Jianlong Fu

Ludan Ruan, Yiyang Ma, Huan Yang et al. · microsoft-research

We propose the first joint audio-video generation framework that brings engaging watching and listening experiences simultaneously, towards high-quality realistic videos. To generate joint audio-video pairs, we propose a novel Multi-Modal Diffusion model (i.e., MM-Diffusion), with two-coupled denoising autoencoders. In contrast to existing single-modal diffusion models, MM-Diffusion consists of a sequential multi-modal U-Net for a joint denoising process by design. Two subnets for audio and video learn to gradually generate aligned audio-video pairs from Gaussian noises. To ensure semantic consistency across modalities, we propose a novel random-shift based attention block bridging over the two subnets, which enables efficient cross-modal alignment, and thus reinforces the audio-video fidelity for each other. Extensive experiments show superior results in unconditional audio-video generation, and zero-shot conditional tasks (e.g., video-to-audio). In particular, we achieve the best FVD and FAD on Landscape and AIST++ dancing datasets. Turing tests of 10k votes further demonstrate dominant preferences for our model. The code and pre-trained models can be downloaded at https://github.com/researchmm/MM-Diffusion.

Chengxu Liu, Huan Yang, Jianlong Fu et al. · microsoft-research

Video super-resolution (VSR) aims to restore a sequence of high-resolution (HR) frames from their low-resolution (LR) counterparts. Although some progress has been made, there are grand challenges to effectively utilize temporal dependency in entire video sequences. Existing approaches usually align and aggregate video frames from limited adjacent frames (e.g., 5 or 7 frames), which prevents these approaches from satisfactory results. In this paper, we take one step further to enable effective spatio-temporal learning in videos. We propose a novel Trajectory-aware Transformer for Video Super-Resolution (TTVSR). In particular, we formulate video frames into several pre-aligned trajectories which consist of continuous visual tokens. For a query token, self-attention is only learned on relevant visual tokens along spatio-temporal trajectories. Compared with vanilla vision Transformers, such a design significantly reduces the computational cost and enables Transformers to model long-range features. We further propose a cross-scale feature tokenization module to overcome scale-changing problems that often occur in long-range videos. Experimental results demonstrate the superiority of the proposed TTVSR over state-of-the-art models, by extensive quantitative and qualitative evaluations in four widely-used video super-resolution benchmarks. Both code and pre-trained models can be downloaded at https://github.com/researchmm/TTVSR.

Yuchong Sun, Hongwei Xue, Ruihua Song et al. · microsoft-research

Large-scale video-language pre-training has shown significant improvement in video-language understanding tasks. Previous studies of video-language pretraining mainly focus on short-form videos (i.e., within 30 seconds) and sentences, leaving long-form video-language pre-training rarely explored. Directly learning representation from long-form videos and language may benefit many long-form video-language understanding tasks. However, it is challenging due to the difficulty of modeling long-range relationships and the heavy computational burden caused by more frames. In this paper, we introduce a Long-Form VIdeo-LAnguage pre-training model (LF-VILA) and train it on a large-scale long-form video and paragraph dataset constructed from an existing public dataset. To effectively capture the rich temporal dynamics and to better align video and language in an efficient end-to-end manner, we introduce two novel designs in our LF-VILA model. We first propose a Multimodal Temporal Contrastive (MTC) loss to learn the temporal relation across different modalities by encouraging fine-grained alignment between long-form videos and paragraphs. Second, we propose a Hierarchical Temporal Window Attention (HTWA) mechanism to effectively capture long-range dependency while reducing computational cost in Transformer. We fine-tune the pre-trained LF-VILA model on seven downstream long-form video-language understanding tasks of paragraph-to-video retrieval and long-form video question-answering, and achieve new state-of-the-art performances. Specifically, our model achieves 16.1% relative improvement on ActivityNet paragraph-to-video retrieval task and 2.4% on How2QA task, respectively. We release our code, dataset, and pre-trained models at https://github.com/microsoft/XPretrain.

Zhongwei Qiu, Huan Yang, Jianlong Fu et al. · microsoft-research

Compressed video super-resolution (VSR) aims to restore high-resolution frames from compressed low-resolution counterparts. Most recent VSR approaches often enhance an input frame by borrowing relevant textures from neighboring video frames. Although some progress has been made, there are grand challenges to effectively extract and transfer high-quality textures from compressed videos where most frames are usually highly degraded. In this paper, we propose a novel Frequency-Transformer for compressed video super-resolution (FTVSR) that conducts self-attention over a joint space-time-frequency domain. First, we divide a video frame into patches, and transform each patch into DCT spectral maps in which each channel represents a frequency band. Such a design enables a fine-grained level self-attention on each frequency band, so that real visual texture can be distinguished from artifacts, and further utilized for video frame restoration. Second, we study different self-attention schemes, and discover that a divided attention which conducts a joint space-frequency attention before applying temporal attention on each frequency band, leads to the best video enhancement quality. Experimental results on two widely-used video super-resolution benchmarks show that FTVSR outperforms state-of-the-art approaches on both uncompressed and compressed videos with clear visual margins. Code is available at https://github.com/researchmm/FTVSR.

Tiankai Hang, Huan Yang, Bei Liu et al. · microsoft-research

Recent works on language-guided image manipulation have shown great power of language in providing rich semantics, especially for face images. However, the other natural information, motions, in language is less explored. In this paper, we leverage the motion information and study a novel task, language-guided face animation, that aims to animate a static face image with the help of languages. To better utilize both semantics and motions from languages, we propose a simple yet effective framework. Specifically, we propose a recurrent motion generator to extract a series of semantic and motion information from the language and feed it along with visual information to a pre-trained StyleGAN to generate high-quality frames. To optimize the proposed framework, three carefully designed loss functions are proposed including a regularization loss to keep the face identity, a path length regularization loss to ensure motion smoothness, and a contrastive loss to enable video synthesis with various language guidance in one single model. Extensive experiments with both qualitative and quantitative evaluations on diverse domains (\textit{e.g.,} human face, anime face, and dog face) demonstrate the superiority of our model in generating high-quality and realistic videos from one still image with the guidance of language. Code will be available at https://github.com/TiankaiHang/language-guided-animation.git.

Yiyang Ma, Huan Yang, Bei Liu et al. · microsoft-research

AI illustrator aims to automatically design visually appealing images for books to provoke rich thoughts and emotions. To achieve this goal, we propose a framework for translating raw descriptions with complex semantics into semantically corresponding images. The main challenge lies in the complexity of the semantics of raw descriptions, which may be hard to be visualized (e.g., "gloomy" or "Asian"). It usually poses challenges for existing methods to handle such descriptions. To address this issue, we propose a Prompt-based Cross-Modal Generation Framework (PCM-Frame) to leverage two powerful pre-trained models, including CLIP and StyleGAN. Our framework consists of two components: a projection module from Text Embeddings to Image Embeddings based on prompts, and an adapted image generation module built on StyleGAN which takes Image Embeddings as inputs and is trained by combined semantic consistency losses. To bridge the gap between realistic images and illustration designs, we further adopt a stylization model as post-processing in our framework for better visual effects. Benefiting from the pre-trained models, our method can handle complex descriptions and does not require external paired data for training. Furthermore, we have built a benchmark that consists of 200 raw descriptions. We conduct a user study to demonstrate our superiority over the competing methods with complicated texts. We release our code at https://github.com/researchmm/AI_Illustrator.

