Text-to-video diffusion models have made remarkable advancements. Driven by their ability to generate temporally coherent videos, research on zero-shot video editing using these fundamental models has expanded rapidly. To enhance editing quality, structural controls are frequently employed in video editing. Among these techniques, cross-attention mask control stands out for its effectiveness and efficiency. However, when cross-attention masks are naively applied to video editing, they can introduce artifacts such as blurring and flickering. Our experiments uncover a critical factor overlooked in previous video editing research: cross-attention masks are not consistently clear but vary with model structure and denoising timestep. To address this issue, we propose the metric Mask Matching Cost (MMC) that quantifies this variability and propose FreeMask, a method for selecting optimal masks tailored to specific video editing tasks. Using MMC-selected masks, we further improve the masked fusion mechanism within comprehensive attention features, e.g., temp, cross, and self-attention modules. Our approach can be seamlessly integrated into existing zero-shot video editing frameworks with better performance, requiring no control assistance or parameter fine-tuning but enabling adaptive decoupling of unedited semantic layouts with mask precision control. Extensive experiments demonstrate that FreeMask achieves superior semantic fidelity, temporal consistency, and editing quality compared to state-of-the-art methods.

As a fundamental backbone for video generation, diffusion models are challenged by low inference speed due to the sequential nature of denoising. Previous methods speed up the models by caching and reusing model outputs at uniformly selected timesteps. However, such a strategy neglects the fact that differences among model outputs are not uniform across timesteps, which hinders selecting the appropriate model outputs to cache, leading to a poor balance between inference efficiency and visual quality. In this study, we introduce Timestep Embedding Aware Cache (TeaCache), a training-free caching approach that estimates and leverages the fluctuating differences among model outputs across timesteps. Rather than directly using the time-consuming model outputs, TeaCache focuses on model inputs, which have a strong correlation with the modeloutputs while incurring negligible computational cost. TeaCache first modulates the noisy inputs using the timestep embeddings to ensure their differences better approximating those of model outputs. TeaCache then introduces a rescaling strategy to refine the estimated differences and utilizes them to indicate output caching. Experiments show that TeaCache achieves up to 4.41× acceleration over Open-Sora-Plan with negligible (-0.07% Vbench score) degradation of visual quality.

The current text-to-video (T2V) generation has made significant progress in synthesizing realistic general videos, but it is still under-explored in identity-specific human video generation with customized ID images. The key challenge lies in maintaining high ID fidelity consistently while preserving the original motion dynamic and semantic following after the identity injection. Current video identity customization methods mainly rely on reconstructing given identity images on text-to-image models, which have a divergent distribution with the T2V model. This process introduces a tuning-inference gap, leading to dynamic and semantic degradation. To tackle this problem, we propose a novel framework, dubbed $\textbf{PersonalVideo}$, that applies a mixture of reward supervision on synthesized videos instead of the simple reconstruction objective on images. Specifically, we first incorporate identity consistency reward to effectively inject the reference's identity without the tuning-inference gap. Then we propose a novel semantic consistency reward to align the semantic distribution of the generated videos with the original T2V model, which preserves its dynamic and semantic following capability during the identity injection. With the non-reconstructive reward training, we further employ simulated prompt augmentation to reduce overfitting by supervising generated results in more semantic scenarios, gaining good robustness even with only a single reference image. Extensive experiments demonstrate our method's superiority in delivering high identity faithfulness while preserving the inherent video generation qualities of the original T2V model, outshining prior methods.

Visual diffusion models achieve remarkable progress, yet they are typically trained at limited resolutions due to the lack of high-resolution data and constrained computation resources, hampering their ability to generate high-fidelity images or videos at higher resolutions. Recent efforts have explored tuning-free strategies to exhibit the untapped potential higher-resolution visual generation of pre-trained models. However, these methods are still prone to producing low-quality visual content with repetitive patterns. The key obstacle lies in the inevitable increase in high-frequency information when the model generates visual content exceeding its training resolution, leading to undesirable repetitive patterns deriving from the accumulated errors. To tackle this challenge, we propose FreeScale, a tuning-free inference paradigm to enable higher-resolution visual generation via scale fusion. Specifically, FreeScale processes information from different receptive scales and then fuses it by extracting desired frequency components. Extensive experiments validate the superiority of our paradigm in extending the capabilities of higher-resolution visual generation for both image and video models. Notably, compared with previous best-performing methods, FreeScale unlocks the 8k-resolution text-to-image generation for the first time.

