This paper proposes a novel deep architecture to address multi-label image recognition, fundamental and practical task towards general visual understanding. Current solutions for this usually rely on an extra step of extracting hypothesis regions (i.e., region proposals), resulting in redundant computation sub-optimal performance. In work, we achieve the interpretable contextualized classification by developing recurrent memorized-attention module. module consists two alternately performed...

Social relationships (e.g., friends, couple etc.) form the basis of social network in our daily life. Automatically interpreting such bears a great potential for intelligent systems to understand human behavior depth and better interact with people at level. Human beings interpret within group not only based on alone, interplay between contextual information around also plays significant role. However, these additional cues are largely overlooked by previous studies. We found that two...

We observed that recent state-of-the-art results on single image human pose estimation were achieved by multistage Convolution Neural Networks (CNN). Notwithstanding the superior performance static images, application of these models videos is not only computationally intensive, it also suffers from degeneration and flicking. Such suboptimal are mainly attributed to inability imposing sequential geometric consistency, handling severe quality degradation (e.g. motion blur occlusion) as well...

Blind face restoration is to recover a high-quality image from unknown degradations. As contains abundant contextual information, we propose method, RestoreFormer, which explores fully-spatial attentions model information and surpasses existing works that use local operators. RestoreFormer has several benefits compared prior arts. First, unlike the conventional multi-head self-attention in previous Vision Transformers (ViTs), incorporates cross-attention layer learn interactions between...

Filmmaking and animation production often require sophisticated techniques for coordinating camera transitions object movements, typically involving labor-intensive real-world capturing. Despite advancements in generative AI video creation, achieving precise control over motion interactive asset generation remains challenging. To this end, we propose Image Conductor, a method of movements to generate assets from single image. An well-cultivated training strategy is proposed separate distinct...

This paper proposes a novel deep architecture to address multi-label image recognition, fundamental and practical task towards general visual understanding. Current solutions for this usually rely on an extra step of extracting hypothesis regions (i.e., region proposals), resulting in redundant computation sub-optimal performance. In work, we achieve the interpretable contextualized classification by developing recurrent memorized-attention module. module consists two alternately performed...

Blind face restoration aims at recovering high-quality images from those with unknown degradations. Current algorithms mainly introduce priors to complement details and achieve impressive progress. However, most of these ignore abundant contextual information in the its interplay priors, leading sub-optimal performance. Moreover, they pay less attention gap between synthetic real-world scenarios, limiting robustness generalization applications. In this work, we propose RestoreFormer++, which...

This paper presents a LoRA-free method for stylized image generation that takes text prompt and style reference images as inputs produces an output in single pass. Unlike existing methods rely on training separate LoRA each style, our can adapt to various styles with unified model. However, this poses two challenges: 1) the loses controllability over generated content, 2) inherits both semantic features of image, compromising its content fidelity. To address these challenges, we introduce...

Automatically selecting exposure bracketing (images exposed differently) is important to obtain a high dynamic range image by using multi-exposure fusion. Unlike previous methods that have many restrictions such as requiring camera response function, sensor noise model, and stream of preview images with different exposures (not accessible in some scenarios e.g. mobile applications), we propose novel deep neural network automatically select bracketing, named EBSNet, which sufficiently...

Social relationships (e.g., friends, couple etc.) form the basis of social network in our daily life. Automatically interpreting such bears a great potential for intelligent systems to understand human behavior depth and better interact with people at level. Human beings interpret within group not only based on alone, interplay between contextual information around also plays significant role. However, these additional cues are largely overlooked by previous studies. We found that two...

In the past a few years, we witnessed rapid advancement in face super-resolution from very low resolution(VLR) images. However, most of previous studies focus on solving such problem without explicitly considering impact severe real-life image degradation (e.g. blur and noise). We can show that robustly recover details VLR images is task beyond ability current state-of-the-art method. this paper, borrow ideas "facial composite" propose an alternative approach to tackle problem. endow...

Due to the limitation of event sensors, spatial resolution data is relatively low compared conventional frame-based camera. However, low-spatial-resolution events recorded by cameras are rich in temporal information which helpful for image deblurring, while intensity images captured frame high and have potential promote quality events. Considering complementarity between images, an alternately performed model proposed this paper deblur high-resolution with help low-resolution This composed...

Diffusion model has demonstrated remarkable capability in video generation, which further sparks interest introducing trajectory control into the generation process. While existing works mainly focus on training-based methods (e.g., conditional adapter), we argue that diffusion itself allows decent over generated content without requiring any training. In this study, introduce a tuning-free framework to achieve trajectory-controllable by imposing guidance both noise construction and...

Filmmaking and animation production often require sophisticated techniques for coordinating camera transitions object movements, typically involving labor-intensive real-world capturing. Despite advancements in generative AI video creation, achieving precise control over motion interactive asset generation remains challenging. To this end, we propose Image Conductor, a method of movements to generate assets from single image. An well-cultivated training strategy is proposed separate distinct...

Although deep learning-based image restoration methods have made significant progress, they still struggle with limited generalization to real-world scenarios due the substantial domain gap caused by training on synthetic data. Existing address this issue improving data synthesis pipelines, estimating degradation kernels, employing internal learning, and performing adaptation regularization. Previous sought bridge learning domain-invariant knowledge in either feature or pixel space. However,...

Recent progress in blind face restoration has resulted producing high-quality restored results for static images. However, efforts to extend these advancements video scenarios have been minimal, partly because of the absence benchmarks that allow a comprehensive and fair comparison. In this work, we first present evaluation benchmark, which introduce Real-world Low-Quality Face Video benchmark (RFV-LQ), evaluate several leading image-based algorithms, conduct thorough systematical analysis...

We observed that recent state-of-the-art results on single image human pose estimation were achieved by multi-stage Convolution Neural Networks (CNN). Notwithstanding the superior performance static images, application of these models videos is not only computationally intensive, it also suffers from degeneration and flicking. Such suboptimal are mainly attributed to inability imposing sequential geometric consistency, handling severe quality degradation (e.g. motion blur occlusion) as well...

Blind face restoration aims at recovering high-quality images from those with unknown degradations. Current algorithms mainly introduce priors to complement details and achieve impressive progress. However, most of these ignore abundant contextual information in the its interplay priors, leading sub-optimal performance. Moreover, they pay less attention gap between synthetic real-world scenarios, limiting robustness generalization applications. In this work, we propose RestoreFormer++, which...

Motions in a video primarily consist of camera motion, induced by movement, and object resulting from movement. Accurate control both motion is essential for generation. However, existing works either mainly focus on one type or do not clearly distinguish between the two, limiting their capabilities diversity. Therefore, this paper presents MotionCtrl, unified flexible controller generation designed to effectively independently motion. The architecture training strategy MotionCtrl are...

Recovering degraded low-resolution text images is challenging, especially for Chinese with complex strokes and severe degradation in real-world scenarios. Ensuring both fidelity style realness crucial high-quality image super-resolution. Recently, diffusion models have achieved great success natural synthesis restoration due to their powerful data distribution modeling abilities generation capabilities. In this work, we propose an Image Diffusion Model (IDM) restore realistic styles. For...

Automatically selecting exposure bracketing (images exposed differently) is important to obtain a high dynamic range image by using multi-exposure fusion. Unlike previous methods that have many restrictions such as requiring camera response function, sensor noise model, and stream of preview images with different exposures (not accessible in some scenarios e.g. mobile applications), we propose novel deep neural network automatically select bracketing, named EBSNet, which sufficiently...