Zixi Tuo, Huan Yang, Jianlong Fu et al. · microsoft-research

Existing real-world video super-resolution (VSR) methods focus on designing a general degradation pipeline for open-domain videos while ignoring data intrinsic characteristics which strongly limit their performance when applying to some specific domains (eg., animation videos). In this paper, we thoroughly explore the characteristics of animation videos and leverage the rich priors in real-world animation data for a more practical animation VSR model. In particular, we propose a multi-scale Vector-Quantized Degradation model for animation video Super-Resolution (VQD-SR) to decompose the local details from global structures and transfer the degradation priors in real-world animation videos to a learned vector-quantized codebook for degradation modeling. A rich-content Real Animation Low-quality (RAL) video dataset is collected for extracting the priors. We further propose a data enhancement strategy for high-resolution (HR) training videos based on our observation that existing HR videos are mostly collected from the Web which contains conspicuous compression artifacts. The proposed strategy is valid to lift the upper bound of animation VSR performance, regardless of the specific VSR model. Experimental results demonstrate the superiority of the proposed VQD-SR over state-of-the-art methods, through extensive quantitative and qualitative evaluations of the latest animation video super-resolution benchmark. The code and pre-trained models can be downloaded at https://github.com/researchmm/VQD-SR.

Hongwei Xue, Yuchong Sun, Bei Liu et al.

The pre-trained image-text models, like CLIP, have demonstrated the strong power of vision-language representation learned from a large scale of web-collected image-text data. In light of the well-learned visual features, some existing works transfer image representation to video domain and achieve good results. However, how to utilize image-language pre-trained model (e.g., CLIP) for video-language pre-training (post-pretraining) is still under explored. In this paper, we investigate two questions: 1) what are the factors hindering post-pretraining CLIP to further improve the performance on video-language tasks? and 2) how to mitigate the impact of these factors? Through a series of comparative experiments and analyses, we find that the data scale and domain gap between language sources have great impacts. Motivated by these, we propose a Omnisource Cross-modal Learning method equipped with a Video Proxy mechanism on the basis of CLIP, namely CLIP-ViP. Extensive results show that our approach improves the performance of CLIP on video-text retrieval by a large margin. Our model also achieves SOTA results on a variety of datasets, including MSR-VTT, DiDeMo, LSMDC, and ActivityNet. We will release our code and pre-trained CLIP-ViP models at https://github.com/microsoft/XPretrain/tree/main/CLIP-ViP.

Kan Wu, Jinnian Zhang, Houwen Peng et al.

Vision transformer (ViT) recently has drawn great attention in computer vision due to its remarkable model capability. However, most prevailing ViT models suffer from huge number of parameters, restricting their applicability on devices with limited resources. To alleviate this issue, we propose TinyViT, a new family of tiny and efficient small vision transformers pretrained on large-scale datasets with our proposed fast distillation framework. The central idea is to transfer knowledge from large pretrained models to small ones, while enabling small models to get the dividends of massive pretraining data. More specifically, we apply distillation during pretraining for knowledge transfer. The logits of large teacher models are sparsified and stored in disk in advance to save the memory cost and computation overheads. The tiny student transformers are automatically scaled down from a large pretrained model with computation and parameter constraints. Comprehensive experiments demonstrate the efficacy of TinyViT. It achieves a top-1 accuracy of 84.8% on ImageNet-1k with only 21M parameters, being comparable to Swin-B pretrained on ImageNet-21k while using 4.2 times fewer parameters. Moreover, increasing image resolutions, TinyViT can reach 86.5% accuracy, being slightly better than Swin-L while using only 11% parameters. Last but not the least, we demonstrate a good transfer ability of TinyViT on various downstream tasks. Code and models are available at https://github.com/microsoft/Cream/tree/main/TinyViT.

Zhongwei Qiu, Huan Yang, Jianlong Fu et al. · microsoft-research

Video Super-Resolution (VSR) aims to restore high-resolution (HR) videos from low-resolution (LR) videos. Existing VSR techniques usually recover HR frames by extracting pertinent textures from nearby frames with known degradation processes. Despite significant progress, grand challenges are remained to effectively extract and transmit high-quality textures from high-degraded low-quality sequences, such as blur, additive noises, and compression artifacts. In this work, a novel Frequency-Transformer (FTVSR) is proposed for handling low-quality videos that carry out self-attention in a combined space-time-frequency domain. First, video frames are split into patches and each patch is transformed into spectral maps in which each channel represents a frequency band. It permits a fine-grained self-attention on each frequency band, so that real visual texture can be distinguished from artifacts. Second, a novel dual frequency attention (DFA) mechanism is proposed to capture the global frequency relations and local frequency relations, which can handle different complicated degradation processes in real-world scenarios. Third, we explore different self-attention schemes for video processing in the frequency domain and discover that a ``divided attention'' which conducts a joint space-frequency attention before applying temporal-frequency attention, leads to the best video enhancement quality. Extensive experiments on three widely-used VSR datasets show that FTVSR outperforms state-of-the-art methods on different low-quality videos with clear visual margins. Code and pre-trained models are available at https://github.com/researchmm/FTVSR.

Shengming Yin, Chenfei Wu, Huan Yang et al. · microsoft-research, pku

In this paper, we propose NUWA-XL, a novel Diffusion over Diffusion architecture for eXtremely Long video generation. Most current work generates long videos segment by segment sequentially, which normally leads to the gap between training on short videos and inferring long videos, and the sequential generation is inefficient. Instead, our approach adopts a ``coarse-to-fine'' process, in which the video can be generated in parallel at the same granularity. A global diffusion model is applied to generate the keyframes across the entire time range, and then local diffusion models recursively fill in the content between nearby frames. This simple yet effective strategy allows us to directly train on long videos (3376 frames) to reduce the training-inference gap, and makes it possible to generate all segments in parallel. To evaluate our model, we build FlintstonesHD dataset, a new benchmark for long video generation. Experiments show that our model not only generates high-quality long videos with both global and local coherence, but also decreases the average inference time from 7.55min to 26s (by 94.26\%) at the same hardware setting when generating 1024 frames. The homepage link is \url{https://msra-nuwa.azurewebsites.net/}

Junchen Zhu, Huan Yang, Huiguo He et al. · microsoft-research

In this paper, we present MovieFactory, a powerful framework to generate cinematic-picture (3072$\times$1280), film-style (multi-scene), and multi-modality (sounding) movies on the demand of natural languages. As the first fully automated movie generation model to the best of our knowledge, our approach empowers users to create captivating movies with smooth transitions using simple text inputs, surpassing existing methods that produce soundless videos limited to a single scene of modest quality. To facilitate this distinctive functionality, we leverage ChatGPT to expand user-provided text into detailed sequential scripts for movie generation. Then we bring scripts to life visually and acoustically through vision generation and audio retrieval. To generate videos, we extend the capabilities of a pretrained text-to-image diffusion model through a two-stage process. Firstly, we employ spatial finetuning to bridge the gap between the pretrained image model and the new video dataset. Subsequently, we introduce temporal learning to capture object motion. In terms of audio, we leverage sophisticated retrieval models to select and align audio elements that correspond to the plot and visual content of the movie. Extensive experiments demonstrate that our MovieFactory produces movies with realistic visuals, diverse scenes, and seamlessly fitting audio, offering users a novel and immersive experience. Generated samples can be found in YouTube or Bilibili (1080P).

