Weilin Huang

Sheng Guo, Zihua Xiong, Yujie Zhong et al.

In this paper, we present a new cross-architecture contrastive learning (CACL) framework for self-supervised video representation learning. CACL consists of a 3D CNN and a video transformer which are used in parallel to generate diverse positive pairs for contrastive learning. This allows the model to learn strong representations from such diverse yet meaningful pairs. Furthermore, we introduce a temporal self-supervised learning module able to predict an Edit distance explicitly between two video sequences in the temporal order. This enables the model to learn a rich temporal representation that compensates strongly to the video-level representation learned by the CACL. We evaluate our method on the tasks of video retrieval and action recognition on UCF101 and HMDB51 datasets, where our method achieves excellent performance, surpassing the state-of-the-art methods such as VideoMoCo and MoCo+BE by a large margin. The code is made available at https://github.com/guoshengcv/CACL.

Team Seedance, De Chen, Liyang Chen et al. · gatech

Seedance 2.0 is a new native multi-modal audio-video generation model, officially released in China in early February 2026. Compared with its predecessors, Seedance 1.0 and 1.5 Pro, Seedance 2.0 adopts a unified, highly efficient, and large-scale architecture for multi-modal audio-video joint generation. This allows it to support four input modalities: text, image, audio, and video, by integrating one of the most comprehensive suites of multi-modal content reference and editing capabilities available in the industry to date. It delivers substantial, well-rounded improvements across all key sub-dimensions of video and audio generation. In both expert evaluations and public user tests, the model has demonstrated performance on par with the leading levels in the field. Seedance 2.0 supports direct generation of audio-video content with durations ranging from 4 to 15 seconds, with native output resolutions of 480p and 720p. For multi-modal inputs as reference, its current open platform supports up to 3 video clips, 9 images, and 3 audio clips. In addition, we provide Seedance 2.0 Fast version, an accelerated variant of Seedance 2.0 designed to boost generation speed for low-latency scenarios. Seedance 2.0 has delivered significant improvements to its foundational generation capabilities and multi-modal generation performance, bringing an enhanced creative experience for end users.

Fengyuan Shi, Ruopeng Gao, Weilin Huang et al.

Multimodal transformer exhibits high capacity and flexibility to align image and text for visual grounding. However, the existing encoder-only grounding framework (e.g., TransVG) suffers from heavy computation due to the self-attention operation with quadratic time complexity. To address this issue, we present a new multimodal transformer architecture, coined as Dynamic Mutilmodal DETR (Dynamic MDETR), by decoupling the whole grounding process into encoding and decoding phases. The key observation is that there exists high spatial redundancy in images. Thus, we devise a new dynamic multimodal transformer decoder by exploiting this sparsity prior to speed up the visual grounding process. Specifically, our dynamic decoder is composed of a 2D adaptive sampling module and a text guided decoding module. The sampling module aims to select these informative patches by predicting the offsets with respect to a reference point, while the decoding module works for extracting the grounded object information by performing cross attention between image features and text features. These two modules are stacked alternatively to gradually bridge the modality gap and iteratively refine the reference point of grounded object, eventually realizing the objective of visual grounding. Extensive experiments on five benchmarks demonstrate that our proposed Dynamic MDETR achieves competitive trade-offs between computation and accuracy. Notably, using only 9% feature points in the decoder, we can reduce ~44% GFLOPs of the multimodal transformer, but still get higher accuracy than the encoder-only counterpart. In addition, to verify its generalization ability and scale up our Dynamic MDETR, we build the first one-stage CLIP empowered visual grounding framework, and achieve the state-of-the-art performance on these benchmarks.

Shiwen Zhang, Shuai Xiao, Weilin Huang

Text-guided image editing on real or synthetic images, given only the original image itself and the target text prompt as inputs, is a very general and challenging task. It requires an editing model to estimate by itself which part of the image should be edited, and then perform either rigid or non-rigid editing while preserving the characteristics of original image. In this paper, we design a novel text-guided image editing method, named as Forgedit. First, we propose a vision-language joint optimization framework capable of reconstructing the original image in 30 seconds, much faster than previous SOTA and much less overfitting. Then we propose a novel vector projection mechanism in text embedding space of Diffusion Models, which is capable to control the identity similarity and editing strength seperately. Finally, we discovered a general property of UNet in Diffusion Models, i.e., Unet encoder learns space and structure, Unet decoder learns appearance and identity. With such a property, we design forgetting mechanisms to successfully tackle the fatal and inevitable overfitting issues when fine-tuning Diffusion Models on one image, thus significantly boosting the editing capability of Diffusion Models. Our method, Forgedit, built on Stable Diffusion, achieves new state-of-the-art results on the challenging text-guided image editing benchmark: TEdBench, surpassing the previous SOTA methods such as Imagic with Imagen, in terms of both CLIP score and LPIPS score. Codes are available at https://github.com/witcherofresearch/Forgedit

Chen Ju, Haicheng Wang, Haozhe Cheng et al.

Vision-Language Large Models (VLMs) recently become primary backbone of AI, due to the impressive performance. However, their expensive computation costs, i.e., throughput and delay, impede potentials in the real-world scenarios. To achieve acceleration for VLMs, most existing methods focus on the model perspective: pruning, distillation, quantization, but completely overlook the data-perspective redundancy. To fill the overlook, this paper pioneers the severity of data redundancy, and designs one plug-and-play Turbo module guided by information degree to prune inefficient tokens from visual or textual data. In pursuit of efficiency-performance trade-offs, information degree takes two crucial factors into consideration: mutual redundancy and semantic value. Concretely, the former evaluates data duplication between sequential tokens; while the latter evaluates each token by its contribution to the overall semantics. As a result, tokens with high information degree carry less redundancy and stronger semantics. For VLMs' calculation, Turbo works as a user-friendly plug-in that sorts data referring to information degree, utilizing only top-level ones to save costs. Its advantages are multifaceted, e.g., being generally compatible to various VLMs across understanding and generation, simple use without re-training and trivial engineering efforts. On multiple VLMs benchmarks, we fully experiment to demonstrate the good acceleration of Turbo, under negligible performance drop.

Junke Wang, Zhi Tian, Xun Wang et al.

This work presents SimpleAR, a vanilla autoregressive visual generation framework without complex architecure modifications. Through careful exploration of training and inference optimization, we demonstrate that: 1) with only 0.5B parameters, our model can generate 1024x1024 resolution images with high fidelity, and achieve competitive results on challenging text-to-image benchmarks, e.g., 0.59 on GenEval and 79.66 on DPG; 2) both supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) training could lead to significant improvements on generation aesthectics and prompt alignment; and 3) when optimized with inference acceleraton techniques like vLLM, the time for SimpleAR to generate an 1024x1024 image could be reduced to around 14 seconds. By sharing these findings and open-sourcing the code, we hope to reveal the potential of autoregressive visual generation and encourage more participation in this research field. Code is available at https://github.com/wdrink/SimpleAR.

Jiacheng Zhang, Jie Wu, Weifeng Chen et al.