Scaling visual generation models is essential for real-world content creation, yet requires substantial training and computational expenses. Alternatively, test-time scaling has garnered growing attention due to resource efficiency and promising performance. In this work, we present the first general test-time scaling framework for visual auto-regressive (VAR) models, TTS-VAR, modeling the generation process as a path searching problem. Inspired by VAR's hierarchical coarse-to-fine multi-scale generation, our framework integrates two key components: (i) At coarse scales, we observe that generated tokens are hard for evaluation, possibly leading to erroneous acceptance of inferior samples or rejection of superior samples. Noticing that the coarse scales contain sufficient structural information, we propose clustering-based diversity search. It preserves structural variety through semantic feature clustering, enabling later selection on samples with higher potential. (ii) In fine scales, resampling-based potential selection prioritizes promising candidates using potential scores, which are defined as reward functions incorporating multi-scale generation history. To dynamically balance computational efficiency with exploration capacity, we further introduce an adaptive descending batch size schedule throughout the causal generation process. Experiments on the powerful VAR model Infinity2B show a notable 8.7% GenEval score improvement (0.69→0.75). Key insights reveal that early-stage structural features effectively influence final quality, and resampling efficacy varies across generation scales.

We present Wan-Move, a simple and scalable framework that brings motion control to video generative models. Existing motion-controllable methods typically suffer from coarse control granularity and limited scalability, leaving their outputs insufficient for practical use. We narrow this gap by achieving precise and high-quality motion control. Our core idea is to directly make the original condition features motion-aware for guiding video synthesis. To this end, we first represent object motions with dense point trajectories, allowing fine-grained control over the scene. We then project these trajectories into latent space and propagate the first frame's features along each trajectory, producing an aligned spatiotemporal feature map that tells how each scene element should move. This feature map serves as the updated latent condition, which is naturally integrated into the off-the-shelf image-to-video model, e.g., Wan-I2V-14B, as motion guidance without any architecture change. It removes the need for auxiliary motion encoders and makes fine-tuning base models easily scalable. Through scaled training, Wan-Move generates 5-second, 480p videos whose motion controllability rivals Kling 1.5 Pro's commercial Motion Brush, as indicated by user studies. To support comprehensive evaluation, we further design MoveBench, a rigorously curated benchmark featuring diverse content categories and hybrid-verified annotations. It is distinguished by larger data volume, longer video durations, and high-quality motion annotations. Extensive experiments on MoveBench and the public dataset consistently show Wan-Move's superior motion quality. Code, models, and benchmark data are made available.

Video generation has emerged as a promising tool for world simulation, leveraging visual data to replicate real-world environments. Within this context, egocentric video generation, which centers on the human perspective, holds significant potential for enhancing applications in virtual reality, augmented reality, and gaming. However, the generation of egocentric videos presents substantial challenges due to the dynamic nature of first-person viewpoints, the intricate diversity of actions, and the complex variety of scenes encountered. Existing datasets are inadequate for addressing these challenges effectively. To bridge this gap, we present EgoVid-5M, the first high-quality dataset specifically curated for egocentric video generation. EgoVid-5M encompasses over 5 million egocentric video clips and is enriched with detailed action annotations, including fine-grained kinematic control and high-level textual descriptions. To ensure the integrity and usability of the dataset, we implement a sophisticated data cleansing pipeline designed to maintain frame consistency, action coherence, and motion smoothness under egocentric conditions. Furthermore, we introduce EgoDreamer, which is capable of generating egocentric videos driven simultaneously by action descriptions and kinematic control signals. The EgoVid-5M dataset, associated action annotations, and all data cleansing metadata will be released for the advancement of research in egocentric video generation.