Yiyang Ma, Huan Yang, Wenjing Wang et al. · microsoft-research

Language-guided image generation has achieved great success nowadays by using diffusion models. However, texts can be less detailed to describe highly-specific subjects such as a particular dog or a certain car, which makes pure text-to-image generation not accurate enough to satisfy user requirements. In this work, we present a novel Unified Multi-Modal Latent Diffusion (UMM-Diffusion) which takes joint texts and images containing specified subjects as input sequences and generates customized images with the subjects. To be more specific, both input texts and images are encoded into one unified multi-modal latent space, in which the input images are learned to be projected to pseudo word embedding and can be further combined with text to guide image generation. Besides, to eliminate the irrelevant parts of the input images such as background or illumination, we propose a novel sampling technique of diffusion models used by the image generator which fuses the results guided by multi-modal input and pure text input. By leveraging the large-scale pre-trained text-to-image generator and the designed image encoder, our method is able to generate high-quality images with complex semantics from both aspects of input texts and images.

Chengxu Liu, Huan Yang, Jianlong Fu et al. · microsoft-research

Image enhancement aims at improving the aesthetic visual quality of photos by retouching the color and tone, and is an essential technology for professional digital photography. Recent years deep learning-based image enhancement algorithms have achieved promising performance and attracted increasing popularity. However, typical efforts attempt to construct a uniform enhancer for all pixels' color transformation. It ignores the pixel differences between different content (e.g., sky, ocean, etc.) that are significant for photographs, causing unsatisfactory results. In this paper, we propose a novel learnable context-aware 4-dimensional lookup table (4D LUT), which achieves content-dependent enhancement of different contents in each image via adaptively learning of photo context. In particular, we first introduce a lightweight context encoder and a parameter encoder to learn a context map for the pixel-level category and a group of image-adaptive coefficients, respectively. Then, the context-aware 4D LUT is generated by integrating multiple basis 4D LUTs via the coefficients. Finally, the enhanced image can be obtained by feeding the source image and context map into fused context-aware 4D~LUT via quadrilinear interpolation. Compared with traditional 3D LUT, i.e., RGB mapping to RGB, which is usually used in camera imaging pipeline systems or tools, 4D LUT, i.e., RGBC(RGB+Context) mapping to RGB, enables finer control of color transformations for pixels with different content in each image, even though they have the same RGB values. Experimental results demonstrate that our method outperforms other state-of-the-art methods in widely-used benchmarks.

Jianbo Wang, Huan Yang, Jianlong Fu et al. · microsoft-research

With the development of the convolutional neural network, image style transfer has drawn increasing attention. However, most existing approaches adopt a global feature transformation to transfer style patterns into content images (e.g., AdaIN and WCT). Such a design usually destroys the spatial information of the input images and fails to transfer fine-grained style patterns into style transfer results. To solve this problem, we propose a novel STyle TRansformer (STTR) network which breaks both content and style images into visual tokens to achieve a fine-grained style transformation. Specifically, two attention mechanisms are adopted in our STTR. We first propose to use self-attention to encode content and style tokens such that similar tokens can be grouped and learned together. We then adopt cross-attention between content and style tokens that encourages fine-grained style transformations. To compare STTR with existing approaches, we conduct user studies on Amazon Mechanical Turk (AMT), which are carried out with 50 human subjects with 1,000 votes in total. Extensive evaluations demonstrate the effectiveness and efficiency of the proposed STTR in generating visually pleasing style transfer results.

Junchen Zhu, Huan Yang, Wenjing Wang et al. · microsoft-research

Videos for mobile devices become the most popular access to share and acquire information recently. For the convenience of users' creation, in this paper, we present a system, namely MobileVidFactory, to automatically generate vertical mobile videos where users only need to give simple texts mainly. Our system consists of two parts: basic and customized generation. In the basic generation, we take advantage of the pretrained image diffusion model, and adapt it to a high-quality open-domain vertical video generator for mobile devices. As for the audio, by retrieving from our big database, our system matches a suitable background sound for the video. Additionally to produce customized content, our system allows users to add specified screen texts to the video for enriching visual expression, and specify texts for automatic reading with optional voices as they like.

Chengxu Liu, Huan Yang, Jianlong Fu et al. · microsoft-research

Video frame interpolation (VFI) aims to synthesize an intermediate frame between two consecutive frames. State-of-the-art approaches usually adopt a two-step solution, which includes 1) generating locally-warped pixels by flow-based motion estimations, 2) blending the warped pixels to form a full frame through deep neural synthesis networks. However, due to the inconsistent warping from the two consecutive frames, the warped features for new frames are usually not aligned, which leads to distorted and blurred frames, especially when large and complex motions occur. To solve this issue, in this paper we propose a novel Trajectory-aware Transformer for Video Frame Interpolation (TTVFI). In particular, we formulate the warped features with inconsistent motions as query tokens, and formulate relevant regions in a motion trajectory from two original consecutive frames into keys and values. Self-attention is learned on relevant tokens along the trajectory to blend the pristine features into intermediate frames through end-to-end training. Experimental results demonstrate that our method outperforms other state-of-the-art methods in four widely-used VFI benchmarks. Both code and pre-trained models will be released soon.

Bolin Ni, Houwen Peng, Minghao Chen et al.

Contrastive language-image pretraining has shown great success in learning visual-textual joint representation from web-scale data, demonstrating remarkable "zero-shot" generalization ability for various image tasks. However, how to effectively expand such new language-image pretraining methods to video domains is still an open problem. In this work, we present a simple yet effective approach that adapts the pretrained language-image models to video recognition directly, instead of pretraining a new model from scratch. More concretely, to capture the long-range dependencies of frames along the temporal dimension, we propose a cross-frame attention mechanism that explicitly exchanges information across frames. Such module is lightweight and can be plugged into pretrained language-image models seamlessly. Moreover, we propose a video-specific prompting scheme, which leverages video content information for generating discriminative textual prompts. Extensive experiments demonstrate that our approach is effective and can be generalized to different video recognition scenarios. In particular, under fully-supervised settings, our approach achieves a top-1 accuracy of 87.1% on Kinectics-400, while using 12 times fewer FLOPs compared with Swin-L and ViViT-H. In zero-shot experiments, our approach surpasses the current state-of-the-art methods by +7.6% and +14.9% in terms of top-1 accuracy under two popular protocols. In few-shot scenarios, our approach outperforms previous best methods by +32.1% and +23.1% when the labeled data is extremely limited. Code and models are available at https://aka.ms/X-CLIP

Huiguo He, Tianfu Wang, Huan Yang et al. · microsoft-research

We study the task of generating profitable Non-Fungible Token (NFT) images from user-input texts. Recent advances in diffusion models have shown great potential for image generation. However, existing works can fall short in generating visually-pleasing and highly-profitable NFT images, mainly due to the lack of 1) plentiful and fine-grained visual attribute prompts for an NFT image, and 2) effective optimization metrics for generating high-quality NFT images. To solve these challenges, we propose a Diffusion-based generation framework with Multiple Visual-Policies as rewards (i.e., Diffusion-MVP) for NFT images. The proposed framework consists of a large language model (LLM), a diffusion-based image generator, and a series of visual rewards by design. First, the LLM enhances a basic human input (such as "panda") by generating more comprehensive NFT-style prompts that include specific visual attributes, such as "panda with Ninja style and green background." Second, the diffusion-based image generator is fine-tuned using a large-scale NFT dataset to capture fine-grained image styles and accessory compositions of popular NFT elements. Third, we further propose to utilize multiple visual-policies as optimization goals, including visual rarity levels, visual aesthetic scores, and CLIP-based text-image relevances. This design ensures that our proposed Diffusion-MVP is capable of minting NFT images with high visual quality and market value. To facilitate this research, we have collected the largest publicly available NFT image dataset to date, consisting of 1.5 million high-quality images with corresponding texts and market values. Extensive experiments including objective evaluations and user studies demonstrate that our framework can generate NFT images showing more visually engaging elements and higher market value, compared with SOTA approaches.

