-incremental learning, but it still lacks sufficient exploration in the video domain. In this paper, we will fill this gap by learning multiple prompts based on a powerful image-language pre-trained model, i.e., CLIP, making it fit for video class-incremental learning (VCIL). For this purpose, we present a space-time prompting approach (ST-Prompt) which contains two kinds of prompts, i.e., task-specific prompts and task-agnostic prompts. The task-specific prompts are to address the catastrophic forgetting problem by learning multi-grained prompts, i.e., spatial prompts, temporal prompts and comprehensive prompts, for accurate task identification. The task-agnostic prompts maintain a globally-shared prompt pool, which can empower the pre-trained image models with temporal perception abilities by exchanging contexts between frames. By this means, ST-Prompt can transfer the plentiful knowledge in the image-language pre-trained models to the VCIL task with only a tiny set of prompts to be optimized. To evaluate ST-Prompt, we conduct extensive experiments on three standard benchmarks. The results show that ST-Prompt can significantly surpass the state-of-the-art VCIL methods, especially it gains 9.06% on HMDB51 dataset under the 1 × 25 stage setting.

Recently, large-scale pre-trained language-image models like CLIP have shown extraordinary capabilities for understanding spatial contents, but naively transferring such models to video recognition still suffers from unsatisfactory temporal modeling capabilities. Existing methods insert tunable structures into or in parallel with the pre-trained model, which either requires back-propagation through the whole pre-trained model and is thus resource-demanding, or is limited by the temporal reasoning capability of the pre-trained structure. In this work, we present DiST, which disentangles the learning of spatial and temporal aspects of videos. Specifically, DiST uses a dual-encoder structure, where a pre-trained foundation model acts as the spatial encoder, and a lightweight network is introduced as the temporal encoder. An integration branch is inserted between the encoders to fuse spatio-temporal information. The disentangled spatial and temporal learning in DiST is highly efficient because it avoids the back-propagation of massive pre-trained parameters. Meanwhile, we empirically show that disentangled learning with an extra network for integration benefits both spatial and temporal understanding. Extensive experiments on five benchmarks show that DiST delivers better performance than existing state-of-the-art methods by convincing gaps. When pre-training on the large-scale Kinetics-710, we achieve 89.7% on Kinetics-400 with a frozen ViT-L model, which verifies the scalability of DiST. Codes and models can be found in https://github.com/alibaba-mmai-research/DiST.

Relational Language-Image Pre-training (RLIP) aims to align vision representations with relational texts, thereby advancing the capability of relational reasoning in computer vision tasks. However, hindered by the slow convergence of RLIPv11architecture and the limited availability of existing scene graph data, scaling RLIPv1 is challenging. In this paper, we propose RLIPv2, a fast converging model that enables the scaling of relational pre-training to large-scale pseudo-labelled scene graph data. To enable fast scaling, RLIPv2 introduces Asymmetric Language-Image Fusion (ALIF), a mechanism that facilitates earlier and deeper gated cross-modal fusion with sparsified language encoding layers. ALIF leads to comparable or better performance than RLIPv1 in a fraction of the time for pre-training and fine-tuning. To obtain scene graph data at scale, we extend object detection datasets with free-form relation labels by introducing a captioner (e.g., BLIP) and a designed Relation Tagger. The Relation Tagger assigns BLIP-generated relation texts to region pairs, thus enabling larger-scale relational pre-training. Through extensive experiments conducted on Human-Object Interaction Detection and Scene Graph Generation, RLIPv2 shows state-of-the-art performance on three benchmarks under fully-finetuning, few-shot and zero-shot settings. Notably, the largest RLIPv2 achieves 23.29mAP on HICO-DET without any fine-tuning, yields 32.22mAP with just 1% data and yields 45.09mAP with 100% data. Code and models are publicly available at https://github.com/JacobYuan7/RLIPv2.