In recent years, the field of text-to-video (T2V) generation has made significant strides. Despite this progress, there is still a gap between theoretical advancements and practical application, amplified by issues like degraded image quality and flickering artifacts. Recent advancements in enhancing the video diffusion model (VDM) through feedback learning have shown promising results. However, these methods still exhibit notable limitations, such as misaligned feedback and inferior scalability. To tackle these issues, we introduce OnlineVPO, a more efficient preference learning approach tailored specifically for video diffusion models. Our method features two novel designs, firstly, instead of directly using image-based reward feedback, we leverage the video quality assessment (VQA) model trained on synthetic data as the reward model to provide distribution and modality-aligned feedback on the video diffusion model. Additionally, we introduce an online DPO algorithm to address the off-policy optimization and scalability issue in existing video preference learning frameworks. By employing the video reward model to offer concise video feedback on the fly, OnlineVPO offers effective and efficient preference guidance. Extensive experiments on the open-source video-diffusion model demonstrate OnlineVPO as a simple yet effective and more importantly scalable preference learning algorithm for video diffusion models, offering valuable insights for future advancements in this domain.

Jie Liu, Zilyu Ye, Linxiao Yuan et al.

Unified models capable of interleaved generation have emerged as a promising paradigm, with the community increasingly converging on autoregressive modeling for text and flow matching for image generation. To advance this direction, we propose a unified reinforcement learning framework tailored for interleaved generation. We validate our approach on its fundamental unit: a single round of reasoning-driven image generation, where the model first expands the user prompt through reasoning, followed by image synthesis. Formulating this multimodal generation process as a Markov Decision Process with sparse terminal rewards, we introduce UniGRPO to jointly optimize text and image generation policies using GRPO. Adopting a minimalist methodology to avoid over-design, we leverage established training recipes for both modalities by seamlessly integrating standard GRPO for reasoning and FlowGRPO for visual synthesis. To ensure scalability to multi-round interleaved generation, we introduce two critical modifications to the original FlowGRPO: (1) eliminating classifier-free guidance to maintain linear, unbranched rollouts, which is essential for scaling to complex scenarios involving multi-turn interactions and multi-condition generation (e.g., editing); and (2) replacing the standard latent KL penalty with an MSE penalty directly on the velocity fields, providing a more robust and direct regularization signal to mitigate reward hacking effectively. Our experiments demonstrate that this unified training recipe significantly enhances image generation quality through reasoning, providing a robust and scalable baseline for the future post-training of fully interleaved models.

Yuwei Guo, Ceyuan Yang, Hao He et al.

Autoregressive video diffusion models hold promise for world simulation but are vulnerable to exposure bias arising from the train-test mismatch. While recent works address this via post-training, they typically rely on a bidirectional teacher model or online discriminator. To achieve an end-to-end solution, we introduce Resampling Forcing, a teacher-free framework that enables training autoregressive video models from scratch and at scale. Central to our approach is a self-resampling scheme that simulates inference-time model errors on history frames during training. Conditioned on these degraded histories, a sparse causal mask enforces temporal causality while enabling parallel training with frame-level diffusion loss. To facilitate efficient long-horizon generation, we further introduce history routing, a parameter-free mechanism that dynamically retrieves the top-k most relevant history frames for each query. Experiments demonstrate that our approach achieves performance comparable to distillation-based baselines while exhibiting superior temporal consistency on longer videos owing to native-length training.

Khalil Mrini, Hanlin Lu, Linjie Yang et al.

Text-to-image generation has advanced rapidly, yet aligning complex textual prompts with generated visuals remains challenging, especially with intricate object relationships and fine-grained details. This paper introduces Fast Prompt Alignment (FPA), a prompt optimization framework that leverages a one-pass approach, enhancing text-to-image alignment efficiency without the iterative overhead typical of current methods like OPT2I. FPA uses large language models (LLMs) for single-iteration prompt paraphrasing, followed by fine-tuning or in-context learning with optimized prompts to enable real-time inference, reducing computational demands while preserving alignment fidelity. Extensive evaluations on the COCO Captions and PartiPrompts datasets demonstrate that FPA achieves competitive text-image alignment scores at a fraction of the processing time, as validated through both automated metrics (TIFA, VQA) and human evaluation. A human study with expert annotators further reveals a strong correlation between human alignment judgments and automated scores, underscoring the robustness of FPA's improvements. The proposed method showcases a scalable, efficient alternative to iterative prompt optimization, enabling broader applicability in real-time, high-demand settings. The codebase is provided to facilitate further research: https://github.com/tiktok/fast_prompt_alignment

Xiaojing Dong, Weilin Huang, Liang Li et al.

Rapid advances in text-to-image (T2I) generation have raised higher requirements for evaluation methodologies. Existing benchmarks center on objective capabilities and dimensions, but lack an application-scenario perspective, limiting external validity. Moreover, current evaluations typically rely on either ELO for overall ranking or MOS for dimension-specific scoring, yet both methods have inherent shortcomings and limited interpretability. Therefore, we introduce the Magic Evaluation Framework (MEF), a systematic and practical approach for evaluating T2I models. First, we propose a structured taxonomy encompassing user scenarios, elements, element compositions, and text expression forms to construct the Magic-Bench-377, which supports label-level assessment and ensures a balanced coverage of both user scenarios and capabilities. On this basis, we combine ELO and dimension-specific MOS to generate model rankings and fine-grained assessments respectively. This joint evaluation method further enables us to quantitatively analyze the contribution of each dimension to user satisfaction using multivariate logistic regression. By applying MEF to current T2I models, we obtain a leaderboard and key characteristics of the leading models. We release our evaluation framework and make Magic-Bench-377 fully open-source to advance research in the evaluation of visual generative models.

Zelu Deng, Yujie Zhong, Sheng Guo et al.

This work aims at improving instance retrieval with self-supervision. We find that fine-tuning using the recently developed self-supervised (SSL) learning methods, such as SimCLR and MoCo, fails to improve the performance of instance retrieval. In this work, we identify that the learnt representations for instance retrieval should be invariant to large variations in viewpoint and background etc., whereas self-augmented positives applied by the current SSL methods can not provide strong enough signals for learning robust instance-level representations. To overcome this problem, we propose InsCLR, a new SSL method that builds on the \textit{instance-level} contrast, to learn the intra-class invariance by dynamically mining meaningful pseudo positive samples from both mini-batches and a memory bank during training. Extensive experiments demonstrate that InsCLR achieves similar or even better performance than the state-of-the-art SSL methods on instance retrieval. Code is available at https://github.com/zeludeng/insclr.

Chengjian Feng, Yujie Zhong, Yu Gao et al.

One-stage object detection is commonly implemented by optimizing two sub-tasks: object classification and localization, using heads with two parallel branches, which might lead to a certain level of spatial misalignment in predictions between the two tasks. In this work, we propose a Task-aligned One-stage Object Detection (TOOD) that explicitly aligns the two tasks in a learning-based manner. First, we design a novel Task-aligned Head (T-Head) which offers a better balance between learning task-interactive and task-specific features, as well as a greater flexibility to learn the alignment via a task-aligned predictor. Second, we propose Task Alignment Learning (TAL) to explicitly pull closer (or even unify) the optimal anchors for the two tasks during training via a designed sample assignment scheme and a task-aligned loss. Extensive experiments are conducted on MS-COCO, where TOOD achieves a 51.1 AP at single-model single-scale testing. This surpasses the recent one-stage detectors by a large margin, such as ATSS (47.7 AP), GFL (48.2 AP), and PAA (49.0 AP), with fewer parameters and FLOPs. Qualitative results also demonstrate the effectiveness of TOOD for better aligning the tasks of object classification and localization. Code is available at https://github.com/fcjian/TOOD.