Jaeseok Byun, Taebaek Hwang, Jianlong Fu et al.

Most of the currently existing vision and language pre-training (VLP) methods have mainly focused on how to extract and align vision and text features. In contrast to the mainstream VLP methods, we highlight that two routinely applied steps during pre-training have crucial impact on the performance of the pre-trained model: in-batch hard negative sampling for image-text matching (ITM) and assigning the large masking probability for the masked language modeling (MLM). After empirically showing the unexpected effectiveness of above two steps, we systematically devise our GRIT-VLP, which adaptively samples mini-batches for more effective mining of hard negative samples for ITM while maintaining the computational cost for pre-training. Our method consists of three components: 1) GRouped mIni-baTch sampling (GRIT) strategy that collects similar examples in a mini-batch, 2) ITC consistency loss for improving the mining ability, and 3) enlarged masking probability for MLM. Consequently, we show our GRIT-VLP achieves a new state-of-the-art performance on various downstream tasks with much less computational cost. Furthermore, we demonstrate that our model is essentially in par with ALBEF, the previous state-of-the-art, only with one-third of training epochs on the same training data. Code is available at https://github.com/jaeseokbyun/GRIT-VLP.

Fuzhi Yang, Huan Yang, Yanhong Zeng et al. · microsoft-research

Blind super-resolution (SR) aims to recover high-quality visual textures from a low-resolution (LR) image, which is usually degraded by down-sampling blur kernels and additive noises. This task is extremely difficult due to the challenges of complicated image degradations in the real-world. Existing SR approaches either assume a predefined blur kernel or a fixed noise, which limits these approaches in challenging cases. In this paper, we propose a Degradation-guided Meta-restoration network for blind Super-Resolution (DMSR) that facilitates image restoration for real cases. DMSR consists of a degradation extractor and meta-restoration modules. The extractor estimates the degradations in LR inputs and guides the meta-restoration modules to predict restoration parameters for different degradations on-the-fly. DMSR is jointly optimized by a novel degradation consistency loss and reconstruction losses. Through such an optimization, DMSR outperforms SOTA by a large margin on three widely-used benchmarks. A user study including 16 subjects further validates the superiority of DMSR in real-world blind SR tasks.

Jinnian Zhang, Houwen Peng, Kan Wu et al.

Vision Transformer (ViT) models have recently drawn much attention in computer vision due to their high model capability. However, ViT models suffer from huge number of parameters, restricting their applicability on devices with limited memory. To alleviate this problem, we propose MiniViT, a new compression framework, which achieves parameter reduction in vision transformers while retaining the same performance. The central idea of MiniViT is to multiplex the weights of consecutive transformer blocks. More specifically, we make the weights shared across layers, while imposing a transformation on the weights to increase diversity. Weight distillation over self-attention is also applied to transfer knowledge from large-scale ViT models to weight-multiplexed compact models. Comprehensive experiments demonstrate the efficacy of MiniViT, showing that it can reduce the size of the pre-trained Swin-B transformer by 48\%, while achieving an increase of 1.0\% in Top-1 accuracy on ImageNet. Moreover, using a single-layer of parameters, MiniViT is able to compress DeiT-B by 9.7 times from 86M to 9M parameters, without seriously compromising the performance. Finally, we verify the transferability of MiniViT by reporting its performance on downstream benchmarks. Code and models are available at here.

Zhongwei Qiu, Kai Qiu, Jianlong Fu et al.

Modern deep learning-based 3D pose estimation approaches require plenty of 3D pose annotations. However, existing 3D datasets lack diversity, which limits the performance of current methods and their generalization ability. Although existing methods utilize 2D pose annotations to help 3D pose estimation, they mainly focus on extracting 2D structural constraints from 2D poses, ignoring the 3D information hidden in the images. In this paper, we propose a novel method to extract weak 3D information directly from 2D images without 3D pose supervision. Firstly, we utilize 2D pose annotations and perspective prior knowledge to generate the relationship of that keypoint is closer or farther from the camera, called relative depth. We collect a 2D pose dataset (MCPC) and generate relative depth labels. Based on MCPC, we propose a weakly-supervised pre-training (WSP) strategy to distinguish the depth relationship between two points in an image. WSP enables the learning of the relative depth of two keypoints on lots of in-the-wild images, which is more capable of predicting depth and generalization ability for 3D human pose estimation. After fine-tuning on 3D pose datasets, WSP achieves state-of-the-art results on two widely-used benchmarks.

Jiange Yang, Wenhui Tan, Chuhao Jin et al.

Improving the generalization capabilities of general-purpose robotic manipulation agents in the real world has long been a significant challenge. Existing approaches often rely on collecting large-scale robotic data which is costly and time-consuming, such as the RT-1 dataset. However, due to insufficient diversity of data, these approaches typically suffer from limiting their capability in open-domain scenarios with new objects and diverse environments. In this paper, we propose a novel paradigm that effectively leverages language-reasoning segmentation mask generated by internet-scale foundation models, to condition robot manipulation tasks. By integrating the mask modality, which incorporates semantic, geometric, and temporal correlation priors derived from vision foundation models, into the end-to-end policy model, our approach can effectively and robustly perceive object pose and enable sample-efficient generalization learning, including new object instances, semantic categories, and unseen backgrounds. We first introduce a series of foundation models to ground natural language demands across multiple tasks. Secondly, we develop a two-stream 2D policy model based on imitation learning, which processes raw images and object masks to predict robot actions with a local-global perception manner. Extensive realworld experiments conducted on a Franka Emika robot arm demonstrate the effectiveness of our proposed paradigm and policy architecture. Demos can be found in our submitted video, and more comprehensive ones can be found in link1 or link2.

Ruoxuan Zhang, Qiaoqiao Wan, Zhengguang Wang et al.