Prompt tuning and adapter tuning have shown great potential in transferring pre-trained vision-language models (VLMs) to various downstream tasks. In this work, we design a new type of tuning method, termed as regularized mask tuning, which masks the network parameters through a learnable selection. Inspired by neural pathways, we argue that the knowledge required by a downstream task already exists in the pre-trained weights but just gets concealed in the upstream pre-training stage. To bring the useful knowledge back into light, we first identify a set of parameters that are important to a given downstream task, then attach a binary mask to each parameter, and finally optimize these masks on the downstream data with the parameters frozen. When updating the mask, we introduce a novel gradient dropout strategy to regularize the parameter selection, in order to prevent the model from forgetting old knowledge and overfitting the downstream data. Experimental results on 11 datasets demonstrate the consistent superiority of our method over previous alternatives. It is noteworthy that we manage to deliver 18.73% performance improvement compared to the zero-shot CLIP via masking an average of only 2.56% parameters. Furthermore, our method is synergistic with most existing parameter-efficient tuning methods and can boost the performance on top of them. Project page can be found here.

This work presents an easy-to-use regularizer for GAN training, which helps explicitly link some axes of the latent space to a set of pixels in the synthesized image. Establishing such a connection facilitates a more convenient local control of GAN generation, where users can alter the image content only within a spatial area simply by partially resampling the latent code. Experimental results confirm four appealing properties of our regularizer, which we call LinkGAN. (1) The latent-pixel linkage is applicable to either a fixed region (i.e., same for all instances) or a particular semantic category (i.e., varying across instances), like the sky. (2) Two or multiple regions can be independently linked to different latent axes, which further supports joint control. (3) Our regularizer can improve the spatial controllability of both 2D and 3D-aware GAN models, barely sacrificing the synthesis performance. (4) The models trained with our regularizer are compatible with GAN inversion techniques and maintain editability on real images. Project page can be found here.

Recent large-scale generative models learned on big data are capable of synthesizing incredible images yet suffer from limited controllability. This work offers a new generation paradigm that allows flexible control of the output image, such as spatial layout and palette, while maintaining the synthesis quality and model creativity. With compositionality as the core idea, we first decompose an image into representative factors, and then train a diffusion model with all these factors as the conditions to recompose the input. At the inference stage, the rich intermediate representations work as composable elements, leading to a huge design space (i.e., exponentially proportional to the number of decomposed factors) for customizable content creation. It is noteworthy that our approach, which we call Composer, supports various levels of conditions, such as text description as the global information, depth map and sketch as the local guidance, color histogram for low-level details, etc. Besides improving controllability, we confirm that Composer serves as a general framework and facilitates a wide range of classical generative tasks without retraining. Code and models will be made available.

Human brains respond to semantic features of presented stimuli with different neurons. This raises the question of whether deep neural networks admit a similar behavior pattern. To investigate this phenomenon, this paper identifies a small cluster of neurons associated with a specific subject in a diffusion model. We call those neurons the concept neurons. They can be identified by statistics of network gradients to a stimulation connected with the given subject. The concept neurons demonstrate magnetic properties in interpreting and manipulating generation results. Shutting them can directly yield the related subject contextualized in different scenes. Concatenating multiple clusters of concept neurons can vividly generate all related concepts in a single image. Our method attains impressive performance for multi-subject customization, even four or more subjects. For large-scale applications, the concept neurons are environmentally friendly as we only need to store a sparse cluster of int index instead of dense float32 parameter values, reducing storage consumption by 90% compared with previous customized generation methods. Extensive qualitative and quantitative studies on diverse scenarios show the superiority of our method in interpreting and manipulating diffusion models.

Vision-language representation learning models (e.g., CLIP) have achieved state-of-the-art performance on various downstream tasks, which usually need large-scale training data to learn discriminative representation. Recent progress on generative diffusion models (e.g., DALL-E 2) has demonstrated that diverse high-quality samples can be synthesized by randomly sampling from generative distribution. By virtue of generative capability in this paper, we propose a novel vision-language Representation Learning method with diffusion-based Embedding Generation (RLEG), which exploits diffusion models to generate feature embedding online for learning effective vision-language representation. Specifically, we first adopt image and text encoders to extract the corresponding embeddings. Secondly, pretrained diffusion-based embedding generators are harnessed to transfer the embedding modality online between vision and language domains. The embeddings generated from the generators are then served as augmented embedding-level samples, which are applied to contrastive learning with the variant of the CLIP framework. Experimental results show that the proposed method could learn effective representation and achieve state-of-the-art performance on various tasks including image classification, image-text retrieval, object detection, semantic segmentation, and text-conditional image generation.