Chengjian Feng, Yujie Zhong, Weilin Huang

The conventional detectors tend to make imbalanced classification and suffer performance drop, when the distribution of the training data is severely skewed. In this paper, we propose to use the mean classification score to indicate the classification accuracy for each category during training. Based on this indicator, we balance the classification via an Equilibrium Loss (EBL) and a Memory-augmented Feature Sampling (MFS) method. Specifically, EBL increases the intensity of the adjustment of the decision boundary for the weak classes by a designed score-guided loss margin between any two classes. On the other hand, MFS improves the frequency and accuracy of the adjustment of the decision boundary for the weak classes through over-sampling the instance features of those classes. Therefore, EBL and MFS work collaboratively for finding the classification equilibrium in long-tailed detection, and dramatically improve the performance of tail classes while maintaining or even improving the performance of head classes. We conduct experiments on LVIS using Mask R-CNN with various backbones including ResNet-50-FPN and ResNet-101-FPN to show the superiority of the proposed method. It improves the detection performance of tail classes by 15.6 AP, and outperforms the most recent long-tailed object detectors by more than 1 AP. Code is available at https://github.com/fcjian/LOCE.

Guanting Liu, Yujie Zhong, Sheng Guo et al.

Differentiable architecture search (DAS) has made great progress in searching for high-performance architectures with reduced computational cost. However, DAS-based methods mainly focus on searching for a repeatable cell structure, which is then stacked sequentially in multiple stages to form the networks. This configuration significantly reduces the search space, and ignores the importance of connections between the cells. To overcome this limitation, in this paper, we propose a Hierarchical Differentiable Architecture Search (H-DAS) that performs architecture search both at the cell level and at the stage level. Specifically, the cell-level search space is relaxed so that the networks can learn stage-specific cell structures. For the stage-level search, we systematically study the architectures of stages, including the number of cells in each stage and the connections between the cells. Based on insightful observations, we design several search rules and losses, and mange to search for better stage-level architectures. Such hierarchical search space greatly improves the performance of the networks without introducing expensive search cost. Extensive experiments on CIFAR10 and ImageNet demonstrate the effectiveness of the proposed H-DAS. Moreover, the searched stage-level architectures can be combined with the cell structures searched by existing DAS methods to further boost the performance. Code is available at: https://github.com/MalongTech/research-HDAS

Yuechen Yu, Yilei Xiong, Weilin Huang et al.

Siamese-based trackers have achieved excellent performance on visual object tracking. However, the target template is not updated online, and the features of the target template and search image are computed independently in a Siamese architecture. In this paper, we propose Deformable Siamese Attention Networks, referred to as SiamAttn, by introducing a new Siamese attention mechanism that computes deformable self-attention and cross-attention. The self attention learns strong context information via spatial attention, and selectively emphasizes interdependent channel-wise features with channel attention. The cross-attention is capable of aggregating rich contextual inter-dependencies between the target template and the search image, providing an implicit manner to adaptively update the target template. In addition, we design a region refinement module that computes depth-wise cross correlations between the attentional features for more accurate tracking. We conduct experiments on six benchmarks, where our method achieves new state of-the-art results, outperforming the strong baseline, SiamRPN++ [24], by 0.464->0.537 and 0.415->0.470 EAO on VOT 2016 and 2018. Our code is available at: https://github.com/msight-tech/research-siamattn.

Xun Wang, Haozhi Zhang, Weilin Huang et al.

Mining informative negative instances are of central importance to deep metric learning (DML), however this task is intrinsically limited by mini-batch training, where only a mini-batch of instances is accessible at each iteration. In this paper, we identify a "slow drift" phenomena by observing that the embedding features drift exceptionally slow even as the model parameters are updating throughout the training process. This suggests that the features of instances computed at preceding iterations can be used to considerably approximate their features extracted by the current model. We propose a cross-batch memory (XBM) mechanism that memorizes the embeddings of past iterations, allowing the model to collect sufficient hard negative pairs across multiple mini-batches - even over the whole dataset. Our XBM can be directly integrated into a general pair-based DML framework, where the XBM augmented DML can boost performance considerably. In particular, without bells and whistles, a simple contrastive loss with our XBM can have large R@1 improvements of 12%-22.5% on three large-scale image retrieval datasets, surpassing the most sophisticated state-of-the-art methods, by a large margin. Our XBM is conceptually simple, easy to implement - using several lines of codes, and is memory efficient - with a negligible 0.2 GB extra GPU memory. Code is available at: https://github.com/MalongTech/research-xbm.

Linjie Xing, Zhi Tian, Weilin Huang et al.

Recent progress has been made on developing a unified framework for joint text detection and recognition in natural images, but existing joint models were mostly built on two-stage framework by involving ROI pooling, which can degrade the performance on recognition task. In this work, we propose convolutional character networks, referred as CharNet, which is an one-stage model that can process two tasks simultaneously in one pass. CharNet directly outputs bounding boxes of words and characters, with corresponding character labels. We utilize character as basic element, allowing us to overcome the main difficulty of existing approaches that attempted to optimize text detection jointly with a RNN-based recognition branch. In addition, we develop an iterative character detection approach able to transform the ability of character detection learned from synthetic data to real-world images. These technical improvements result in a simple, compact, yet powerful one-stage model that works reliably on multi-orientation and curved text. We evaluate CharNet on three standard benchmarks, where it consistently outperforms the state-of-the-art approaches [25, 24] by a large margin, e.g., with improvements of 65.33%->71.08% (with generic lexicon) on ICDAR 2015, and 54.0%->69.23% on Total-Text, on end-to-end text recognition. Code is available at: https://github.com/MalongTech/research-charnet.

Miao Kang, Xiaojun Hu, Weilin Huang et al.

We propose a Dual-Stream Pyramid Registration Network (referred as Dual-PRNet) for unsupervised 3D medical image registration. Unlike recent CNN-based registration approaches, such as VoxelMorph, which explores a single-stream encoder-decoder network to compute a registration fields from a pair of 3D volumes, we design a two-stream architecture able to compute multi-scale registration fields from convolutional feature pyramids. Our contributions are two-fold: (i) we design a two-stream 3D encoder-decoder network which computes two convolutional feature pyramids separately for a pair of input volumes, resulting in strong deep representations that are meaningful for deformation estimation; (ii) we propose a pyramid registration module able to predict multi-scale registration fields directly from the decoding feature pyramids. This allows it to refine the registration fields gradually in a coarse-to-fine manner via sequential warping, and enable the model with the capability for handling significant deformations between two volumes, such as large displacements in spatial domain or slice space. The proposed Dual-PRNet is evaluated on two standard benchmarks for brain MRI registration, where it outperforms the state-of-the-art approaches by a large margin, e.g., having improvements over recent VoxelMorph [2] with 0.683->0.778 on the LPBA40, and 0.511->0.631 on the Mindboggle101, in term of average Dice score. Code is available at: https://github.com/kangmiao15/Dual-Stream-PRNet-Plus.