Theory of Mind (ToM) enables an agent to reason about another actor's beliefs, goals, and intentions, which is essential for human-centered embodied assistance. Existing ToM benchmarks have advanced text and multimodal mental-state recognition, but they mostly evaluate offline question answering or final action prediction. They do not fully test whether an embodied agent can stay connected to a changing environment, update actor-specific beliefs, decide when reasoning is needed, and intervene only when help is useful. Building on MindPower, we extend robot-centric ToM reasoning to a real-time closed-loop setting and introduce MindClaw, a framework for embodied mental-state reasoning with precision intervention. MindClaw connects multi-source inputs, belief memory, an embodied cognitive trigger skill, mental reasoning, and action generation, allowing the agent to output helpful actions at the right time while remaining silent when intervention is unnecessary. Experiments show that direct VLM baselines struggle with task awareness and intervention calibration, while MindClaw achieves the best overall performance, demonstrating the importance of trigger-skill optimization for closed-loop embodied ToM assistance.

Yi-Syuan Chen, Yun-Zhu Song, Cheng Yu Yeo et al.

Large Pre-trained Transformers exhibit an intriguing capacity for in-context learning. Without gradient updates, these models can rapidly construct new predictors from demonstrations presented in the inputs. Recent works promote this ability in the vision-language domain by incorporating visual information into large language models that can already make in-context predictions. However, these methods could inherit issues in the language domain, such as template sensitivity and hallucination. Also, the scale of these language models raises a significant demand for computations, making learning and operating these models resource-intensive. To this end, we raise a question: ``How can we enable in-context learning without relying on the intrinsic in-context ability of large language models?". To answer it, we propose a succinct and general framework, Self-supervised IN-Context learning (SINC), that introduces a meta-model to learn on self-supervised prompts consisting of tailored demonstrations. The learned models can be transferred to downstream tasks for making in-context predictions on-the-fly. Extensive experiments show that SINC outperforms gradient-based methods in various vision-language tasks under few-shot settings. Furthermore, the designs of SINC help us investigate the benefits of in-context learning across different tasks, and the analysis further reveals the essential components for the emergence of in-context learning in the vision-language domain.

Yuchong Sun, Bei Liu, Xu Chen et al.

Engaging video comments play an important role in video social media, as they are the carrier of feelings, thoughts, or humor of the audience. Preliminary works have made initial exploration for video comment generation by adopting caption-style encoder-decoder models. However, comment generation presents some unique challenges distinct from caption generation, which makes these methods somewhat less effective at generating engaging comments. In contrast to the objective and descriptive nature of captions, comments tend to be inherently subjective, making it hard to quantify and evaluate the engagement of comments. Furthermore, the scarcity of truly engaging comments brings difficulty to collecting enough high-quality training examples. In this paper, we propose ViCo with three novel designs to tackle the above challenges for generating engaging Video Comments. Firstly, to quantify the engagement of comments, we utilize the number of "likes" each comment receives as a proxy of human preference after an appropriate debiasing procedure. Secondly, to automatically evaluate the engagement of comments, we train a reward model to align its judgment to the above proxy. Our user studies indicate that this reward model effectively aligns with human judgments. Lastly, to alleviate the scarcity of high-quality comments, an initial generator is trained on readily available but noisy data to generate comments. Then the reward model is employed to offer feedback on the generated comments, thus optimizing the initial generator. To facilitate the research of video commenting, we collect a large video comment-dataset (ViCo-20k) with rich metadata from a popular video website. Experiments on ViCo-20k show that the comments generated by our ViCo model exhibit the best performance in terms of both quantitative and qualitative results, particularly when engagement is considered.

Sipeng Zheng, Qi Zhang, Bei Liu et al.

In this paper we provide the technique report of Ego4D natural language query challenge in CVPR 2022. Natural language query task is challenging due to the requirement of comprehensive understanding of video contents. Most previous works address this task based on third-person view datasets while few research interest has been placed in the ego-centric view by far. Great progress has been made though, we notice that previous works can not adapt well to ego-centric view datasets e.g., Ego4D mainly because of two reasons: 1) most queries in Ego4D have a excessively small temporal duration (e.g., less than 5 seconds); 2) queries in Ego4D are faced with much more complex video understanding of long-term temporal orders. Considering these, we propose our solution of this challenge to solve the above issues.

Seogkyu Jeon, Bei Liu, Pilhyeon Lee et al.

Training deep generative models usually requires a large amount of data. To alleviate the data collection cost, the task of zero-shot GAN adaptation aims to reuse well-trained generators to synthesize images of an unseen target domain without any further training samples. Due to the data absence, the textual description of the target domain and the vision-language models, e.g., CLIP, are utilized to effectively guide the generator. However, with only a single representative text feature instead of real images, the synthesized images gradually lose diversity as the model is optimized, which is also known as mode collapse. To tackle the problem, we propose a novel method to find semantic variations of the target text in the CLIP space. Specifically, we explore diverse semantic variations based on the informative text feature of the target domain while regularizing the uncontrolled deviation of the semantic information. With the obtained variations, we design a novel directional moment loss that matches the first and second moments of image and text direction distributions. Moreover, we introduce elastic weight consolidation and a relation consistency loss to effectively preserve valuable content information from the source domain, e.g., appearances. Through extensive experiments, we demonstrate the efficacy of the proposed methods in ensuring sample diversity in various scenarios of zero-shot GAN adaptation. We also conduct ablation studies to validate the effect of each proposed component. Notably, our model achieves a new state-of-the-art on zero-shot GAN adaptation in terms of both diversity and quality.

Qixiu Li, Yaobo Liang, Zeyu Wang et al.

The advancement of large Vision-Language-Action (VLA) models has significantly improved robotic manipulation in terms of language-guided task execution and generalization to unseen scenarios. While existing VLAs adapted from pretrained large Vision-Language-Models (VLM) have demonstrated promising generalizability, their task performance is still unsatisfactory as indicated by the low tasks success rates in different environments. In this paper, we present a new advanced VLA architecture derived from VLM. Unlike previous works that directly repurpose VLM for action prediction by simple action quantization, we propose a omponentized VLA architecture that has a specialized action module conditioned on VLM output. We systematically study the design of the action module and demonstrates the strong performance enhancement with diffusion action transformers for action sequence modeling, as well as their favorable scaling behaviors. We also conduct comprehensive experiments and ablation studies to evaluate the efficacy of our models with varied designs. The evaluation on 5 robot embodiments in simulation and real work shows that our model not only significantly surpasses existing VLAs in task performance and but also exhibits remarkable adaptation to new robots and generalization to unseen objects and backgrounds. It exceeds the average success rates of OpenVLA which has similar model size (7B) with ours by over 35% in simulated evaluation and 55% in real robot experiments. It also outperforms the large RT-2-X model (55B) by 18% absolute success rates in simulation. Code and models can be found on our project page (https://cogact.github.io/).

Wenjing Wang, Huan Yang, Jianlong Fu et al.

Understanding illumination and reducing the need for supervision pose a significant challenge in low-light enhancement. Current approaches are highly sensitive to data usage during training and illumination-specific hyper-parameters, limiting their ability to handle unseen scenarios. In this paper, we propose a new zero-reference low-light enhancement framework trainable solely with normal light images. To accomplish this, we devise an illumination-invariant prior inspired by the theory of physical light transfer. This prior serves as the bridge between normal and low-light images. Then, we develop a prior-to-image framework trained without low-light data. During testing, this framework is able to restore our illumination-invariant prior back to images, automatically achieving low-light enhancement. Within this framework, we leverage a pretrained generative diffusion model for model ability, introduce a bypass decoder to handle detail distortion, as well as offer a lightweight version for practicality. Extensive experiments demonstrate our framework's superiority in various scenarios as well as good interpretability, robustness, and efficiency. Code is available on our project homepage: http://daooshee.github.io/QuadPrior-Website/

Wenjing Wang, Huan Yang, Zixi Tuo et al.

