Recent advances in visual generative models have enabled high-fidelity image editing guided by human instructions. However, these models often struggle with complex instructions involving combinatorial editing operations or inter-step dependencies. This difficulty stems from the limitations of two canonical paradigms: (1) single-turn editing, which attempts to apply all instructed edits in one pass, often fails to parse the complex instruction accurately and causes undesired edits; and (2) sequential editing can decompose the task into simpler steps but suffers from compounding errors introduced by the sequential execution, leading to low-fidelity results. To derive a robust solution for complex image editing, we examine editing behaviors of different paradigms under a unified in-context editing framework, and study how the benefits of sequential decomposition can be balanced against its error-accumulation drawbacks. We further develop a synthetic data pipeline that constructs editing tasks of varying instruction complexity, allowing us to curate a large-scale editing dataset with high-quality decomposed sequences. By finetuning on synthetic data, we discovered that with properly designed editing paradigms, sequential decomposition yields robust improvements even as task complexity increases. Furthermore, the decomposition skills learned from synthetic tasks can transfer to real images by co-training with real-world editing data, demonstrating the promise of sim-to-real generalization for tackling complex image editing across broader domains.

We present MMCORE, a unified framework designed for multimodal image generation and editing. MMCORE leverages a pre-trained Vision-Language Model (VLM) to predict semantic visual embeddings via learnable query tokens, which subsequently serve as conditioning signals for a diffusion model. This streamlined design effectively transfers the rich understanding and reasoning capabilities of VLMs into the visual generation process. By obviating the need for deep fusion between autoregressive and diffusion models or training from scratch, MMCORE significantly reduces computational overhead while maintaining high-fidelity synthesis. MMCORE seamlessly integrates text-to-image synthesis with interleaved image generation, demonstrating robust multimodal comprehension in complex scenarios such as spatial reasoning and visual grounding. Comprehensive evaluations indicate that MMCORE consistently outperforms state-of-the-art baselines across a broad spectrum of text-to-image and single/multi-image editing benchmarks.

Rapid advancement of diffusion models has catalyzed remarkable progress in the field of image generation. However, prevalent models such as Flux, SD3.5 and Midjourney, still grapple with issues like model bias, limited text rendering capabilities, and insufficient understanding of Chinese cultural nuances. To address these limitations, we present Seedream 2.0, a native Chinese-English bilingual image generation foundation model that excels across diverse dimensions, which adeptly manages text prompt in both Chinese and English, supporting bilingual image generation and text rendering. We develop a powerful data system that facilitates knowledge integration, and a caption system that balances the accuracy and richness for image description. Particularly, Seedream is integrated with a self-developed bilingual large language model as a text encoder, allowing it to learn native knowledge directly from massive data. This enable it to generate high-fidelity images with accurate cultural nuances and aesthetic expressions described in either Chinese or English. Beside, Glyph-Aligned ByT5 is applied for flexible character-level text rendering, while a Scaled ROPE generalizes well to untrained resolutions. Multi-phase post-training optimizations, including SFT and RLHF iterations, further improve the overall capability. Through extensive experimentation, we demonstrate that Seedream 2.0 achieves state-of-the-art performance across multiple aspects, including prompt-following, aesthetics, text rendering, and structural correctness. Furthermore, Seedream 2.0 has been optimized through multiple RLHF iterations to closely align its output with human preferences, as revealed by its outstanding ELO score. In addition, it can be readily adapted to an instruction-based image editing model, such as SeedEdit, with strong editing capability that balances instruction-following and image consistency.

We introduce SeedEdit 3.0, in companion with our T2I model Seedream 3.0 [22], which significantly improves over our previous version [27] in both aspects of edit instruction following and image content (e.g., ID/IP) preservation on real image inputs. Additional to model upgrading with T2I, in this report, we present several key improvements. First, we develop an enhanced data curation pipeline with a meta-info paradigm and meta-info embedding strategy that help mix images from multiple data sources. This allows us to scale editing data effectively, and meta information is helpfult to connect VLM with diffusion model more closely. Second, we introduce a joint learning pipeline for computing a diffusion loss and a reward loss. Finally, we evaluate SeedEdit 3.0 on our testing benchmarks, for real image editing, where it achieves a best trade-off between multiple aspects, yielding a high usability rate of 56.1 compared to SeedEdit 1.6 (38.4

The semantic segmentation of nighttime scenes is a challenging problem that is key to impactful applications like self-driving cars. Yet, it has received little attention compared to its daytime counterpart. In this paper, we propose NightLab, a novel nighttime segmentation framework that leverages multiple deep learning models imbued with night-aware features to yield State-of-The-Art (SoTA) performance on multiple night segmentation benchmarks. Notably, NightLab contains models at two levels of granularity, i.e. image and regional, and each level is composed of light adaptation and segmentation modules. Given a nighttime image, the image level model provides an initial segmentation estimate while, in parallel, a hardness detection module identifies regions and their surrounding context that need further analysis. A regional level model focuses on these difficult regions to provide a significantly improved segmentation. All the models in NightLab are trained end-to-end using a set of proposed night-aware losses without handcrafted heuristics. Extensive experiments on the NightCity [44] and BDD100K [59] datasets show NightLab achieves SoTA performance compared to concurrent methods. Code and dataset are available at https://github.com/xdeng7/NightLab.