Video inpainting and editing have long been challenging tasks in the video generation community, requiring extensive computational resources and large datasets to train models with satisfactory performance. Recent breakthroughs in large-scale video foundation models have greatly enhanced text-to-video generation capabilities. This naturally leads to the idea of leveraging the prior knowledge from these powerful generators to facilitate video inpainting and editing. In this work, we investigate the feasibility of employing pre-trained text-to-video foundation models for high-quality video inpainting and editing without additional training. Specifically, we introduce a model-agnostic denoising sampler that optimizes the trajectory by maximizing the log-likelihood expectation conditioned on the known video segments. To enable efficient dynamic object removal and replacement, we propose a latent mask fuser that performs accurate video masking directly in latent space, eliminating the need for explicit VAE decoding and encoding. We implement our approach in widely-used foundation generators such as CogVideoX and HunyuanVideo, demonstrating the model-agnostic nature of our sampler. Comprehensive quantitative and qualitative evaluations confirm that our method achieves outstanding video inpainting and editing performance in a plug-and-play fashion.

Audio-driven talking head synthesis is a promising topic with wide applications in digital human, film making and virtual reality. Recent NeRF-based approaches have shown superiority in quality and fidelity compared to previous studies. However, when it comes to few-shot talking head generation, a practical scenario where only few seconds of talking video is available for one identity, two limitations emerge: 1) they either have no base model, which serves as a facial prior for fast convergence, or ignore the importance of audio when building the prior; 2) most of them overlook the degree of correlation between different face regions and audio, e.g., mouth is audio related, while ear is audio independent. In this paper, we present Audio Enhanced Neural Radiance Field (AE-NeRF) to tackle the above issues, which can generate realistic portraits of a new speaker with few-shot dataset. Specifically, we introduce an Audio Aware Aggregation module into the feature fusion stage of the reference scheme, where the weight is determined by the similarity of audio between reference and target image. Then, an Audio-Aligned Face Generation strategy is proposed to model the audio related and audio independent regions respectively, with a dual-NeRF framework. Extensive experiments have shown AE-NeRF surpasses the state-of-the-art on image fidelity, audio-lip synchronization, and generalization ability, even in limited training set or training iterations.

Customized generation using diffusion models has made impressive progress in image generation, but remains unsatisfactory in the challenging video generation task, as it requires the controllability of both subjects and motions. To that end, we present DreamVideo, a novel approach to generating personalized videos from a few static images of the desired subject and a few videos of target motion. DreamVideo decouples this task into two stages, subject learning and motion learning, by leveraging a pre-trained video diffusion model. The subject learning aims to accurately capture the fine appearance of the subject from provided images, which is achieved by combining textual inversion and fine-tuning of our carefully designed identity adapter. In motion learning, we architect a motion adapter and fine-tune it on the given videos to effectively model the target motion pattern. Combining these two lightweight and efficient adapters allows for flexible customization of any subject with any motion. Extensive experimental results demonstrate the superior performance of our DreamVideo over the state-of-the-art methods for customized video generation. Our project page is at https://dreamvideo-t2v.github.io.

Despite diffusion models having shown powerful abilities to generate photorealistic images, generating videos that are realistic and diverse still remains in its infancy. One of the key reasons is that current methods intertwine spatial content and temporal dynamics together, leading to a notably increased complexity of text-to-video generation (T2V). In this work, we propose HiGen, a diffusion model-based method that improves performance by decoupling the spatial and temporal factors of videos from two perspectives, i.e., structure level and content level. At the structure level, we decompose the T2V task into two steps, including spatial reasoning and temporal reasoning, using a unified denoiser. Specifically, we generate spatially coherent priors using text during spatial reasoning and then generate temporally coherent motions from these priors during temporal reasoning. At the content level, we extract two subtle cues from the content of the input video that can express motion and appearance changes, respectively. These two cues then guide the model's training for generating videos, enabling flexible content variations and enhancing temporal stability. Through the decoupled paradigm, HiGen can effectively reduce the complexity of this task and generate realistic videos with semantics accuracy and motion stability. Extensive experiments demonstrate the superior performance of HiGen over the state-of-the-art T2V methods. We have released our source code and models.