Parameter-efficient tuning has become a trend in transferring large-scale foundation models to downstream applications. Existing methods typically embed some light-weight tuners into the backbone, where both the design and the learning of the tuners are highly dependent on the base model. This work offers a new tuning paradigm, dubbed Res-Tuning, which intentionally unbinds tuners from the backbone. With both theoretical and empirical evidence, we show that popular tuning approaches have their equivalent counterparts under our unbinding formulation, and hence can be integrated into our framework effortlessly. Thanks to the structural disentanglement, we manage to free the design of tuners from the network architecture, facilitating flexible combination of various tuning strategies. We further propose a memory-efficient variant of Res-Tuning, where the bypass i.e., formed by a sequence of tuners) is effectively detached from the main branch, such that the gradients are back-propagated only to the tuners but not to the backbone. Such a detachment also allows one-time backbone forward for multi-task inference. Extensive experiments on both discriminative and generative tasks demonstrate the superiority of our method over existing alternatives from the perspectives of efficacy and efficiency. Project page: https://res-tuning.github.io/.

Video moment retrieval pursues an efficient and generalized solution to identify the specific temporal segments within an untrimmed video that correspond to a given language description. To achieve this goal, we provide a generative diffusion-based framework called MomentDiff, which simulates a typical human retrieval process from random browsing to gradual localization. Specifically, we first diffuse the real span to random noise, and learn to denoise the random noise to the original span with the guidance of similarity between text and video. This allows the model to learn a mapping from arbitrary random locations to real moments, enabling the ability to locate segments from random initialization. Once trained, MomentDiff could sample random temporal segments as initial guesses and iteratively refine them to generate an accurate temporal boundary. Different from discriminative works (e.g., based on learnable proposals or queries), MomentDiff with random initialized spans could resist the temporal location biases from datasets. To evaluate the influence of the temporal location biases, we propose two ``anti-bias'' datasets with location distribution shifts, named Charades-STA-Len and Charades-STA-Mom. The experimental results demonstrate that our efficient framework consistently outperforms state-of-the-art methods on three public benchmarks, and exhibits better generalization and robustness on the proposed anti-bias datasets. The code, model, and anti-bias evaluation datasets will be released publicly.

Synthesizing images with user-specified subjects has received growing attention due to its practical applications. Despite the recent success in single subject customization, existing algorithms suffer from high training cost and low success rate along with increased number of subjects. Towards controllable image synthesis with multiple subjects as the constraints, this work studies how to efficiently represent a particular subject as well as how to appropriately compose different subjects. We find that the text embedding regarding the subject token already serves as a simple yet effective representation that supports arbitrary combinations without any model tuning. Through learning a residual on top of the base embedding, we manage to robustly shift the raw subject to the customized subject given various text conditions. We then propose to employ layout, a very abstract and easy-to-obtain prior, as the spatial guidance for subject arrangement. By rectifying the activations in the cross-attention map, the layout appoints and separates the location of different subjects in the image, significantly alleviating the interference across them. Using cross-attention map as the intermediary, we could strengthen the signal of target subjects and weaken the signal of irrelevant subjects within a certain region, significantly alleviating the interference across subjects. Both qualitative and quantitative experimental results demonstrate our superiority over state-of-the-art alternatives under a variety of settings for multi-subject customization.

The pursuit of controllability as a higher standard of visual content creation has yielded remarkable progress in customizable image synthesis. However, achieving controllable video synthesis remains challenging due to the large variation of temporal dynamics and the requirement of cross-frame temporal consistency. Based on the paradigm of compositional generation, this work presents VideoComposer that allows users to flexibly compose a video with textual conditions, spatial conditions, and more importantly temporal conditions. Specifically, considering the characteristic of video data, we introduce the motion vector from compressed videos as an explicit control signal to provide guidance regarding temporal dynamics. In addition, we develop a Spatio-Temporal Condition encoder (STC-encoder) that serves as a unified interface to effectively incorporate the spatial and temporal relations of sequential inputs, with which the model could make better use of temporal conditions and hence achieve higher inter-frame consistency. Extensive experimental results suggest that VideoComposer is able to control the spatial and temporal patterns simultaneously within a synthesized video in various forms, such as text description, sketch sequence, reference video, or even simply hand-crafted motions. The code and models are publicly available at https://videocomposer.github.io.