High-resolution representations (HR) are essential for dense prediction tasks such as segmentation, detection, and pose estimation. Learning HR representations is typically ignored in previous Neural Architecture Search (NAS) methods that focus on image classification. This work proposes a novel NAS method, called HR-NAS, which is able to find efficient and accurate networks for different tasks, by effectively encoding multiscale contextual information while maintaining high-resolution representations. In HR-NAS, we renovate the NAS search space as well as its searching strategy. To better encode multiscale image contexts in the search space of HR-NAS, we first carefully design a lightweight transformer, whose computational complexity can be dynamically changed with respect to different objective functions and computation budgets. To maintain high-resolution representations of the learned networks, HR-NAS adopts a multi-branch architecture that provides convolutional encoding of multiple feature resolutions, inspired by HRNet [73]. Last, we proposed an efficient fine-grained search strategy to train HR-NAS, which effectively explores the search space, and finds optimal architectures given various tasks and computation resources. As shown in Fig. 1 (a), HR-NAS is capable of achieving state-of-the-art trade-offs between performance and FLOPs for three dense prediction tasks and an image classification task, given only small computational budgets. For example, HR-NAS surpasses SqueezeNAS [63] that is specially designed for semantic segmentation while improving efficiency by 45.9%. Code is available at https://github.com/dingmyu/HR-NAS.

Current neural architecture search (NAS) algorithms still require expert knowledge and effort to design a search space for network construction. In this paper, we consider automating the search space design to minimize human interference, which however faces two challenges: the ex-plosive complexity of the exploration space and the expensive computation cost to evaluate the quality of different search spaces. To solve them, we propose a novel differentiable evolutionary framework named AutoSpace, which evolves the search space to an optimal one with following novel techniques: a differentiable fitness scoring function to efficiently evaluate the performance of cells and a reference architecture to speedup the evolution procedure and avoid falling into sub-optimal solutions. The frame-work is generic and compatible with additional computational constraints, making it feasible to learn specialized search spaces that fit different computational bud-gets. With the learned search space, the performance of recent NAS algorithms can be improved significantly compared with using previously manually designed spaces. Remarkably, the models generated from the new search space achieve 77.8% top-1 accuracy on ImageNet under the mobile setting (MAdds 500M), outperforming previous SOTA EfficientNet-B0 by≤0.7%. https://github.com/zhoudaquan/AutoSpace.git.

Vision transformers (ViTs) have been successfully applied in image classification tasks recently. In this paper, we show that, unlike convolution neural networks (CNNs)that can be improved by stacking more convolutional layers, the performance of ViTs saturate fast when scaled to be deeper. More specifically, we empirically observe that such scaling difficulty is caused by the attention collapse issue: as the transformer goes deeper, the attention maps gradually become similar and even much the same after certain layers. In other words, the feature maps tend to be identical in the top layers of deep ViT models. This fact demonstrates that in deeper layers of ViTs, the self-attention mechanism fails to learn effective concepts for representation learning and hinders the model from getting expected performance gain. Based on above observation, we propose a simple yet effective method, named Re-attention, to re-generate the attention maps to increase their diversity at different layers with negligible computation and memory cost. The pro-posed method makes it feasible to train deeper ViT models with consistent performance improvements via minor modification to existing ViT models. Notably, when training a deep ViT model with 32 transformer blocks, the Top-1 classification accuracy can be improved by 1.6% on ImageNet. Code is publicly available at https://github.com/zhoudaquan/dvit_repo.

Search space design is very critical to neural architecture search (NAS) algorithms. We propose a fine-grained search space comprised of atomic blocks, a minimal search unit that is much smaller than the ones used in recent NAS algorithms. This search space allows a mix of operations by composing different types of atomic blocks, while the search space in previous methods only allows homogeneous operations. Based on this search space, we propose a resource-aware architecture search framework which automatically assigns the computational resources (e.g., output channel numbers) for each operation by jointly considering the performance and the computational cost. In addition, to accelerate the search process, we propose a dynamic network shrinkage technique which prunes the atomic blocks with negligible influence on outputs on the fly. Instead of a search-and-retrain two-stage paradigm, our method simultaneously searches and trains the target architecture. Our method achieves state-of-the-art performance under several FLOPs configurations on ImageNet with a small searching cost. We open our entire codebase at: https://github.com/meijieru/AtomNAS.