Diffusion-based text-to-video generation has witnessed impressive progress in the past year yet still falls behind text-to-image generation. One of the key reasons is the limited scale of publicly available data (e.g., 10M video-text pairs in WebVid10m vs. 5B image-text pairs in LAION), considering the high cost of video captioning. Instead, it could be far easier to collect unlabeled clips from video platforms like YouTube. Motivated by this, we come up with a novel text-to-video generation framework, termed TF-T2V, which can directly learn with text-free videos. The rationale behind is to separate the process of text decoding from that of temporal modeling. To this end, we employ a content branch and a motion branch, which are jointly optimized with weights shared. Following such a pipeline, we study the effect of doubling the scale of training set (i.e., video-only WebVid10M) with some randomly collected text-free videos and are encouraged to observe the performance improvement (FID from 9.67 to 8.19 and FVD from 484 to 441), demonstrating the scalability of our approach. We also find that our model could enjoy sustainable performance gain (FID from 8.19 to 7.64 and FVD from 441 to 366) after reintroducing some text labels for training. Finally, we validate the effectiveness and generalizability of our ideology on both native text-to-video generation and compositional video synthesis paradigms. Code and models will be publicly available at here.

Diffusion models have emerged as the de facto paradigm for video generation. However, their reliance on web-scale data of varied quality often yields results that are visually unappealing and misaligned with the textual prompts. To tackle this problem, we propose InstructVideo to instruct text-to-video diffusion models with human feedback by reward fine-tuning. InstructVideo has two key ingredients: 1) To ameliorate the cost of reward fine-tuning induced by generating through the full DDIM sampling chain, we recast reward fine-tuning as editing. By leveraging the diffusion process to corrupt a sampled video, InstructVideo requires only partial inference of the DDIM sampling chain, reducing fine-tuning cost while improving fine-tuning efficiency. 2) To mitigate the absence of a dedicated video reward model for human preferences, we repurpose established image reward models, e.g., HPSv2. To this end, we propose Segmental Video Reward, a mechanism to provide reward signals based on segmental sparse sampling, and Temporally Attenuated Reward, a method that mitigates temporal modeling degradation during fine-tuning. Extensive experiments, both qualitative and quantitative, validate the practicality and efficacy of using image reward models in InstructVideo, significantly enhancing the visual quality of generated videos without compromising generalization capabilities. Code and models can be accessed through our project page https://instructvideo.github.io/.

Recent advancements in generation models have showcased remarkable capabilities in generating fantastic content. However, most of them are trained on proprietary high-quality data, and some models withhold their parameters and only provide accessible application programming interfaces (APIs), limiting their benefits for downstream tasks. To explore the feasibility of training a text-to-image generation model comparable to advanced models using publicly available resources, we introduce EvolveDirector. This framework interacts with advanced models through their public APIs to obtain text-image data pairs to train a base model. Our experiments with extensive data indicate that the model trained on generated data of the advanced model can approximate its generation capability. However, it requires large-scale samples of 10 million or more. This incurs significant expenses in time, computational resources, and especially the costs associated with calling fee-based APIs. To address this problem, we leverage pre-trained large vision-language models (VLMs) to guide the evolution of the base model. VLM continuously evaluates the base model during training and dynamically updates and refines the training dataset by the discrimination, expansion, deletion, and mutation operations. Experimental results show that this paradigm significantly reduces the required data volume. Furthermore, when approaching multiple advanced models, EvolveDirector can select the best samples generated by them to learn powerful and balanced abilities. The final trained model Edgen is demonstrated to outperform these advanced models. The code and model weights are available at https://github.com/showlab/EvolveDirector.

Open-set action recognition is to reject unknown human action cases which are out of the distribution of the training set. Existing methods mainly focus on learning better uncertainty scores but dismiss the importance of feature representations. We find that features with richer semantic diversity can significantly improve the open-set performance under the same uncertainty scores. In this paper, we begin with analyzing the feature representation behavior in the open-set action recognition (OSAR) problem based on the information bottleneck (IB) theory, and propose to enlarge the instance-specific (IS) and classspecific (CS) information contained in the feature for better performance. To this end, a novel Prototypical Similarity Learning (PSL) framework is proposed to keep the instance variance within the same class to retain more IS information. Besides, we notice that unknown samples sharing similar appearances to known samples are easily misclassified as known classes. To alleviate this issue, video shuffling is further introduced in our PSL to learn distinct temporal information between original and shuffled samples, which we find enlarges the CS information. Extensive experiments demonstrate that the proposed PSL can significantly boost both the open-set and closed-set performance and achieves state-of-the-art results on multiple benchmarks. Code is available at https://github.com/Jun-CEN/PSL.

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.

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.

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.