Sheng Guo, Weilin Huang, Xiao Zhang et al.

Large-scale image databases such as ImageNet have significantly advanced image classification and other visual recognition tasks. However much of these datasets are constructed only for single-label and coarse object-level classification. For real-world applications, multiple labels and fine-grained categories are often needed, yet very few such datasets exist publicly, especially those of large-scale and high quality. In this work, we contribute to the community a new dataset called iMaterialist Fashion Attribute (iFashion-Attribute) to address this problem in the fashion domain. The dataset is constructed from over one million fashion images with a label space that includes 8 groups of 228 fine-grained attributes in total. Each image is annotated by experts with multiple, high-quality fashion attributes. The result is the first known million-scale multi-label and fine-grained image dataset. We conduct extensive experiments and provide baseline results with modern deep Convolutional Neural Networks (CNNs). Additionally, we demonstrate models pre-trained on iFashion-Attribute achieve superior transfer learning performance on fashion related tasks compared with pre-training from ImageNet or other fashion datasets. Data is available at: https://github.com/visipedia/imat_fashion_comp

Xun Wang, Xintong Han, Weilin Huang et al.

A family of loss functions built on pair-based computation have been proposed in the literature which provide a myriad of solutions for deep metric learning. In this paper, we provide a general weighting framework for understanding recent pair-based loss functions. Our contributions are three-fold: (1) we establish a General Pair Weighting (GPW) framework, which casts the sampling problem of deep metric learning into a unified view of pair weighting through gradient analysis, providing a powerful tool for understanding recent pair-based loss functions; (2) we show that with GPW, various existing pair-based methods can be compared and discussed comprehensively, with clear differences and key limitations identified; (3) we propose a new loss called multi-similarity loss (MS loss) under the GPW, which is implemented in two iterative steps (i.e., mining and weighting). This allows it to fully consider three similarities for pair weighting, providing a more principled approach for collecting and weighting informative pairs. Finally, the proposed MS loss obtains new state-of-the-art performance on four image retrieval benchmarks, where it outperforms the most recent approaches, such as ABE\cite{Kim_2018_ECCV} and HTL by a large margin: 60.6% to 65.7% on CUB200, and 80.9% to 88.0% on In-Shop Clothes Retrieval dataset at Recall@1. Code is available at https://github.com/MalongTech/research-ms-loss.

Sheng Guo, Weilin Huang, Haozhi Zhang et al.

We present a simple yet efficient approach capable of training deep neural networks on large-scale weakly-supervised web images, which are crawled raw from the Internet by using text queries, without any human annotation. We develop a principled learning strategy by leveraging curriculum learning, with the goal of handling a massive amount of noisy labels and data imbalance effectively. We design a new learning curriculum by measuring the complexity of data using its distribution density in a feature space, and rank the complexity in an unsupervised manner. This allows for an efficient implementation of curriculum learning on large-scale web images, resulting in a high-performance CNN model, where the negative impact of noisy labels is reduced substantially. Importantly, we show by experiments that those images with highly noisy labels can surprisingly improve the generalization capability of the model, by serving as a manner of regularization. Our approaches obtain state-of-the-art performance on four benchmarks: WebVision, ImageNet, Clothing-1M and Food-101. With an ensemble of multiple models, we achieved a top-5 error rate of 5.2% on the WebVision challenge for 1000-category classification. This result was the top performance by a wide margin, outperforming second place by a nearly 50% relative error rate. Code and models are available at: https://github.com/MalongTech/CurriculumNet .

Limin Wang, Sheng Guo, Weilin Huang et al.

Convolutional Neural Networks (CNNs) have made remarkable progress on scene recognition, partially due to these recent large-scale scene datasets, such as the Places and Places2. Scene categories are often defined by multi-level information, including local objects, global layout, and background environment, thus leading to large intra-class variations. In addition, with the increasing number of scene categories, label ambiguity has become another crucial issue in large-scale classification. This paper focuses on large-scale scene recognition and makes two major contributions to tackle these issues. First, we propose a multi-resolution CNN architecture that captures visual content and structure at multiple levels. The multi-resolution CNNs are composed of coarse resolution CNNs and fine resolution CNNs, which are complementary to each other. Second, we design two knowledge guided disambiguation techniques to deal with the problem of label ambiguity. (i) We exploit the knowledge from the confusion matrix computed on validation data to merge ambiguous classes into a super category. (ii) We utilize the knowledge of extra networks to produce a soft label for each image. Then the super categories or soft labels are employed to guide CNN training on the Places2. We conduct extensive experiments on three large-scale image datasets (ImageNet, Places, and Places2), demonstrating the effectiveness of our approach. Furthermore, our method takes part in two major scene recognition challenges, and achieves the second place at the Places2 challenge in ILSVRC 2015, and the first place at the LSUN challenge in CVPR 2016. Finally, we directly test the learned representations on other scene benchmarks, and obtain the new state-of-the-art results on the MIT Indoor67 (86.7\%) and SUN397 (72.0\%). We release the code and models at~\url{https://github.com/wanglimin/MRCNN-Scene-Recognition}.

Zeyue Xue, Jie Wu, Yu Gao et al.

Recent advances in generative AI have revolutionized visual content creation, yet aligning model outputs with human preferences remains a critical challenge. While Reinforcement Learning (RL) has emerged as a promising approach for fine-tuning generative models, existing methods like DDPO and DPOK face fundamental limitations - particularly their inability to maintain stable optimization when scaling to large and diverse prompt sets, severely restricting their practical utility. This paper presents DanceGRPO, a framework that addresses these limitations through an innovative adaptation of Group Relative Policy Optimization (GRPO) for visual generation tasks. Our key insight is that GRPO's inherent stability mechanisms uniquely position it to overcome the optimization challenges that plague prior RL-based approaches on visual generation. DanceGRPO establishes several significant advances: First, it demonstrates consistent and stable policy optimization across multiple modern generative paradigms, including both diffusion models and rectified flows. Second, it maintains robust performance when scaling to complex, real-world scenarios encompassing three key tasks and four foundation models. Third, it shows remarkable versatility in optimizing for diverse human preferences as captured by five distinct reward models assessing image/video aesthetics, text-image alignment, video motion quality, and binary feedback. Our comprehensive experiments reveal that DanceGRPO outperforms baseline methods by up to 181\% across multiple established benchmarks, including HPS-v2.1, CLIP Score, VideoAlign, and GenEval. Our results establish DanceGRPO as a robust and versatile solution for scaling Reinforcement Learning from Human Feedback (RLHF) tasks in visual generation, offering new insights into harmonizing reinforcement learning and visual synthesis.

Yu Gao, Lixue Gong, Qiushan Guo et al.