With the explosive popularity of AI-generated content (AIGC), video generation has recently received a lot of attention. Generating videos guided by text instructions poses significant challenges, such as modeling the complex relationship between space and time, and the lack of large-scale text-video paired data. Existing text-video datasets suffer from limitations in both content quality and scale, or they are not open-source, rendering them inaccessible for study and use. For model design, previous approaches extend pretrained text-to-image generation models by adding temporal 1D convolution/attention modules for video generation. However, these approaches overlook the importance of jointly modeling space and time, inevitably leading to temporal distortions and misalignment between texts and videos. In this paper, we propose a novel approach that strengthens the interaction between spatial and temporal perceptions. In particular, we utilize a swapped cross-attention mechanism in 3D windows that alternates the "query" role between spatial and temporal blocks, enabling mutual reinforcement for each other. Moreover, to fully unlock model capabilities for high-quality video generation and promote the development of the field, we curate a large-scale and open-source video dataset called HD-VG-130M. This dataset comprises 130 million text-video pairs from the open-domain, ensuring high-definition, widescreen and watermark-free characters. A smaller-scale yet more meticulously cleaned subset further enhances the data quality, aiding models in achieving superior performance. Experimental quantitative and qualitative results demonstrate the superiority of our approach in terms of per-frame quality, temporal correlation, and text-video alignment, with clear margins.

Minghao Chen, Kan Wu, Bolin Ni et al.

Vision Transformer has shown great visual representation power in substantial vision tasks such as recognition and detection, and thus been attracting fast-growing efforts on manually designing more effective architectures. In this paper, we propose to use neural architecture search to automate this process, by searching not only the architecture but also the search space. The central idea is to gradually evolve different search dimensions guided by their E-T Error computed using a weight-sharing supernet. Moreover, we provide design guidelines of general vision transformers with extensive analysis according to the space searching process, which could promote the understanding of vision transformer. Remarkably, the searched models, named S3 (short for Searching the Search Space), from the searched space achieve superior performance to recently proposed models, such as Swin, DeiT and ViT, when evaluated on ImageNet. The effectiveness of S3 is also illustrated on object detection, semantic segmentation and visual question answering, demonstrating its generality to downstream vision and vision-language tasks. Code and models will be available at https://github.com/microsoft/Cream.

Yupan Huang, Bei Liu, Jianlong Fu et al.

A creative image-and-text generative AI system mimics humans' extraordinary abilities to provide users with diverse and comprehensive caption suggestions, as well as rich image creations. In this work, we demonstrate such an AI creation system to produce both diverse captions and rich images. When users imagine an image and associate it with multiple captions, our system paints a rich image to reflect all captions faithfully. Likewise, when users upload an image, our system depicts it with multiple diverse captions. We propose a unified multi-modal framework to achieve this goal. Specifically, our framework jointly models image-and-text representations with a Transformer network, which supports rich image creation by accepting multiple captions as input. We consider the relations among input captions to encourage diversity in training and adopt a non-autoregressive decoding strategy to enable real-time inference. Based on these, our system supports both diverse captions and rich images generations. Our code is available online.

Kan Wu, Houwen Peng, Minghao Chen et al.

Relative position encoding (RPE) is important for transformer to capture sequence ordering of input tokens. General efficacy has been proven in natural language processing. However, in computer vision, its efficacy is not well studied and even remains controversial, e.g., whether relative position encoding can work equally well as absolute position? In order to clarify this, we first review existing relative position encoding methods and analyze their pros and cons when applied in vision transformers. We then propose new relative position encoding methods dedicated to 2D images, called image RPE (iRPE). Our methods consider directional relative distance modeling as well as the interactions between queries and relative position embeddings in self-attention mechanism. The proposed iRPE methods are simple and lightweight. They can be easily plugged into transformer blocks. Experiments demonstrate that solely due to the proposed encoding methods, DeiT and DETR obtain up to 1.5% (top-1 Acc) and 1.3% (mAP) stable improvements over their original versions on ImageNet and COCO respectively, without tuning any extra hyperparameters such as learning rate and weight decay. Our ablation and analysis also yield interesting findings, some of which run counter to previous understanding. Code and models are open-sourced at https://github.com/microsoft/Cream/tree/main/iRPE.

Minghao Chen, Houwen Peng, Jianlong Fu et al.

Recently, pure transformer-based models have shown great potentials for vision tasks such as image classification and detection. However, the design of transformer networks is challenging. It has been observed that the depth, embedding dimension, and number of heads can largely affect the performance of vision transformers. Previous models configure these dimensions based upon manual crafting. In this work, we propose a new one-shot architecture search framework, namely AutoFormer, dedicated to vision transformer search. AutoFormer entangles the weights of different blocks in the same layers during supernet training. Benefiting from the strategy, the trained supernet allows thousands of subnets to be very well-trained. Specifically, the performance of these subnets with weights inherited from the supernet is comparable to those retrained from scratch. Besides, the searched models, which we refer to AutoFormers, surpass the recent state-of-the-arts such as ViT and DeiT. In particular, AutoFormer-tiny/small/base achieve 74.7%/81.7%/82.4% top-1 accuracy on ImageNet with 5.7M/22.9M/53.7M parameters, respectively. Lastly, we verify the transferability of AutoFormer by providing the performance on downstream benchmarks and distillation experiments. Code and models are available at https://github.com/microsoft/AutoML.

Bin Yan, Houwen Peng, Kan Wu et al.

Object tracking has achieved significant progress over the past few years. However, state-of-the-art trackers become increasingly heavy and expensive, which limits their deployments in resource-constrained applications. In this work, we present LightTrack, which uses neural architecture search (NAS) to design more lightweight and efficient object trackers. Comprehensive experiments show that our LightTrack is effective. It can find trackers that achieve superior performance compared to handcrafted SOTA trackers, such as SiamRPN++ and Ocean, while using much fewer model Flops and parameters. Moreover, when deployed on resource-constrained mobile chipsets, the discovered trackers run much faster. For example, on Snapdragon 845 Adreno GPU, LightTrack runs $12\times$ faster than Ocean, while using $13\times$ fewer parameters and $38\times$ fewer Flops. Such improvements might narrow the gap between academic models and industrial deployments in object tracking task. LightTrack is released at https://github.com/researchmm/LightTrack.

Yanhong Zeng, Jianlong Fu, Hongyang Chao et al.