While numerous effective decentralized algorithms have been proposed with theoretical guarantees and empirical successes, the performance limits in decentralized optimization, especially the influence of network topology and its associated weight matrix on the optimal convergence rate, have not been fully understood. While Lu and Sa have recently provided an optimal rate for non-convex stochastic decentralized optimization using weight matrices associated with linear graphs, the optimal rate with general weight matrices remains unclear. This paper revisits non-convex stochastic decentralized optimization and establishes an optimal convergence rate with general weight matrices. In addition, we also establish the first optimal rate when non-convex loss functions further satisfy the Polyak-Lojasiewicz (PL) condition. Following existing lines of analysis in literature cannot achieve these results. Instead, we leverage the Ring-Lattice graph to admit general weight matrices while maintaining the optimal relation between the graph diameter and weight matrix connectivity. Lastly, we develop a new decentralized algorithm to attain the above two optimal rates up to logarithm factors.

Researches have demonstrated that low bit-width (e.g., INT8) quantization can be employed to accelerate the inference process. It makes the gradient quantization very promising since the backward propagation requires approximately twice more computation than forward one. Due to the variability and uncertainty of gradient distribution, a lot of methods have been proposed to attain training stability. However, most of them ignore the channel-wise gradient distributions and the impact of gradients with different magnitudes, resulting in the degradation of final accuracy. In this paper, we propose a novel INT8 quantization training framework for convolutional neural network to address the above issues. Specifically, we adopt Gradient Vectorized Quantization to quantize the gradient, based on the observation that layer-wise gradients contain multiple distributions along the channel dimension. Then, Magnitude-aware Clipping Strategy is introduced by taking the magnitudes of gradients into consideration when minimizing the quantization error, and we present a theoretical derivation to solve the quantization parameters of different distributions. Experimental results on broad range of computer vision tasks, such as image classification, object detection and video classification, demonstrate that the proposed Distribution Adaptive INT8 Quantization training method has achieved almost lossless training accuracy for different backbones, including ResNet, MobileNetV2, InceptionV3, VGG and AlexNet, which is superior to the state-of-the-art techniques. Moreover, we further implement the INT8 kernel that can accelerate the training iteration more than 200% under the latest Turing architecture, i.e., our method excels on both training accuracy and speed.

Gradient compression has been widely adopted in data-parallel distributed training of deep neural networks to reduce communication overhead. Some literatures have demonstrated that large gradients are more important than small ones because they contain more information, such as Top-k compressor. Other mainstream methods, like random-k compressor and gradient quantization, usually treat all gradients equally. Different from all of them, we regard large and small gradients selection as the exploitation and exploration of gradient information, respectively. And we find taking both of them into consideration is the key to boost the final accuracy. So, we propose a novel gradient compressor: Gradient Sampling with Bayes Prior in this paper. Specifically, we sample important/large gradients based on the global gradient distribution, which is periodically updated across multiple workers. Then we introduce Bayes Prior into distribution model to further explore the gradients. We prove the convergence of our method for smooth non-convex problems in the distributed system. Compared with methods that running after high compression ratio at the expense of accuracy, we pursue no loss of accuracy and the actual acceleration benefit in practice. Experimental comparisons on a variety of computer vision tasks (e.g. image classification and object detection) and backbones (ResNet, MobileNetV2, InceptionV3 and AlexNet) show that our approach outperforms the state-of-the-art techniques in terms of both speed and accuracy, with the limitation of 100× compression ratio.

The scale of deep learning nowadays calls for efficient distributed training algorithms. Decentralized momentum SGD (DmSGD), in which each node averages only with its neighbors, is more communication efficient than vanilla Parallel momentum SGD that incurs global average across all computing nodes. On the other hand, the large-batch training has been demonstrated critical to achieve runtime speedup. This motivates us to investigate how DmSGD performs in the large-batch scenario.In this work, we find the momentum term can amplify the inconsistency bias in DmSGD. Such bias becomes more evident as batch-size grows large and hence results in severe performance degradation. We next propose DecentLaM, a novel decentralized large-batch momentum SGD to remove the momentum-incurred bias. The convergence rate for both strongly convex and non-convex scenarios is established. Our theoretical results justify the superiority of DecentLaM to DmSGD especially in the large-batch scenario. Experimental results on a a variety of computer vision tasks and models show that DecentLaM promises both efficient and high-quality training.