Non-Local (NL) blocks have been widely studied in various vision tasks. However, it has been rarely explored to embed the NL blocks in mobile neural networks, mainly due to the following challenges: 1) NL blocks generally have heavy computation cost which makes it difficult to be applied in applications where computational resources are limited, and 2) it is an open problem to discover an optimal configuration to embed NL blocks into mobile neural networks. We propose AutoNL to overcome the above two obstacles. Firstly, we propose a Lightweight Non-Local (LightNL) block by squeezing the transformation operations and incorporating compact features. With the novel design choices, the proposed LightNL block is 400x computationally cheaper than its conventional counterpart without sacrificing the performance. Secondly, by relaxing the structure of the LightNL block to be differentiable during training, we propose an efficient neural architecture search algorithm to learn an optimal configuration of LightNL blocks in an end-to-end manner. Notably, using only 32 GPU hours, the searched AutoNL model achieves 77.7% top-1 accuracy on ImageNet under a typical mobile setting (350M FLOPs), significantly outperforming previous mobile models including MobileNetV2 (+5.7%), FBNet (+2.8%) and MnasNet (+2.1%). Code and models are available a thttps://github.com/LiYingwei/AutoNL.

Recently there has been a rising interest in training agents, embodied in virtual environments, to perform language-directed tasks by deep reinforcement learning. In this paper, we propose a simple but effective neural language grounding module for embodied agents that can be trained end to end from scratch taking raw pixels, unstructured linguistic commands, and sparse rewards as the inputs. We model the language grounding process as a language-guided transformation of visual features, where latent sentence embeddings are used as the transformation matrices. In several language-directed navigation tasks that feature challenging partial observability and require simple reasoning, our module significantly outperforms the state of the art. We also release XWorld3D, an easy-to-customize 3D environment that can potentially be modified to evaluate a variety of embodied agents.

Local information is very crucial in many image and video analysis tasks. In this thesis,we introduce four representative works in exploiting local information. We first introduce aset of per-pixel labeling datasets, which provide a good platform for studies of using localinformation in image analysis. Based on this dataset, we propose a novel segmentationmethod which utilizes local appearance consistency for car semantic part parsing task. Wethen address the attention issue in video action recognition tasks, by designing a latentattention module, which is jointly learned with video recognition components. Last, weimprove the attention mechanism to explicitly detect spatial and spatio-temporal regionsthat are related to actions (ROIs).

In this paper, we address the task of instance-level semantic boundary detection. To this end, we generate a large database consisting of more than 10k images (which is 20 bigger than existing edge detection databases) along with ground truth boundaries between 459 semantic classes including instances from both foreground objects and different types of background, and call it the PASCAL Boundaries dataset. This is a timely introduction of the dataset since results on the existing standard dataset for measuring edge detection performance, i.e. BSDS500, have started to saturate. In addition to generating a new dataset, we propose a deep network-based multi-scale semantic boundary detector and name it Multi-scale Deep Semantic Boundary Detector (M-DSBD). We evaluate M-DSBD on PASCAL boundaries and compare it to baselines. We test the transfer capabilities of our model by evaluating on MS-COCO and BSDS500 and show that our model transfers well.

Author(s): Lian, Xiaochen | Advisor(s): Yuille, Alan Loddon | Abstract: In this paper, we propose an approach to classify action sequences. We observe that in action sequences the critical features for discriminating between actions occur only within sub-regions of the image. Hence deep network approaches will address the entire image are at a disadvantage. This motivates our strategy which uses static and spatio-temporal visual cues to isolate static and spatio-temporal regions of interest (ROIs). We then use weakly supervised learning to train deep network classifiers using the ROIs as input. More specifically, we combine multiple instance learning (MIL) with convolutional neural networks (CNNs) to select discriminative action cues. This yields classifiers for static images, using the static ROIs, as well as classifiers for short image sequences (16 frames), using spatio-temporal ROIs. Extensive experiments performed on the UCF101 and HMDB51 benchmarks show that both these types of classifiers perform well individually and achieve state of the art performance when combined together. We also show qualitatively that our ROIs (selected by the algorithms) capture the most relevant parts of the image sequences.