State-of-the-art image inpainting approaches can suffer from generating distorted structures and blurry textures in high-resolution images (e.g., 512x512). The challenges mainly drive from (1) image content reasoning from distant contexts, and (2) fine-grained texture synthesis for a large missing region. To overcome these two challenges, we propose an enhanced GAN-based model, named Aggregated COntextual-Transformation GAN (AOT-GAN), for high-resolution image inpainting. Specifically, to enhance context reasoning, we construct the generator of AOT-GAN by stacking multiple layers of a proposed AOT block. The AOT blocks aggregate contextual transformations from various receptive fields, allowing to capture both informative distant image contexts and rich patterns of interest for context reasoning. For improving texture synthesis, we enhance the discriminator of AOT-GAN by training it with a tailored mask-prediction task. Such a training objective forces the discriminator to distinguish the detailed appearances of real and synthesized patches, and in turn, facilitates the generator to synthesize clear textures. Extensive comparisons on Places2, the most challenging benchmark with 1.8 million high-resolution images of 365 complex scenes, show that our model outperforms the state-of-the-art by a significant margin in terms of FID with 38.60% relative improvement. A user study including more than 30 subjects further validates the superiority of AOT-GAN. We further evaluate the proposed AOT-GAN in practical applications, e.g., logo removal, face editing, and object removal. Results show that our model achieves promising completions in the real world. We release code and models in https://github.com/researchmm/AOT-GAN-for-Inpainting.

Minghao Chen, Houwen Peng, Jianlong Fu et al.

Despite remarkable progress achieved, most neural architecture search (NAS) methods focus on searching for one single accurate and robust architecture. To further build models with better generalization capability and performance, model ensemble is usually adopted and performs better than stand-alone models. Inspired by the merits of model ensemble, we propose to search for multiple diverse models simultaneously as an alternative way to find powerful models. Searching for ensembles is non-trivial and has two key challenges: enlarged search space and potentially more complexity for the searched model. In this paper, we propose a one-shot neural ensemble architecture search (NEAS) solution that addresses the two challenges. For the first challenge, we introduce a novel diversity-based metric to guide search space shrinking, considering both the potentiality and diversity of candidate operators. For the second challenge, we enable a new search dimension to learn layer sharing among different models for efficiency purposes. The experiments on ImageNet clearly demonstrate that our solution can improve the supernet's capacity of ranking ensemble architectures, and further lead to better search results. The discovered architectures achieve superior performance compared with state-of-the-arts such as MobileNetV3 and EfficientNet families under aligned settings. Moreover, we evaluate the generalization ability and robustness of our searched architecture on the COCO detection benchmark and achieve a 3.1% improvement on AP compared with MobileNetV3. Codes and models are available at https://github.com/researchmm/NEAS.

Bin Yan, Houwen Peng, Jianlong Fu et al.

In this paper, we present a new tracking architecture with an encoder-decoder transformer as the key component. The encoder models the global spatio-temporal feature dependencies between target objects and search regions, while the decoder learns a query embedding to predict the spatial positions of the target objects. Our method casts object tracking as a direct bounding box prediction problem, without using any proposals or predefined anchors. With the encoder-decoder transformer, the prediction of objects just uses a simple fully-convolutional network, which estimates the corners of objects directly. The whole method is end-to-end, does not need any postprocessing steps such as cosine window and bounding box smoothing, thus largely simplifying existing tracking pipelines. The proposed tracker achieves state-of-the-art performance on five challenging short-term and long-term benchmarks, while running at real-time speed, being 6x faster than Siam R-CNN. Code and models are open-sourced at https://github.com/researchmm/Stark.

Houwen Peng, Hao Du, Hongyuan Yu et al.

One-shot weight sharing methods have recently drawn great attention in neural architecture search due to high efficiency and competitive performance. However, weight sharing across models has an inherent deficiency, i.e., insufficient training of subnetworks in hypernetworks. To alleviate this problem, we present a simple yet effective architecture distillation method. The central idea is that subnetworks can learn collaboratively and teach each other throughout the training process, aiming to boost the convergence of individual models. We introduce the concept of prioritized path, which refers to the architecture candidates exhibiting superior performance during training. Distilling knowledge from the prioritized paths is able to boost the training of subnetworks. Since the prioritized paths are changed on the fly depending on their performance and complexity, the final obtained paths are the cream of the crop. We directly select the most promising one from the prioritized paths as the final architecture, without using other complex search methods, such as reinforcement learning or evolution algorithms. The experiments on ImageNet verify such path distillation method can improve the convergence ratio and performance of the hypernetwork, as well as boosting the training of subnetworks. The discovered architectures achieve superior performance compared to the recent MobileNetV3 and EfficientNet families under aligned settings. Moreover, the experiments on object detection and more challenging search space show the generality and robustness of the proposed method. Code and models are available at https://github.com/microsoft/cream.git.

Yanhong Zeng, Jianlong Fu, Hongyang Chao

High-quality video inpainting that completes missing regions in video frames is a promising yet challenging task. State-of-the-art approaches adopt attention models to complete a frame by searching missing contents from reference frames, and further complete whole videos frame by frame. However, these approaches can suffer from inconsistent attention results along spatial and temporal dimensions, which often leads to blurriness and temporal artifacts in videos. In this paper, we propose to learn a joint Spatial-Temporal Transformer Network (STTN) for video inpainting. Specifically, we simultaneously fill missing regions in all input frames by self-attention, and propose to optimize STTN by a spatial-temporal adversarial loss. To show the superiority of the proposed model, we conduct both quantitative and qualitative evaluations by using standard stationary masks and more realistic moving object masks. Demo videos are available at https://github.com/researchmm/STTN.

Hongyuan Yu, Houwen Peng, Yan Huang et al.

Differentiable ARchiTecture Search, i.e., DARTS, has drawn great attention in neural architecture search. It tries to find the optimal architecture in a shallow search network and then measures its performance in a deep evaluation network. The independent optimization of the search and evaluation networks, however, leaves room for potential improvement by allowing interaction between the two networks. To address the problematic optimization issue, we propose new joint optimization objectives and a novel Cyclic Differentiable ARchiTecture Search framework, dubbed CDARTS. Considering the structure difference, CDARTS builds a cyclic feedback mechanism between the search and evaluation networks with introspective distillation. First, the search network generates an initial architecture for evaluation, and the weights of the evaluation network are optimized. Second, the architecture weights in the search network are further optimized by the label supervision in classification, as well as the regularization from the evaluation network through feature distillation. Repeating the above cycle results in joint optimization of the search and evaluation networks and thus enables the evolution of the architecture to fit the final evaluation network. The experiments and analysis on CIFAR, ImageNet and NAS-Bench-201 demonstrate the effectiveness of the proposed approach over the state-of-the-art ones. Specifically, in the DARTS search space, we achieve 97.52% top-1 accuracy on CIFAR10 and 76.3% top-1 accuracy on ImageNet. In the chain-structured search space, we achieve 78.2% top-1 accuracy on ImageNet, which is 1.1% higher than EfficientNet-B0. Our code and models are publicly available at https://github.com/microsoft/Cream.

Zhipeng Zhang, Houwen Peng, Jianlong Fu et al.

Anchor-based Siamese trackers have achieved remarkable advancements in accuracy, yet the further improvement is restricted by the lagged tracking robustness. We find the underlying reason is that the regression network in anchor-based methods is only trained on the positive anchor boxes (i.e., $IoU \geq0.6$). This mechanism makes it difficult to refine the anchors whose overlap with the target objects are small. In this paper, we propose a novel object-aware anchor-free network to address this issue. First, instead of refining the reference anchor boxes, we directly predict the position and scale of target objects in an anchor-free fashion. Since each pixel in groundtruth boxes is well trained, the tracker is capable of rectifying inexact predictions of target objects during inference. Second, we introduce a feature alignment module to learn an object-aware feature from predicted bounding boxes. The object-aware feature can further contribute to the classification of target objects and background. Moreover, we present a novel tracking framework based on the anchor-free model. The experiments show that our anchor-free tracker achieves state-of-the-art performance on five benchmarks, including VOT-2018, VOT-2019, OTB-100, GOT-10k and LaSOT. The source code is available at https://github.com/researchmm/TracKit.