Communication overhead hinders the scalability of large-scale distributed training. Gossip SGD, where each node averages only with its neighbors, is more communication-efficient than the prevalent parallel SGD. However, its convergence rate is reversely proportional to quantity $1-\beta$ which measures the network connectivity. On large and sparse networks where $1-\beta \to 0$, Gossip SGD requires more iterations to converge, which offsets against its communication benefit. This paper introduces Gossip-PGA, which adds Periodic Global Averaging to accelerate Gossip SGD. Its transient stage, i.e., the iterations required to reach asymptotic linear speedup stage, improves from $\Omega(\beta^4 n^3/(1-\beta)^4)$ to $\Omega(\beta^4 n^3 H^4)$ for non-convex problems. The influence of network topology in Gossip-PGA can be controlled by the averaging period $H$. Its transient-stage complexity is also superior to local SGD which has order $\Omega(n^3 H^4)$. Empirical results of large-scale training on image classification (ResNet50) and language modeling (BERT) validate our theoretical findings.

In contrast to regular convolutions with local receptive fields, non-local operations have widely proven an effective method for modeling long-range dependencies. Although lots of prior works have been proposed, prohibitive computation and GPU memory occupation are still the major concerns. Different from that carrying out non-local operations pixel-wise or channel-wise in a computation intensive way, we argue that we can achieve effective non-local operation using a more compact high-order statistic, which can be computed more efficiently and may convey some high-level information. In this paper, we propose an extremely compact non-local learning module (CoNL) with high-order reasoning based on a graph convolution as the core. In our CoNL, a global Hadamard pooling (GHP) as a non-local operation is used to extract a compact second-order feature vector from the input tensor. With the help of a light-weight graph convolution network (GCN), this high-order compact vector is further refined with high-level reasoning. After the GCN refinement, the compact high-order vector intuitively indicates some global semantic characteristics, and is eventually applied to enhance the input tensor through a channel scaling operation. The CoNL module is designed easily pluggable to upgrade existing networks. Extensive experiments on a wide range of tasks demonstrate the effectiveness and efficiency of our work. The proposed CoNL can achieve comparable or superior performance over previous state-of-the-art baselines on video recognition, semantic segmentation, object detection and instance segmentation tasks. For a 96 x 96 x 2048 input, our block consumes 13.6 x less in computational cost than non-local block while 7.6 x smaller in GPU memory occupation.

Videos grow to be one of the largest mediums on the Internet. E-commerce platforms like Alibaba need to process millions of video data across multimedia (e.g., visual, audio, image, and text) and on a variety of tasks (e.g., retrieval, tagging, and summary) every day. In this work, we aim to develop a once and for all pretraining technique for diverse modalities and downstream tasks. To achieve this, we make the following contributions: (1) We propose a self-supervised multi-modal co-training framework. It takes cross-modal pseudo-label consistency as the supervision and can jointly learn representations of multiple modalities. (2) We introduce several novel techniques (e.g., sliding-window subset sampling, coarse-to-fine clustering, fast spatial-temporal convolution and parallel data transmission and processing) to optimize the training process, making billion-scale stable training feasible. (3) We construct a large-scale multi-modal dataset consisting of 1.4 billion videos (~0.5 PB) and train our framework on it. The training takes only 4.6 days on an in-house 256 GPUs cluster, and it simultaneously produces pretrained video, audio, image, motion, and text networks. (4) Finetuning from our pretrained models, we obtain significant performance gains and faster convergence on diverse multimedia tasks at Alibaba. Furthermore, we also validate the learned representation on public datasets. Despite the domain gap between our commodity-centric pretraining and the action-centric evaluation data, we show superior results against state-of-the-arts.