This work presents FaceComposer, a unified generative model that accomplishes a variety of facial content creation tasks, including text-conditioned face synthesis, text-guided face editing, face animation etc. Based on the latent diffusion framework, FaceComposer follows the paradigm of compositional generation and employs diverse face-specific conditions, e.g., Identity Feature and Projected Normalized Coordinate Code, to release the model creativity at all possible. To support text control and animation, we clean up some existing face image datasets and collect around 500 hours of talking-face videos, forming a high-quality large-scale multi-modal face database. A temporal self-attention module is incorporated into the U-Net structure, which allows learning the denoising process on the mixture of images and videos. Extensive experiments suggest that our approach not only achieves comparable or even better performance than state-of-the-arts on each single task, but also facilitates some combined tasks with one-time forward, demonstrating its potential in serving as a foundation generative model in face domain. We further develop an interface such that users can enjoy our one-step service to create, edit, and animate their own characters. Code, dataset, model, and interface will be made publicly available.

The development of online economics arouses the demand of generating images of models on product clothes, to display new clothes and promote sales. However, the expensive proprietary model images challenge the existing image virtual try-on methods in this scenario, as most of them need to be trained on considerable amounts of model images accompanied with paired clothes images. In this paper, we propose a cheap yet scalable weakly-supervised method called Deep Generative Projection (DGP) to address this specific scenario. Lying in the heart of the proposed method is to imitate the process of human predicting the wearing effect, which is an unsupervised imagination based on life experience rather than computation rules learned from supervisions. Here a pretrained StyleGAN is used to capture the practical experience of wearing. Experiments show that projecting the rough alignment of clothing and body onto the StyleGAN space can yield photo-realistic wearing results. Experiments on real scene proprietary model images demonstrate the superiority of DGP over several state-of-the-art supervised methods when generating clothing model images.

Human can extrapolate well, generalize daily knowledge into unseen scenarios, raise and answer counterfactual questions. To imitate this ability via generative models, previous works have extensively studied explicitly encoding Structural Causal Models (SCMs) into architectures of generator networks. This methodology, however, limits the flexibility of the generator as they must be carefully crafted to follow the causal graph, and demands a ground truth SCM with strong ignorability assumption as prior, which is a nontrivial assumption in many real scenarios. Thus, many current causal GAN methods fail to generate high fidelity counterfactual results as they cannot easily leverage state-of-the-art generative models. In this paper, we propose to study counterfactual synthesis from a new perspective of knowledge extrapolation, where a given knowledge dimension of the data distribution is extrapolated, but the remaining knowledge is kept indistinguishable from the original distribution. We show that an adversarial game with a closed-form discriminator can be used to address the knowledge extrapolation problem, and a novel principal knowledge descent method can efficiently estimate the extrapolated distribution through the adversarial game. Our method enjoys both elegant theoretical guarantees and superior performance in many scenarios.

Despite the rapid advancement of semantic discovery in the latent space of Generative Adversarial Networks (GANs), existing approaches either are limited to finding global attributes or rely on a number of segmentation masks to identify local attributes. In this work, we present a highly efficient algorithm to factorize the latent semantics learned by GANs concerning an arbitrary image region. Concretely, we revisit the task of local manipulation with pre-trained GANs and formulate region-based semantic discovery as a dual optimization problem. Through an appropriately defined generalized Rayleigh quotient, we manage to solve such a problem without any annotations or training. Experimental results on various state-of-the-art GAN models demonstrate the effectiveness of our approach, as well as its superiority over prior arts regarding precise control, region robustness, speed of implementation, and simplicity of use.

The rank of neural networks measures information flowing across layers. It is an instance of a key structural condition that applies across broad domains of machine learning. In particular, the assumption of low-rank feature representations led to algorithmic developments in many architectures. For neural networks, however, the intrinsic mechanism that yields low-rank structures remains vague and unclear. To fill this gap, we perform a rigorous study on the behavior of network rank, focusing particularly on the notion of rank deficiency. We theoretically establish a universal monotone decreasing property of network ranks from the basic rules of differential and algebraic composition, and uncover rank deficiency of network blocks and deep function coupling. By virtue of our numerical tools, we provide the first empirical analysis of the per-layer behavior of network ranks in realistic settings, \ieno, ResNets, deep MLPs, and Transformers on ImageNet. These empirical results are in direct accord with our theory. Furthermore, we reveal a novel phenomenon of independence deficit caused by the rank deficiency of deep networks, where classification confidence of a given category can be linearly decided by the confidence of a handful of other categories. The theoretical results of this work, together with the empirical findings, may advance understanding of the inherent principles of deep neural networks. Code to detect the rank behavior of networks can be found in https://github.com/RuiLiFeng/Rank-Diminishing-in-Deep-Neural-Networks.