We present Seedream 3.0, a high-performance Chinese-English bilingual image generation foundation model. We develop several technical improvements to address existing challenges in Seedream 2.0, including alignment with complicated prompts, fine-grained typography generation, suboptimal visual aesthetics and fidelity, and limited image resolutions. Specifically, the advancements of Seedream 3.0 stem from improvements across the entire pipeline, from data construction to model deployment. At the data stratum, we double the dataset using a defect-aware training paradigm and a dual-axis collaborative data-sampling framework. Furthermore, we adopt several effective techniques such as mixed-resolution training, cross-modality RoPE, representation alignment loss, and resolution-aware timestep sampling in the pre-training phase. During the post-training stage, we utilize diversified aesthetic captions in SFT, and a VLM-based reward model with scaling, thereby achieving outputs that well align with human preferences. Furthermore, Seedream 3.0 pioneers a novel acceleration paradigm. By employing consistent noise expectation and importance-aware timestep sampling, we achieve a 4 to 8 times speedup while maintaining image quality. Seedream 3.0 demonstrates significant improvements over Seedream 2.0: it enhances overall capabilities, in particular for text-rendering in complicated Chinese characters which is important to professional typography generation. In addition, it provides native high-resolution output (up to 2K), allowing it to generate images with high visual quality.

Yu Gao, Haoyuan Guo, Tuyen Hoang et al.

Notable breakthroughs in diffusion modeling have propelled rapid improvements in video generation, yet current foundational model still face critical challenges in simultaneously balancing prompt following, motion plausibility, and visual quality. In this report, we introduce Seedance 1.0, a high-performance and inference-efficient video foundation generation model that integrates several core technical improvements: (i) multi-source data curation augmented with precision and meaningful video captioning, enabling comprehensive learning across diverse scenarios; (ii) an efficient architecture design with proposed training paradigm, which allows for natively supporting multi-shot generation and jointly learning of both text-to-video and image-to-video tasks. (iii) carefully-optimized post-training approaches leveraging fine-grained supervised fine-tuning, and video-specific RLHF with multi-dimensional reward mechanisms for comprehensive performance improvements; (iv) excellent model acceleration achieving ~10x inference speedup through multi-stage distillation strategies and system-level optimizations. Seedance 1.0 can generate a 5-second video at 1080p resolution only with 41.4 seconds (NVIDIA-L20). Compared to state-of-the-art video generation models, Seedance 1.0 stands out with high-quality and fast video generation having superior spatiotemporal fluidity with structural stability, precise instruction adherence in complex multi-subject contexts, native multi-shot narrative coherence with consistent subject representation.

Chao Liao, Liyang Liu, Xun Wang et al.

Recent progress in unified models for image understanding and generation has been impressive, yet most approaches remain limited to single-modal generation conditioned on multiple modalities. In this paper, we present Mogao, a unified framework that advances this paradigm by enabling interleaved multi-modal generation through a causal approach. Mogao integrates a set of key technical improvements in architecture design, including a deep-fusion design, dual vision encoders, interleaved rotary position embeddings, and multi-modal classifier-free guidance, which allow it to harness the strengths of both autoregressive models for text generation and diffusion models for high-quality image synthesis. These practical improvements also make Mogao particularly effective to process interleaved sequences of text and images arbitrarily. To further unlock the potential of unified models, we introduce an efficient training strategy on a large-scale, in-house dataset specifically curated for joint text and image generation. Extensive experiments show that Mogao not only achieves state-of-the-art performance in multi-modal understanding and text-to-image generation, but also excels in producing high-quality, coherent interleaved outputs. Its emergent capabilities in zero-shot image editing and compositional generation highlight Mogao as a practical omni-modal foundation model, paving the way for future development and scaling the unified multi-modal systems.

Yichun Shi, Peng Wang, Weilin Huang

We introduce SeedEdit, a diffusion model that is able to revise a given image with any text prompt. In our perspective, the key to such a task is to obtain an optimal balance between maintaining the original image, i.e. image reconstruction, and generating a new image, i.e. image re-generation. To this end, we start from a weak generator (text-to-image model) that creates diverse pairs between such two directions and gradually align it into a strong image editor that well balances between the two tasks. SeedEdit can achieve more diverse and stable editing capability over prior image editing methods, enabling sequential revision over images generated by diffusion models.

Hanzhong Guo, Jie Wu, Jie Liu et al.

While Reinforcement Learning from Human Feedback (RLHF) has become a pivotal paradigm for text-to-image generation, its application to image editing remains largely unexplored. A key bottleneck is the lack of a robust general reward model for all editing tasks. Existing edit reward models usually give overall scores without detailed checks, ignoring different instruction requirements and causing biased rewards. To address this, we argue that the key is to move from a simple scorer to a reasoning verifier. We introduce Edit-R1, a framework that builds a chain-of-thought (CoT) verifier-based reasoning reward model (RRM) and then leverages it for downstream image editing. The Edit-RRM breaks instructions into distinct principles, evaluates the edited image against each principle, and aggregates these checks into an interpretable, fine-grained reward. To build such an RRM, we first apply supervised fine-tuning (SFT) as a ``cold-start'' to generate CoT reward trajectories. Then, we introduce Group Contrastive Preference Optimization (GCPO), a reinforcement learning algorithm that leverages human pairwise preference data to reinforce our pointwise RRM. After building the RRM, we use GRPO to train editing models with this non-differentiable yet powerful reward model. Extensive experiments demonstrate that our Edit-RRM surpasses powerful VLMs such as Seed-1.5-VL and Seed-1.6-VL as an editing-specific reward model, and we observe a clear scaling trend, with performance consistently improving from 3B to 7B parameters. Moreover, Edit-R1 delivers gains to editing models like FLUX.1-kontext, highlighting its effectiveness in enhancing image editing.

Lixue Gong, Xiaoxia Hou, Fanshi Li et al.

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.

Ceyuan Yang, Zhijie Lin, Yang Zhao et al.

We present Omni, a unified multimodal model natively trained on diverse modalities, including text, images, videos, 3D geometry, and hidden representations. We find that such training enables Context Unrolling, where the model explicitly reasons across multiple modal representations before producing predictions. This process enables the model to aggregate complementary information across heterogeneous modalities, facilitating a more faithful approximation of the shared multimodal knowledge manifold and improving downstream reasoning fidelity. As a result, Omni achieves strong performance on both multimodal generation and understanding benchmarks, while demonstrating advanced multimodal reasoning capabilities, including in-context generation of text, image, video, and 3D geometry.

Team Seedream, Yunpeng Chen, Yu Gao et al.

We introduce Seedream 4.0, an efficient and high-performance multimodal image generation system that unifies text-to-image (T2I) synthesis, image editing, and multi-image composition within a single framework. We develop a highly efficient diffusion transformer with a powerful VAE which also can reduce the number of image tokens considerably. This allows for efficient training of our model, and enables it to fast generate native high-resolution images (e.g., 1K-4K). Seedream 4.0 is pretrained on billions of text-image pairs spanning diverse taxonomies and knowledge-centric concepts. Comprehensive data collection across hundreds of vertical scenarios, coupled with optimized strategies, ensures stable and large-scale training, with strong generalization. By incorporating a carefully fine-tuned VLM model, we perform multi-modal post-training for training both T2I and image editing tasks jointly. For inference acceleration, we integrate adversarial distillation, distribution matching, and quantization, as well as speculative decoding. It achieves an inference time of up to 1.8 seconds for generating a 2K image (without a LLM/VLM as PE model). Comprehensive evaluations reveal that Seedream 4.0 can achieve state-of-the-art results on both T2I and multimodal image editing. In particular, it demonstrates exceptional multimodal capabilities in complex tasks, including precise image editing and in-context reasoning, and also allows for multi-image reference, and can generate multiple output images. This extends traditional T2I systems into an more interactive and multidimensional creative tool, pushing the boundary of generative AI for both creativity and professional applications. Seedream 4.0 is now accessible on https://www.volcengine.com/experience/ark?launch=seedream.