Thanks to the popularity of GPU and the growth of its computational power, more and more deep learning tasks, such as face recognition, image retrieval and word embedding, can take advantage of extreme classification to improve accuracy. However, it remains a big challenge to train a deep model with millions of classes efficiently due to the huge memory and computation consumption in the last layer. By sampling a small set of classes to avoid the total classes calculation, sampling-based approaches have been proved to be an effective solution. But most of them suffer from the following two issues: i) the important classes are ignored or only partly sampled, such as the methods using random sampling scheme or retrieval techniques of low recall (e.g., locality-sensitive hashing), resulting in the degradation of accuracy; ii) inefficient implementation owing to incompatibility with GPU, like selective softmax. It uses hashing forest to help select classes, but the search process is implemented in CPU. To address the above issues, we propose a new sampling-based softmax called ANN Softmax in this paper. Specifically, we employ binary quantization with inverted file system to improve the recall of important classes. With the help of dedicated kernel design, it can be totally parallelized in mainstream training framework. Then, we find the size of important classes that are recalled by each training sample has a great impact on the final accuracy, so we introduce sample grouping optimization to well approximate the full classes training. Experimental evaluations on two tasks (Embedding Learning and Classification) and ten datasets (e.g., MegaFace, ImageNet, SKU datasets) demonstrate our proposed method maintains the same precision as Full Softmax for different loss objectives, including cross entropy loss, ArcFace, CosFace and D-Softmax loss, with only 1/10 sampled classes, which outperforms the state-of-the-art techniques. Moreover, we implement ANN Soft-max in a complete GPU pipeline that can accelerate the training more than 4.3X. Equipped our method with a 256 GPUs cluster, the time of training a classifier of 300 million classes on our SKU-300M dataset can be reduced to ten days.

Deep neural networks (DNNs) have gained tremendous attractions as compelling solutions for applications such as image classification, object detection, speech recognition, and so forth. Its great success comes with excessive trainings to make sure the model accuracy is good enough for those applications. Nowadays, it becomes challenging to train a DNN model because of 1) the model size and data size keep increasing, which usually needs more iterations to train; 2) DNN algorithms evolve rapidly, which requires the training phase to be short for a quick deployment. To address those challenges, distributed training platforms have been proposed to leverage massive server nodes for training, with the hope of significant training time reduction. Therefore, scalability is a critical performance metric to evaluate a distributed training platform. Nevertheless, our analysis reveals that traditional server clusters have poor scalability for training due to the traffic congestions within the server and beyond. The intra-server traffic on the I/O fabric can result in severe congestions and skewed quality of service as high performance devices are competing with each other. Moreover, the traffic congestions on the Ethernet for inter-server communication could also incur significant performance degradation. In this work, we devise a novel distributed training platform, EFLOPS, that adopts an algorithm and system co-design methodology to achieve good scalability. A new server architecture is proposed to alleviate the intra-server congestions. Moreover, a new network topology, BiGraph, is proposed to divide the network into two separate parts, so that there is always a direct connection between any nodes from different parts. Finally, accompany with BiGraph, a topology-aware allreduce algorithm is proposed to eliminate the traffic congestion on the direct connection. The experimental results show that eliminating the congestions on network interface can gain up to 11.3xcommunication speedup. The proposed algorithm and topology can provide further improvement up to 6.08x. The overall performance of ResNet-50 training achieves near-linear scalability, and is competitive to the top-rankings of MLPerf results.

In the last decades, extreme classification has become an essential topic for deep learning. It has achieved great success in many areas, especially in computer vision and natural language processing (NLP). However, it is very challenging to train a deep model with millions of classes due to the memory and computation explosion in the last output layer. In this paper, we propose a large-scale training system to address these challenges. First, we build a hybrid parallel training framework to make the training process feasible. Second, we propose a novel softmax variation named KNN softmax, which reduces both the GPU memory consumption and computation costs and improves the throughput of training. Then, to eliminate the communication overhead, we propose a new overlapping pipeline and a gradient sparsification method. Furthermore, we design a fast continuous convergence strategy to reduce total training iterations by adaptively adjusting learning rate and updating model parameters. With the help of all the proposed methods, we gain 3.9× throughput of our training system and reduce almost 60% of training iterations. The experimental results show that using an in-house 256 GPUs cluster, we could train a classifier of 100 million classes on Alibaba Retail Product Dataset in about five days while achieving a comparable accuracy with the naive softmax training process.

Graph-based approximate nearest neighbor search has attracted more and more attentions due to its online search advantages. Numbers of methods studying the enhancement of speed and recall have been put forward. However, few of them focus on the efficiency and scale of offline graph-construction. For a deployed visual search system with several billions of online images in total, building a billion-scale offline graph in hours is essential, which is almost unachievable by most existing methods.