3D-aware image synthesis aims at learning a generative model that can render photo-realistic 2D images while capturing decent underlying 3D shapes. A popular solution is to adopt the generative adversarial network (GAN) and replace the generator with a 3D renderer, where volume rendering with neural radiance field (NeRF) is commonly used. Despite the advancement of synthesis quality, existing methods fail to obtain moderate 3D shapes. We argue that, considering the two-player game in the formulation of GANs, only making the generator 3D-aware is not enough. In other words, displacing the generative mechanism only offers the capability, but not the guarantee, of producing 3D-aware images, because the supervision of the generator primarily comes from the discriminator. To address this issue, we propose GeoD through learning a geometry-aware discriminator to improve 3D-aware GANs. Concretely, besides differentiating real and fake samples from the 2D image space, the discriminator is additionally asked to derive the geometry information from the inputs, which is then applied as the guidance of the generator. Such a simple yet effective design facilitates learning substantially more accurate 3D shapes. Extensive experiments on various generator architectures and training datasets verify the superiority of GeoD over state-of-the-art alternatives. Moreover, our approach is registered as a general framework such that a more capable discriminator (i.e., with a third task of novel view synthesis beyond domain classification and geometry extraction) can further assist the generator with a better multi-view consistency. Project page can be found at https://vivianszf.github.io/geod.

Discriminator plays a vital role in training generative adversarial networks (GANs) via distinguishing real and synthesized samples. While the real data distribution remains the same, the synthesis distribution keeps varying because of the evolving generator, and thus effects a corresponding change of the bi-classification task assigned to the discriminator. We argue that a discriminator with an on-the-fly adjustment on its capacity can better accommodate such a time-varying task. A comprehensive empirical study confirms that the proposed training strategy, termed as DynamicD, improves the synthesis performance without incurring any additional computation cost or training objectives. Two capacity adjusting schemes are developed for training GANs under different data regimes: i) given a sufficient amount of training data, the discriminator benefits from a progressively increased learning capacity, and ii) when the training data is limited, gradually decreasing the layer width mitigates the over-fitting issue of the discriminator. Experiments on both 2D and 3D-aware image synthesis tasks conducted on a range of datasets substantiate the generalizability of our DynamicD as well as its substantial improvement over the baselines. Furthermore, DynamicD is synergistic to other discriminator-improving approaches (including data augmentation, regularizers, and pre-training), and brings continuous performance gain when combined with them for learning GANs. Code will be made publicly available.

The compelling synthesis results of Generative Adversarial Networks (GANs) demonstrate rich semantic knowledge in their latent codes. To obtain this knowledge for downstream applications, encoding GANs has been proposed to learn encoders, such that real world data can be encoded to latent codes, which can be fed to generators to reconstruct those data. However, despite the theoretical guarantees of precise reconstruction in previous works, current algorithms generally reconstruct inputs with non-negligible deviations from inputs. In this paper we study this predicament of encoding GANs, which is indispensable research for the GAN community. We prove three uncertainty principles of encoding GANs in practice: a) the ‘perfect’ encoder and generator cannot be continuous at the same time, which implies that current framework of encoding GANs is ill-posed and needs rethinking; b) neural networks cannot approximate the underlying encoder and generator precisely at the same time, which explains why we cannot get ‘perfect’ encoders and generators as promised in previous theories; c) neural networks cannot be stable and accurate at the same time, which demonstrates the difficulty of training and trade-off between fidelity and disentanglement encountered in previous works. Our work may eliminate gaps between previous theories and empirical results, promote the understanding of GANs, and guide network designs for follow-up works.