Peng Wang, Yichun Shi, Xiaochen Lian et al.

We introduce SeedEdit 3.0, in companion with our T2I model Seedream 3.0, which significantly improves over our previous SeedEdit versions 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 reward losses. Finally, we evaluate SeedEdit 3.0 on our testing benchmarks, for real/synthetic 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%), GPT4o (37.1%) and Gemini 2.0 (30.3%).

Shuai Wang, Zhi Tian, Weilin Huang et al.

Diffusion transformers have demonstrated remarkable generation quality, albeit requiring longer training iterations and numerous inference steps. In each denoising step, diffusion transformers encode the noisy inputs to extract the lower-frequency semantic component and then decode the higher frequency with identical modules. This scheme creates an inherent optimization dilemma: encoding low-frequency semantics necessitates reducing high-frequency components, creating tension between semantic encoding and high-frequency decoding. To resolve this challenge, we propose a new \textbf{\color{ddt}D}ecoupled \textbf{\color{ddt}D}iffusion \textbf{\color{ddt}T}ransformer~(\textbf{\color{ddt}DDT}), with a decoupled design of a dedicated condition encoder for semantic extraction alongside a specialized velocity decoder. Our experiments reveal that a more substantial encoder yields performance improvements as model size increases. For ImageNet $256\times256$, Our DDT-XL/2 achieves a new state-of-the-art performance of {1.31 FID}~(nearly $4\times$ faster training convergence compared to previous diffusion transformers). For ImageNet $512\times512$, Our DDT-XL/2 achieves a new state-of-the-art FID of 1.28. Additionally, as a beneficial by-product, our decoupled architecture enhances inference speed by enabling the sharing self-condition between adjacent denoising steps. To minimize performance degradation, we propose a novel statistical dynamic programming approach to identify optimal sharing strategies.

Chen Ju, Haicheng Wang, Zeqian Li et al.

Vision-Language Large Models (VLMs) have become primary backbone of AI, due to the impressive performance. However, their expensive computation costs, i.e., throughput and delay, impede potentials in real-world scenarios. To achieve acceleration for VLMs, most existing methods focus on the model perspective: pruning, distillation, quantification, but completely overlook the data-perspective redundancy. To fill the overlook, this paper pioneers the severity of data redundancy, and designs one plug-and-play Turbo module guided by information degree to prune inefficient tokens from visual or textual data. In pursuit of efficiency-performance trade-offs, information degree takes two key factors into consideration: mutual redundancy and semantic value. Concretely, the former evaluates the data duplication between sequential tokens; while the latter evaluates each token by its contribution to the overall semantics. As a result, tokens with high information degree carry less redundancy and stronger semantics. For VLMs' calculation, Turbo works as a user-friendly plug-in that sorts data referring to information degree, utilizing only top-level ones to save costs. Its advantages are multifaceted, e.g., being generally compatible to various VLMs across understanding and generation, simple use without retraining and trivial engineering efforts. On multiple public VLMs benchmarks, we conduct extensive experiments to reveal the gratifying acceleration of Turbo, under negligible performance drop.

Jie Wu, Yu Gao, Zilyu Ye et al.

Reward Models (RMs) are critical for improving generation models via Reinforcement Learning (RL), yet the RM scaling paradigm in visual generation remains largely unexplored. It primarily due to fundamental limitations in existing approaches: CLIP-based RMs suffer from architectural and input modality constraints, while prevalent Bradley-Terry losses are fundamentally misaligned with the next-token prediction mechanism of Vision-Language Models (VLMs), hindering effective scaling. More critically, the RLHF optimization process is plagued by Reward Hacking issue, where models exploit flaws in the reward signal without improving true quality. To address these challenges, we introduce RewardDance, a scalable reward modeling framework that overcomes these barriers through a novel generative reward paradigm. By reformulating the reward score as the model's probability of predicting a "yes" token, indicating that the generated image outperforms a reference image according to specific criteria, RewardDance intrinsically aligns reward objectives with VLM architectures. This alignment unlocks scaling across two dimensions: (1) Model Scaling: Systematic scaling of RMs up to 26 billion parameters; (2) Context Scaling: Integration of task-specific instructions, reference examples, and chain-of-thought (CoT) reasoning. Extensive experiments demonstrate that RewardDance significantly surpasses state-of-the-art methods in text-to-image, text-to-video, and image-to-video generation. Crucially, we resolve the persistent challenge of "reward hacking": Our large-scale RMs exhibit and maintain high reward variance during RL fine-tuning, proving their resistance to hacking and ability to produce diverse, high-quality outputs. It greatly relieves the mode collapse problem that plagues smaller models.

Yatai Ji, Jiacheng Zhang, Jie Wu et al.

Text-to-video models have made remarkable advancements through optimization on high-quality text-video pairs, where the textual prompts play a pivotal role in determining quality of output videos. However, achieving the desired output often entails multiple revisions and iterative inference to refine user-provided prompts. Current automatic methods for refining prompts encounter challenges such as Modality-Inconsistency, Cost-Discrepancy, and Model-Unaware when applied to text-to-video diffusion models. To address these problem, we introduce an LLM-based prompt adaptation framework, termed as Prompt-A-Video, which excels in crafting Video-Centric, Labor-Free and Preference-Aligned prompts tailored to specific video diffusion model. Our approach involves a meticulously crafted two-stage optimization and alignment system. Initially, we conduct a reward-guided prompt evolution pipeline to automatically create optimal prompts pool and leverage them for supervised fine-tuning (SFT) of the LLM. Then multi-dimensional rewards are employed to generate pairwise data for the SFT model, followed by the direct preference optimization (DPO) algorithm to further facilitate preference alignment. Through extensive experimentation and comparative analyses, we validate the effectiveness of Prompt-A-Video across diverse generation models, highlighting its potential to push the boundaries of video generation.

Jiacheng Zhang, Jie Wu, Yuxi Ren et al.

Latent diffusion models (LDM) have revolutionized text-to-image generation, leading to the proliferation of various advanced models and diverse downstream applications. However, despite these significant advancements, current diffusion models still suffer from several limitations, including inferior visual quality, inadequate aesthetic appeal, and inefficient inference, without a comprehensive solution in sight. To address these challenges, we present UniFL, a unified framework that leverages feedback learning to enhance diffusion models comprehensively. UniFL stands out as a universal, effective, and generalizable solution applicable to various diffusion models, such as SD1.5 and SDXL. Notably, UniFL consists of three key components: perceptual feedback learning, which enhances visual quality; decoupled feedback learning, which improves aesthetic appeal; and adversarial feedback learning, which accelerates inference. In-depth experiments and extensive user studies validate the superior performance of our method in enhancing generation quality and inference acceleration. For instance, UniFL surpasses ImageReward by 17% user preference in terms of generation quality and outperforms LCM and SDXL Turbo by 57% and 20% general preference with 4-step inference.