This paper addresses the problem of semantic part parsing (segmentation) of cars, i.e.assigning every pixel within the car to one of the parts (e.g.body, window, lights, license plates and wheels). We formulate this as a landmark identification problem, where a set of landmarks specifies the boundaries of the parts. A novel mixture of graphical models is proposed, which dynamically couples the landmarks to a hierarchy of segments. When modeling pairwise relation between landmarks, this coupling enables our model to exploit the local image contents in addition to spatial deformation, an aspect that most existing graphical models ignore. In particular, our model enforces appearance consistency between segments within the same part. Parsing the car, including finding the optimal coupling between landmarks and segments in the hierarchy, is performed by dynamic programming. We evaluate our method on a subset of PASCAL VOC 2010 car images and on the car subset of 3D Object Category dataset (CAR3D). We show good results and, in particular, quantify the effectiveness of using the segment appearance consistency in terms of accuracy of part localization and segmentation.

In this paper, we propose a novel gender classification framework, which utilizes not only facial features, but also external information, i.e. hair and clothing. Instead of using the whole face, we consider five facial components: forehead, eyes, nose, mouth and chin. We also design feature extraction methods for hair and clothing; these features have seldom been used in previous work because of their large variability. For each type of feature, we train a single support vector machine classifier with probabilistic output. The outputs of these classifiers are combined using various strategies, namely fuzzy integral, maximal, sum, voting, and product rule. The major contributions of this paper are (1) investigating the gender discriminative ability of clothing information; (2) using facial components instead of the whole face to obtain higher robustness for occlusions and noise; (3) exploiting hair and clothing information to facilitate gender classification. Experimental results show that our proposed framework improves classification accuracy, even when images contain occlusions, noise, and illumination changes.

In this paper, we propose a multi-view gender classification system with a hierarchical framework using facial images as input. The front end of the framework is a classifier, which will properly divides the input images into several groups. To ease the data sparsity problem in the multi-view scenario, facial symmetry is used to reduce the number of views. Moreover, we adopt soft assignment when dividing the input data, which can reduce the errors caused by the boundary effect in hard assignment. Then for each group, we train a gender classifier, called an expert. These experts can be any commonly used classifiers, such as support vector machines or neural networks. In this step, facial components instead of the whole face are used to achieve higher robustness against variations caused by facial alignment, illumination and occlusions. Experimental results demonstrate that our framework significantly improves the performance.

Spectral clustering is an elegant and powerful approach for clustering. However, the underlying eigen-decomposition takes cubic time and quadratic space w.r.t. the data set size. These can be reduced by the Nyström method which samples only a subset of columns from the matrix. However, the manipulation and storage of these sampled columns can still be expensive when the data set is large. In this paper, we propose a time- and space-efficient spectral clustering algorithm which can scale to very large data sets. A general procedure to orthogonalize the approximated eigenvectors is also proposed. Extensive spectral clustering experiments on a number of data sets, ranging in size from a few thousands to several millions, demonstrate the accuracy and scalability of the proposed approach. We further apply it to the task of image segmentation. For images with more than 10 millions pixels, this algorithm can obtain the eigenvectors in 1 minute on a single machine.

Visual dictionary learning and base (binary) classifier training are two basic problems for the recently most popular image categorization framework, which is based on the bag-of-visual-terms (BOV) models and multiclass SVM classifiers. In this paper, we study new algorithms to improve performance of this framework from these two aspects. Typically SVM classifiers are trained with dictionaries fixed, and as a result the traditional loss function can only be minimized with respect to hyperplane parameters (w and b). We propose a novel loss function for a binary classifier, which links the hinge-loss term with dictionary learning. By doing so, we can further optimize the loss function with respect to the dictionary parameters. Thus, this framework is able to further increase margins of binary classifiers, and consequently decrease the error bound of the aggregated classifier. On two benchmark dataset, Graz [1] and the fifteen scene category dataset [2], our experiment results significantly outperformed state-of-the-art works.

Dictionary generation is a core technique of the bag-of-visual-words (BOV) models when applied to image categorization. Most of previous approaches generate dictionaries by unsupervised clustering techniques, e.g. k-means. However, the features obtained by such kind of dictionaries may not be optimal for image classification. In this paper, we propose a probabilistic model for supervised dictionary learning (SDLM) which seamlessly combines an unsupervised model (a Gaussian Mixture Model) and a supervised model (a logistic regression model) in a probabilistic framework. In the model, image category information directly affects the generation of a dictionary. A dictionary obtained by this approach is a trade-off between minimization of distortions of clusters and maximization of discriminative power of image-wise representations, i.e. histogram representations of images. We further extend the model to incorporate spatial information during the dictionary learning process in a spatial pyramid matching like manner. We extensively evaluated the two models on various benchmark dataset and obtained promising results.