Noise injection is an effective way of circumventing overfitting and enhancing generalization in machine learning, the rationale of which has been validated in deep learning as well. Recently, noise injection exhibits surprising effectiveness when generating high-fidelity images in Generative Adversarial Networks (GANs) (e.g. StyleGAN). Despite its successful applications in GANs, the mechanism of its validity is still unclear. In this paper, we propose a geometric framework to theoretically analyze the role of noise injection in GANs. First, we point out the existence of the adversarial dimension trap inherent in GANs, which leads to the difficulty of learning a proper generator. Second, we successfully model the noise injection framework with exponential maps based on Riemannian geometry. Guided by our theories, we propose a general geometric realization for noise injection. Under our novel framework, the simple noise injection used in StyleGAN reduces to the Euclidean case. The goal of our work is to make theoretical steps towards understanding the underlying mechanism of state-of-the-art GAN algorithms. Experiments on image generation and GAN inversion validate our theory in practice.

The latent space of a Generative Adversarial Network (GAN) has been shown to encode rich semantics within some subspaces. To identify these subspaces, researchers typically analyze the statistical information from a collection of synthesized data, and the identified subspaces tend to control image attributes globally (i.e., manipulating an attribute causes the change of an entire image). By contrast, this work introduces low-rank subspaces that enable more precise control of GAN generation. Concretely, given an arbitrary image and a region of interest (e.g., eyes of face images), we manage to relate the latent space to the image region with the Jacobian matrix and then use low-rank factorization to discover steerable latent subspaces. There are three distinguishable strengths of our approach that can be aptly called LowRankGAN. First, compared to analytic algorithms in prior work, our low-rank factorization of Jacobians is able to find the low-dimensional representation of attribute manifold, making image editing more precise and controllable. Second, low-rank factorization naturally yields a null space of attributes such that moving the latent code within it only affects the outer region of interest. Therefore, local image editing can be simply achieved by projecting an attribute vector into the null space without relying on a spatial mask as existing methods do. Third, our method can robustly work with a local region from one image for analysis yet well generalize to other images, making it much easy to use in practice. Extensive experiments on state-of-the-art GAN models (including StyleGAN2 and BigGAN) trained on various datasets demonstrate the effectiveness of our LowRankGAN.

The accurate exposure is the key of capturing high-quality photos in computational photography, especially for mobile phones that are limited by sizes of camera modules. Inspired by luminosity masks usually applied by professional photographers, in this paper, we develop a novel algorithm for learning local exposures with deep reinforcement adversarial learning. To be specific, we segment an image into sub-images that can reflect variations of dynamic range exposures according to raw low-level features. Based on these sub-images, a local exposure for each sub-image is automatically learned by virtue of policy network sequentially while the reward of learning is globally designed for striking a balance of overall exposures. The aesthetic evaluation function is approximated by discriminator in generative adversarial networks. The reinforcement learning and the adversarial learning are trained collaboratively by asynchronous deterministic policy gradient and generative loss approximation. To further simply the algorithmic architecture, we also prove the feasibility of leveraging the discriminator as the value function. Further more, we employ each local exposure to retouch the raw input image respectively, thus delivering multiple retouched images under different exposures which are fused with exposure blending. The extensive experiments verify that our algorithms are superior to state-of-the-art methods in terms of quantitative accuracy and visual illustration.

Linear Dynamical Systems (LDSs) are the fundamental tools for encoding spatio-temporal data in various disciplines. To enhance the performance of LDSs, in this paper, we address the challenging issue of performing sparse coding on the space of LDSs, where both data and dictionary atoms are LDSs. Rather than approximate the extended observability with a finite-order matrix, we represent the space of LDSs by an infinite Grassmannian consisting of the orthonormalized extended observability subspaces. Via a homeomorphic mapping, such Grassmannian is embedded into the space of symmetric matrices, where a tractable objective function can be derived for sparse coding. Then, we propose an efficient method to learn the system parameters of the dictionary atoms explicitly, by imposing the symmetric constraint to the transition matrices of the data and dictionary systems. Moreover, we combine the state covariance into the algorithm formulation, thus further promoting the performance of the models with symmetric transition matrices. Comparative experimental evaluations reveal the superior performance of proposed methods on various tasks including video classification and tactile recognition.