Shuai Wang, Ziteng Gao, Chenhui Zhu et al.

The current success of diffusion transformers heavily depends on the compressed latent space shaped by the pre-trained variational autoencoder(VAE). However, this two-stage training paradigm inevitably introduces accumulated errors and decoding artifacts. To address the aforementioned problems, researchers return to pixel space at the cost of complicated cascade pipelines and increased token complexity. In contrast to their efforts, we propose to model the patch-wise decoding with neural field and present a single-scale, single-stage, efficient, end-to-end solution, coined as pixel neural field diffusion~(PixelNerd). Thanks to the efficient neural field representation in PixNerd, we directly achieved 2.15 FID on ImageNet $256\times256$ and 2.84 FID on ImageNet $512\times512$ without any complex cascade pipeline or VAE. We also extend our PixNerd framework to text-to-image applications. Our PixNerd-XXL/16 achieved a competitive 0.73 overall score on the GenEval benchmark and 80.9 overall score on the DPG benchmark.

Zeqian Li, Shangzhe Di, Zhonghua Zhai et al.

This paper presents a computational model for universal video temporal grounding, which accurately localizes temporal moments in videos based on natural language queries (e.g., questions or descriptions). Unlike existing methods that are often limited to specific video domains or durations, we propose UniTime, a robust and universal video grounding model leveraging the strong vision-language understanding capabilities of generative Multi-modal Large Language Models (MLLMs). Our model effectively handles videos of diverse views, genres, and lengths while comprehending complex language queries. The key contributions include: (i) We consider steering strong MLLMs for temporal grounding in videos. To enable precise timestamp outputs, we incorporate temporal information by interleaving timestamp tokens with video tokens. (ii) By training the model to handle videos with different input granularities through adaptive frame scaling, our approach achieves robust temporal grounding for both short and long videos. (iii) Comprehensive experiments show that UniTime outperforms state-of-the-art approaches in both zero-shot and dataset-specific finetuned settings across five public temporal grounding benchmarks. (iv) When employed as a preliminary moment retriever for long-form video question-answering (VideoQA), UniTime significantly improves VideoQA accuracy, highlighting its value for complex video understanding tasks.

Di Chang, Mingdeng Cao, Yichun Shi et al. · stanford

Editing images with instructions to reflect non-rigid motions, camera viewpoint shifts, object deformations, human articulations, and complex interactions, poses a challenging yet underexplored problem in computer vision. Existing approaches and datasets predominantly focus on static scenes or rigid transformations, limiting their capacity to handle expressive edits involving dynamic motion. To address this gap, we introduce ByteMorph, a comprehensive framework for instruction-based image editing with an emphasis on non-rigid motions. ByteMorph comprises a large-scale dataset, ByteMorph-6M, and a strong baseline model built upon the Diffusion Transformer (DiT), named ByteMorpher. ByteMorph-6M includes over 6 million high-resolution image editing pairs for training, along with a carefully curated evaluation benchmark ByteMorph-Bench. Both capture a wide variety of non-rigid motion types across diverse environments, human figures, and object categories. The dataset is constructed using motion-guided data generation, layered compositing techniques, and automated captioning to ensure diversity, realism, and semantic coherence. We further conduct a comprehensive evaluation of recent instruction-based image editing methods from both academic and commercial domains.

Chenyu Zheng, Xinyu Zhang, Rongzhen Wang et al.

Diffusion Transformers have emerged as the foundation for vision generative models, but their scalability is limited by the high cost of hyperparameter (HP) tuning at large scales. Recently, Maximal Update Parametrization ($μ$P) was proposed for vanilla Transformers, which enables stable HP transfer from small to large language models, and dramatically reduces tuning costs. However, it remains unclear whether $μ$P of vanilla Transformers extends to diffusion Transformers, which differ architecturally and objectively. In this work, we generalize standard $μ$P to diffusion Transformers and validate its effectiveness through large-scale experiments. First, we rigorously prove that $μ$P of mainstream diffusion Transformers, including U-ViT, DiT, PixArt-$α$, and MMDiT, aligns with that of the vanilla Transformer, enabling the direct application of existing $μ$P methodologies. Leveraging this result, we systematically demonstrate that DiT-$μ$P enjoys robust HP transferability. Notably, DiT-XL-2-$μ$P with transferred learning rate achieves 2.9 times faster convergence than the original DiT-XL-2. Finally, we validate the effectiveness of $μ$P on text-to-image generation by scaling PixArt-$α$ from 0.04B to 0.61B and MMDiT from 0.18B to 18B. In both cases, models under $μ$P outperform their respective baselines while requiring small tuning cost, only 5.5% of one training run for PixArt-$α$ and 3% of consumption by human experts for MMDiT-18B. These results establish $μ$P as a principled and efficient framework for scaling diffusion Transformers.

Yuxin Yang, Haoran Zhang, Mingxuan Li et al.

Parameter-Efficient Fine-Tuning (PEFT) techniques, particularly Low-Rank Adaptation (LoRA), have become essential for adapting Large Language Models (LLMs) to downstream tasks. While the recent FlyLoRA framework successfully leverages bio-inspired sparse random projections to mitigate parameter interference, it relies on a static, magnitude-based routing mechanism that is agnostic to input context. In this paper, we propose NeuroLoRA, a novel Mixture-of-Experts (MoE) based LoRA framework inspired by biological neuromodulation -- the dynamic regulation of neuronal excitability based on context. NeuroLoRA retains the computational efficiency of frozen random projections while introducing a lightweight, learnable neuromodulation gate that contextually rescales the projection space prior to expert selection. We further propose a Contrastive Orthogonality Loss to explicitly enforce separation between expert subspaces, enhancing both task decoupling and continual learning capacity. Extensive experiments on MMLU, GSM8K, and ScienceQA demonstrate that NeuroLoRA consistently outperforms FlyLoRA and other strong baselines across single-task adaptation, multi-task model merging, and sequential continual learning scenarios, while maintaining comparable parameter efficiency.

Team Seedance, Heyi Chen, Siyan Chen et al.

Recent strides in video generation have paved the way for unified audio-visual generation. In this work, we present Seedance 1.5 pro, a foundational model engineered specifically for native, joint audio-video generation. Leveraging a dual-branch Diffusion Transformer architecture, the model integrates a cross-modal joint module with a specialized multi-stage data pipeline, achieving exceptional audio-visual synchronization and superior generation quality. To ensure practical utility, we implement meticulous post-training optimizations, including Supervised Fine-Tuning (SFT) on high-quality datasets and Reinforcement Learning from Human Feedback (RLHF) with multi-dimensional reward models. Furthermore, we introduce an acceleration framework that boosts inference speed by over 10X. Seedance 1.5 pro distinguishes itself through precise multilingual and dialect lip-syncing, dynamic cinematic camera control, and enhanced narrative coherence, positioning it as a robust engine for professional-grade content creation. Seedance 1.5 pro is now accessible on Volcano Engine at https://console.volcengine.com/ark/region:ark+cn-beijing/experience/vision?type=GenVideo.