Hokyun Im, Euijin Jeong, Jianlong Fu et al.

Vision-language-action models (VLAs) trained on large-scale robotic datasets have demonstrated strong performance on manipulation tasks, including bimanual tasks. However, because most public datasets focus on single-arm demonstrations, adapting VLAs for bimanual tasks typically requires substantial additional bimanual data and fine-tuning. To address this challenge, we introduce TwinVLA, a modular framework that composes two copies of a pretrained single-arm VLA into a coordinated bimanual VLA. Unlike monolithic cross-embodiment models trained on mixtures of single-arm and bimanual data, TwinVLA improves both data efficiency and performance by composing pretrained single-arm policies. Across diverse bimanual tasks in real-world and simulation settings, TwinVLA outperforms a comparably-sized monolithic RDT-1B model without requiring any bimanual pretraining. Furthermore, it narrows the gap to state-of-the-art model, $π_0$ which rely on extensive proprietary bimanual data and compute cost. These results establish our modular composition approach as a data-efficient and scalable path toward high-performance bimanual manipulation, leveraging public single-arm data.

Hokyun Im, Andrey Kolobov, Jianlong Fu et al.

Offline reinforcement learning (RL) allows robots to learn from offline datasets without risky exploration. Yet, offline RL's performance often hinges on a brittle trade-off between (1) return maximization, which can push policies outside the dataset support, and (2) behavioral constraints, which typically require sensitive hyperparameter tuning. Latent steering offers a structural way to stay within the dataset support during RL, but existing offline adaptations commonly approximate action values using latent-space critics learned via indirect distillation, which can lose information and hinder convergence. We propose Latent Policy Steering (LPS), which enables high-fidelity latent policy improvement by backpropagating original-action-space Q-gradients through a differentiable one-step MeanFlow policy to update a latent-action-space actor. By eliminating proxy latent critics, LPS allows an original-action-space critic to guide end-to-end latent-space optimization, while the one-step MeanFlow policy serves as a behavior-constrained generative prior. This decoupling yields a robust method that works out-of-the-box with minimal tuning. Across OGBench and real-world robotic tasks, LPS achieves state-of-the-art performance and consistently outperforms behavioral cloning and strong latent steering baselines.

Ruoxuan Zhang, Qiyun Zheng, Zhiyu Zhou et al.

Theory of Mind (ToM) refers to the ability to infer others' mental states, such as beliefs, desires, and intentions. Current vision-language embodied agents lack ToM-based decision-making, and existing benchmarks focus solely on human mental states while ignoring the agent's own perspective, hindering coherent decision and action generation. To address this, we propose MindPower, a Robot-Centric framework integrating Perception, Mental Reasoning, Decision Making and Action. Given multimodal inputs, MindPower first perceives the environment and human states, then performs ToM Reasoning to model both self and others, and finally generates decisions and actions guided by inferred mental states. Furthermore, we introduce Mind-Reward, a novel optimization objective that encourages VLMs to produce consistent ToM Reasoning and behavior. Our model outperforms GPT-4o by 12.77% in decision making and 12.49% in action generation.

Bo Zhao, Huan Yang, Jianlong Fu

Face inpainting requires the model to have a precise global understanding of the facial position structure. Benefiting from the powerful capabilities of deep learning backbones, recent works in face inpainting have achieved decent performance in ideal setting (square shape with $512px$). However, existing methods often produce a visually unpleasant result, especially in the position-sensitive details (e.g., eyes and nose), when directly applied to arbitrary-shaped images in real-world scenarios. The visually unpleasant position-sensitive details indicate the shortcomings of existing methods in terms of position information processing capability. In this paper, we propose an \textbf{I}mplicit \textbf{N}eural \textbf{I}npainting \textbf{N}etwork (IN$^2$) to handle arbitrary-shape face images in real-world scenarios by explicit modeling for position information. Specifically, a downsample processing encoder is proposed to reduce information loss while obtaining the global semantic feature. A neighbor hybrid attention block is proposed with a hybrid attention mechanism to improve the facial understanding ability of the model without restricting the shape of the input. Finally, an implicit neural pyramid decoder is introduced to explicitly model position information and bridge the gap between low-resolution features and high-resolution output. Extensive experiments demonstrate the superiority of the proposed method in real-world face inpainting task.

Chuhao Jin, Wenhui Tan, Jiange Yang et al.

We propose a novel framework for learning high-level cognitive capabilities in robot manipulation tasks, such as making a smiley face using building blocks. These tasks often involve complex multi-step reasoning, presenting significant challenges due to the limited paired data connecting human instructions (e.g., making a smiley face) and robot actions (e.g., end-effector movement). Existing approaches relieve this challenge by adopting an open-loop paradigm decomposing high-level instructions into simple sub-task plans, and executing them step-by-step using low-level control models. However, these approaches are short of instant observations in multi-step reasoning, leading to sub-optimal results. To address this issue, we propose to automatically collect a cognitive robot dataset by Large Language Models (LLMs). The resulting dataset AlphaBlock consists of 35 comprehensive high-level tasks of multi-step text plans and paired observation sequences. To enable efficient data acquisition, we employ elaborated multi-round prompt designs that effectively reduce the burden of extensive human involvement. We further propose a closed-loop multi-modal embodied planning model that autoregressively generates plans by taking image observations as input. To facilitate effective learning, we leverage MiniGPT-4 with a frozen visual encoder and LLM, and finetune additional vision adapter and Q-former to enable fine-grained spatial perception for manipulation tasks. We conduct experiments to verify the superiority over existing open and closed-loop methods, and achieve a significant increase in success rate by 21.4% and 14.5% over ChatGPT and GPT-4 based robot tasks. Real-world demos are shown in https://www.youtube.com/watch?v=ayAzID1_qQk .

Yiyang Ma, Huan Yang, Wenhan Yang et al.

Diffusion models, as a kind of powerful generative model, have given impressive results on image super-resolution (SR) tasks. However, due to the randomness introduced in the reverse process of diffusion models, the performances of diffusion-based SR models are fluctuating at every time of sampling, especially for samplers with few resampled steps. This inherent randomness of diffusion models results in ineffectiveness and instability, making it challenging for users to guarantee the quality of SR results. However, our work takes this randomness as an opportunity: fully analyzing and leveraging it leads to the construction of an effective plug-and-play sampling method that owns the potential to benefit a series of diffusion-based SR methods. More in detail, we propose to steadily sample high-quality SR images from pre-trained diffusion-based SR models by solving diffusion ordinary differential equations (diffusion ODEs) with optimal boundary conditions (BCs) and analyze the characteristics between the choices of BCs and their corresponding SR results. Our analysis shows the route to obtain an approximately optimal BC via an efficient exploration in the whole space. The quality of SR results sampled by the proposed method with fewer steps outperforms the quality of results sampled by current methods with randomness from the same pre-trained diffusion-based SR model, which means that our sampling method "boosts" current diffusion-based SR models without any additional training.