The Riemannian three-factor matrix completion (R3MC) algorithm is one of the state-of-the-art geometric optimization methods for the low-rank matrix completion problem. It is a nonlinear conjugate-gradient method optimizing on a quotient Riemannian manifold. In the line search step, R3MC approximates the minimum point on the searching curve by minimizing on the line tangent to the curve. However, finding the exact minimum point by iteration is too expensive. We address this issue by proposing a new retraction with a minimizing property. This special property provides the exact minimization for the line search by establishing correspondences between points on the searching curve and points on the tangent line. Accelerated R3MC, which is R3MC equipped with this new retraction, outperforms the original algorithm and other geometric algorithms for matrix completion in our empirical study.

Optimization on Riemannian manifolds is an intuitive generalization of the traditional optimization algorithms in Euclidean spaces. In these algorithms, minimizing along a search direction becomes minimizing along a search curve lying on a manifold. Computing such a curve to be subsequently searched upon is itself computational intensive. We propose a new minimization scheme aiming to find a better step size utilizing the first order information of the search curve. We prove that this scheme can provide further reduction for the cost function when the retraction and the vector transport are collinear. Then we adapt this scheme to propose a heuristic strategy for line search. In numerical experiments, we apply this heuristic strategy to one of the geometric algorithms for matrix completion and show its feasibility and the potential in accelerating computation.

The essence of distantly supervised relation extraction is that it is an incomplete multi-label classification problem with sparse and noisy features.To tackle the sparsity and noise challenges, we propose solving the classification problem using matrix completion on factorized matrix of minimized rank.We formulate relation classification as completing the unknown labels of testing items (entity pairs) in a sparse matrix that concatenates training and testing textual features with training labels.Our algorithmic framework is based on the assumption that the rank of item-byfeature and item-by-label joint matrix is low.We apply two optimization models to recover the underlying low-rank matrix leveraging the sparsity of feature-label matrix.The matrix completion problem is then solved by the fixed point continuation (FPC) algorithm, which can find the global optimum.Experiments on two widely used datasets with different dimensions of textual features demonstrate that our low-rank matrix completion approach significantly outperforms the baseline and the state-of-the-art methods.

When face images are taken in the wild, the large variations in facial pose, illumination, and expression make face recognition challenging. The most fundamental problem for face recognition is to measure the similarity between faces. The traditional measurements such as various mathematical norms, Hausdorff distance, and approximate geodesic distance cannot accurately capture the structural information between faces in such complex circumstances. To address this issue, we develop a novel face patch network, based on which we define a new similarity measure called the random path (RP) measure. The RP measure is derived from the collective similarity of paths by performing random walks in the network. It can globally characterize the contextual and curved structures of the face space. To apply the RP measure, we construct two kinds of networks: the in-face network and the out-face network. The in-face network is drawn from any two face images and captures the local structural information. The out-face network is constructed from all the training face patches, thereby modeling the global structures of face space. The two face networks are structurally complementary and can be combined together to improve the recognition performance. Experiments on the Multi-PIE and LFW benchmarks show that the RP measure outperforms most of the state-of-art algorithms for face recognition.

The archetype hull model is playing an important role in large-scale data analytics and mining, but rarely applied to vision problems. In this paper, we migrate such a geometric model to address face recognition and verification together through proposing a unified archetype hull ranking framework. Upon a scalable graph characterized by a compact set of archetype exemplars whose convex hull encompasses most of the training images, the proposed framework explicitly captures the relevance between any query and the stored archetypes, yielding a rank vector over the archetype hull. The archetype hull ranking is then executed for every block of face images to generate a block wise similarity measure that is achieved by comparing two different rank vectors with respect to the same archetype hull. After integrating block wise similarity measurements with learned importance weights, we accomplish a sensible face similarity measure which can support robust and effective face recognition and verification. We evaluate the face similarity measure in terms of experiments performed on three benchmark face databases Multi-PIE, Pubfig83, and LFW, demonstrating its performance superior to the state-of-the-arts.