Fengyuan Shi, Weilin Huang, Limin Wang

Video grounding aims to localize the corresponding video moment in an untrimmed video given a language query. Existing methods often address this task in an indirect way, by casting it as a proposal-and-match or fusion-and-detection problem. Solving these surrogate problems often requires sophisticated label assignment during training and hand-crafted removal of near-duplicate results. Meanwhile, existing works typically focus on sparse video grounding with a single sentence as input, which could result in ambiguous localization due to its unclear description. In this paper, we tackle a new problem of dense video grounding, by simultaneously localizing multiple moments with a paragraph as input. From a perspective on video grounding as language conditioned regression, we present an end-to-end parallel decoding paradigm by re-purposing a Transformer-alike architecture (PRVG). The key design in our PRVG is to use languages as queries, and directly regress the moment boundaries based on language-modulated visual representations. Thanks to its simplicity in design, our PRVG framework can be applied in different testing schemes (sparse or dense grounding) and allows for efficient inference without any post-processing technique. In addition, we devise a robust proposal-level attention loss to guide the training of PRVG, which is invariant to moment duration and contributes to model convergence. We perform experiments on two video grounding benchmarks of ActivityNet Captions and TACoS, demonstrating that our PRVG can significantly outperform previous methods. We also perform in-depth studies to investigate the effectiveness of parallel regression paradigm on video grounding.

Haoxian Tan, Sheng Guo, Yujie Zhong et al.

Differentiable architecture search is prevalent in the field of NAS because of its simplicity and efficiency, where two paradigms, multi-path algorithms and single-path methods, are dominated. Multi-path framework (e.g. DARTS) is intuitive but suffers from memory usage and training collapse. Single-path methods (e.g.GDAS and ProxylessNAS) mitigate the memory issue and shrink the gap between searching and evaluation but sacrifice the performance. In this paper, we propose a conceptually simple yet efficient method to bridge these two paradigms, referred as Mutually-aware Sub-Graphs Differentiable Architecture Search (MSG-DAS). The core of our framework is a differentiable Gumbel-TopK sampler that produces multiple mutually exclusive single-path sub-graphs. To alleviate the severer skip-connect issue brought by multiple sub-graphs setting, we propose a Dropblock-Identity module to stabilize the optimization. To make best use of the available models (super-net and sub-graphs), we introduce a memory-efficient super-net guidance distillation to improve training. The proposed framework strikes a balance between flexible memory usage and searching quality. We demonstrate the effectiveness of our methods on ImageNet and CIFAR10, where the searched models show a comparable performance as the most recent approaches.

Haozhi Zhang, Xun Wang, Weilin Huang et al.

Pair-wise loss functions have been extensively studied and shown to continuously improve the performance of deep metric learning (DML). However, they are primarily designed with intuition based on simple toy examples, and experimentally identifying the truly effective design is difficult in complicated, real-world cases. In this paper, we provide a new methodology for systematically studying weighting strategies of various pair-wise loss functions, and rethink pair weighting with an embedding memory. We delve into the weighting mechanisms by decomposing the pair-wise functions, and study positive and negative weights separately using direct weight assignment. This allows us to study various weighting functions deeply and systematically via weight curves, and identify a number of meaningful, comprehensive and insightful facts, which come up with our key observation on memory-based DML: it is critical to mine hard negatives and discard easy negatives which are less informative and redundant, but weighting on positive pairs is not helpful. This results in an efficient but surprisingly simple rule to design the weighting scheme, making it significantly different from existing mini-batch based methods which design various sophisticated loss functions to weight pairs carefully. Finally, we conduct extensive experiments on three large-scale visual retrieval benchmarks, and demonstrate the superiority of memory-based DML over recent mini-batch based approaches, by using a simple contrastive loss with momentum-updated memory.

Yawen Huang, Feng Zheng, Danyang Wang et al.

Recent advances in neuroscience have highlighted the effectiveness of multi-modal medical data for investigating certain pathologies and understanding human cognition. However, obtaining full sets of different modalities is limited by various factors, such as long acquisition times, high examination costs and artifact suppression. In addition, the complexity, high dimensionality and heterogeneity of neuroimaging data remains another key challenge in leveraging existing randomized scans effectively, as data of the same modality is often measured differently by different machines. There is a clear need to go beyond the traditional imaging-dependent process and synthesize anatomically specific target-modality data from a source input. In this paper, we propose to learn dedicated features that cross both intre- and intra-modal variations using a novel CSC$\ell_4$Net. Through an initial unification of intra-modal data in the feature maps and multivariate canonical adaptation, CSC$\ell_4$Net facilitates feature-level mutual transformation. The positive definite Riemannian manifold-penalized data fidelity term further enables CSC$\ell_4$Net to reconstruct missing measurements according to transformed features. Finally, the maximization $\ell_4$-norm boils down to a computationally efficient optimization problem. Extensive experiments validate the ability and robustness of our CSC$\ell_4$Net compared to the state-of-the-art methods on multiple datasets.

Yujie Zhong, Zelu Deng, Sheng Guo et al.

Region Proposal Network (RPN) provides strong support for handling the scale variation of objects in two-stage object detection. For one-stage detectors which do not have RPN, it is more demanding to have powerful sub-networks capable of directly capturing objects of unknown sizes. To enhance such capability, we propose an extremely efficient neural architecture search method, named Fast And Diverse (FAD), to better explore the optimal configuration of receptive fields and convolution types in the sub-networks for one-stage detectors. FAD consists of a designed search space and an efficient architecture search algorithm. The search space contains a rich set of diverse transformations designed specifically for object detection. To cope with the designed search space, a novel search algorithm termed Representation Sharing (RepShare) is proposed to effectively identify the best combinations of the defined transformations. In our experiments, FAD obtains prominent improvements on two types of one-stage detectors with various backbones. In particular, our FAD detector achieves 46.4 AP on MS-COCO (under single-scale testing), outperforming the state-of-the-art detectors, including the most recent NAS-based detectors, Auto-FPN (searched for 16 GPU-days) and NAS-FCOS (28 GPU-days), while significantly reduces the search cost to 0.6 GPU-days. Beyond object detection, we further demonstrate the generality of FAD on the more challenging instance segmentation, and expect it to benefit more tasks.

Yu Gao, Xintong Han, Xun Wang et al.

Fine-grained image categorization is challenging due to the subtle inter-class differences.We posit that exploiting the rich relationships between channels can help capture such differences since different channels correspond to different semantics. In this paper, we propose a channel interaction network (CIN), which models the channel-wise interplay both within an image and across images. For a single image, a self-channel interaction (SCI) module is proposed to explore channel-wise correlation within the image. This allows the model to learn the complementary features from the correlated channels, yielding stronger fine-grained features. Furthermore, given an image pair, we introduce a contrastive channel interaction (CCI) module to model the cross-sample channel interaction with a metric learning framework, allowing the CIN to distinguish the subtle visual differences between images. Our model can be trained efficiently in an end-to-end fashion without the need of multi-stage training and testing. Finally, comprehensive experiments are conducted on three publicly available benchmarks, where the proposed method consistently outperforms the state-of-theart approaches, such as DFL-CNN (Wang, Morariu, and Davis 2018) and NTS (Yang et al. 2018).