This paper introduces the large scale visual search algorithm and system infrastructure at Alibaba. The following challenges are discussed under the E-commercial circumstance at Alibaba (a) how to handle heterogeneous image data and bridge the gap between real-shot images from user query and the online images. (b) how to deal with large scale indexing for massive updating data. (c) how to train deep models for effective feature representation without huge human annotations. (d) how to improve the user engagement by considering the quality of the content. We take advantage of large image collection of Alibaba and state-of-the-art deep learning techniques to perform visual search at scale. We present solutions and implementation details to overcome those problems and also share our learnings from building such a large scale commercial visual search engine. Specifically, model and search-based fusion approach is introduced to effectively predict categories. Also, we propose a deep CNN model for joint detection and feature learning by mining user click behavior. The binary index engine is designed to scale up indexing without compromising recall and precision. Finally, we apply all the stages into an end-to-end system architecture, which can simultaneously achieve highly efficient and scalable performance adapting to real-shot images. Extensive experiments demonstrate the advancement of each module in our system. We hope visual search at Alibaba becomes more widely incorporated into today's commercial applications.

This paper addresses the problem of grouping the data points sampled from a union of multiple subspaces in the presence of outliers. Information theoretic objective functions are proposed to combine structured low-rank representations (LRRs) to capture the global structure of data and information theoretic measures to handle outliers. In theoretical part, we point out that group sparsity-induced measures (ℓ2,1-norm, ℓα-norm, and correntropy) can be justified from the viewpoint of halfquadratic (HQ) optimization, which facilitates both convergence study and algorithmic development. In particular, a general formulation is accordingly proposed to unify HQ-based group sparsity methods into a common framework. In algorithmic part, we develop information theoretic subspace clustering methods via correntropy. With the help of Parzen window estimation, correntropy is used to handle either outliers under any distributions or sample-specific errors in data. Pairwise link constraints are further treated as a prior structure of LRRs. Based on the HQ framework, iterative algorithms are developed to solve the nonconvex information theoretic loss functions. Experimental results on three benchmark databases show that our methods can further improve the robustness of LRR subspace clustering and outperform other state-of-the-art subspace clustering methods.

Subspace clustering has important and wide applications in computer vision and pattern recognition. It is a challenging task to learn low-dimensional subspace structures due to complex noise existing in high-dimensional data. Complex noise has much more complex statistical structures, and is neither Gaussian nor Laplacian noise. Recent subspace clustering methods usually assume a sparse representation of the errors incurred by noise and correct these errors iteratively. However, large corruptions incurred by complex noise cannot be well addressed by these methods. A novel optimization model for robust subspace clustering is proposed in this paper. Its objective function mainly includes two parts. The first part aims to achieve a sparse representation of each high-dimensional data point with other data points. The second part aims to maximize the correntropy between a given data point and its low-dimensional representation with other points. Correntropy is a robust measure so that the influence of large corruptions on subspace clustering can be greatly suppressed. An extension of pairwise link constraints is also proposed as prior information to deal with complex noise. Half-quadratic minimization is provided as an efficient solution to the proposed robust subspace clustering formulations. Experimental results on three commonly used data sets show that our method outperforms state-of-the-art subspace clustering methods.

Subspace clustering has important and wide applications in computer vision and pattern recognition. It is a challenging task to learn low-dimensional subspace structures due to the possible errors (e.g., noise and corruptions) existing in high-dimensional data. Recent subspace clustering methods usually assume a sparse representation of corrupted errors and correct the errors iteratively. However large corruptions in real-world applications can not be well addressed by these methods. A novel optimization model for robust subspace clustering is proposed in this paper. The objective function of our model mainly includes two parts. The first part aims to achieve a sparse representation of each high-dimensional data point with other data points. The second part aims to maximize the correntropy between a given data point and its low-dimensional representation with other points. Correntropy is a robust measure so that the influence of large corruptions on subspace clustering can be greatly suppressed. An extension of our method with explicit introduction of representation error terms into the model is also proposed. Half-quadratic minimization is provided as an efficient solution to the proposed robust subspace clustering formulations. Experimental results on Hopkins 155 dataset and Extended Yale Database B demonstrate that our method outperforms state-of-the-art subspace clustering methods.