In this paper, we develop a geometric framework for linear or nonlinear discriminant subspace learning and classification. In our framework, the structures of classes are conceptualized as a semi-Riemannian manifold which is considered as a submanifold embedded in an ambient semi-Riemannian space. The class structures of original samples can be characterized and deformed by local metrics of the semi-Riemannian space. Semi-Riemannian metrics are uniquely determined by the smoothing of discrete functions and the nullity of the semi-Riemannian space. Based on the geometrization of class structures, optimizing class structures in the feature space is equivalent to maximizing the quadratic quantities of metric tensors in the semi-Riemannian space. Thus supervised discriminant subspace learning reduces to unsupervised semi-Riemannian manifold learning. Based on the proposed framework, a novel algorithm, dubbed as Semi-Riemannian Discriminant Analysis (SRDA), is presented for subspace-based classification. The performance of SRDA is tested on face recognition (singular case) and handwritten capital letter classification (nonsingular case) against existing algorithms. The experimental results show that SRDA works well on recognition and classification, implying that semi-Riemannian geometry is a promising new tool for pattern recognition and machine learning.

Detecting underlying clusters from large-scale data plays a central role in machine learning research. In this paper, we attempt to tackle clustering problems for complex data of multiple distributions and large multi-scales. To this end, we develop an algorithm named Zeta $l$-links, or Zell which consists of two parts: Zeta merging with a similarity graph and an initial set of small clusters derived from local $l$-links of the graph. More specifically, we propose to structurize a cluster using cycles in the associated subgraph. A mathematical tool, Zeta function of a graph, is introduced for the integration of all cycles, leading to a structural descriptor of the cluster in determinantal form. The popularity character of the cluster is conceptualized as the global fusion of variations of the structural descriptor by means of the leave-one-out strategy in the cluster. Zeta merging proceeds, in the agglomerative fashion, according to the maximum incremental popularity among all pairwise clusters. Experiments on toy data, real imagery data, and real sensory data show the promising performance of Zell. The $98.1\%$ accuracy, in the sense of the normalized mutual information, is obtained on the FRGC face data of 16028 samples and 466 facial clusters. The MATLAB codes of Zell will be made publicly available for peer evaluation.

Discriminant feature extraction plays a fundamental role in pattern recognition. In this paper, we propose the linear Laplacian discrimination (LLD) algorithm/or discriminant feature extraction. LLD is an extension of linear discriminant analysis (LDA). Our motivation is to address the issue that LDA cannot work well in cases where sample spaces are non-Euclidean. Specifically, we define the within-class scatter and the between-class scatter using similarities which are based on pairwise distances in sample spaces. Thus the structural information of classes is contained in the within-class and the between-class Laplacian matrices which are free from metrics of sample spaces. The optimal discriminant subspace can be derived by controlling the structural evolution of Laplacian matrices. Experiments are performed on the facial database for FRGC version 2. Experimental results show that LLD is effective in extracting discriminant features.

Structural perception of data plays a fundamental role in pattern analysis and machine learning. In this paper, we develop a new structural perception of data based on local contexts. We first identify the contextual set of a point by finding its nearest neighbors. Then the contextual distance between the point and one of its neighbors is defined by the difference between their contribution to the integrity of the geometric structure of the contextual set, which is depicted by a structural descriptor. The centroid and the coding length are introduced as the examples of descriptors of the contextual set. Furthermore, a directed graph (digraph) is built to model the asymmetry of perception. The edges of the digraph are weighted based on the contextual distances. Thus direction is brought to the undirected data. And the structural perception of data can be performed by mining the properties of the digraph. We also present the method for deriving the global digraph Laplacian from the alignment of the local digraph Laplacians. Experimental results on clustering and ranking of toy problems and real data show the superiority of asymmetric perception.

Dimensionality reduction plays a fundamental role in data processing, for which principal component analysis (PCA) is widely used. In this paper, we develop the Laplacian PCA (LPCA) algorithm which is the extension of PCA to a more general form by locally optimizing the weighted scatter. In addition to the simplicity of PCA, the benefits brought by LPCA are twofold: the strong robustness against noise and the weak metric-dependence on sample spaces. The LPCA algorithm is based on the global alignment of locally Gaussian or linear subspaces via an alignment technique borrowed from manifold learning. Based on the coding length of local samples, the weights can be determined to capture the local principal structure of data. We also give the exemplary application of LPCA to manifold learning. Manifold unfolding (non-linear dimensionality reduction) can be performed by the alignment of tangential maps which are linear transformations of tangent coordinates approximated by LPCA. The superiority of LPCA to PCA and kernel PCA is verified by the experiments on face recognition (FRGC version 2 face database) and manifold (Scherk surface) unfolding.