Songhua Liu

Songhua Liu, Kai Wang, Xingyi Yang et al.

In this paper, we study \xw{dataset distillation (DD)}, from a novel perspective and introduce a \emph{dataset factorization} approach, termed \emph{HaBa}, which is a plug-and-play strategy portable to any existing DD baseline. Unlike conventional DD approaches that aim to produce distilled and representative samples, \emph{HaBa} explores decomposing a dataset into two components: data \emph{Ha}llucination networks and \emph{Ba}ses, where the latter is fed into the former to reconstruct image samples. The flexible combinations between bases and hallucination networks, therefore, equip the distilled data with exponential informativeness gain, which largely increase the representation capability of distilled datasets. To furthermore increase the data efficiency of compression results, we further introduce a pair of adversarial contrastive constraints on the resultant hallucination networks and bases, which increase the diversity of generated images and inject more discriminant information into the factorization. Extensive comparisons and experiments demonstrate that our method can yield significant improvement on downstream classification tasks compared with previous state of the arts, while reducing the total number of compressed parameters by up to 65\%. Moreover, distilled datasets by our approach also achieve \textasciitilde10\% higher accuracy than baseline methods in cross-architecture generalization. Our code is available \href{https://github.com/Huage001/DatasetFactorization}{here}.

Xingyi Yang, Daquan Zhou, Songhua Liu et al.

In this paper, we explore a novel knowledge-transfer task, termed as Deep Model Reassembly (DeRy), for general-purpose model reuse. Given a collection of heterogeneous models pre-trained from distinct sources and with diverse architectures, the goal of DeRy, as its name implies, is to first dissect each model into distinctive building blocks, and then selectively reassemble the derived blocks to produce customized networks under both the hardware resource and performance constraints. Such ambitious nature of DeRy inevitably imposes significant challenges, including, in the first place, the feasibility of its solution. We strive to showcase that, through a dedicated paradigm proposed in this paper, DeRy can be made not only possibly but practically efficiently. Specifically, we conduct the partitions of all pre-trained networks jointly via a cover set optimization, and derive a number of equivalence set, within each of which the network blocks are treated as functionally equivalent and hence interchangeable. The equivalence sets learned in this way, in turn, enable picking and assembling blocks to customize networks subject to certain constraints, which is achieved via solving an integer program backed up with a training-free proxy to estimate the task performance. The reassembled models, give rise to gratifying performances with the user-specified constraints satisfied. We demonstrate that on ImageNet, the best reassemble model achieves 78.6% top-1 accuracy without fine-tuning, which could be further elevated to 83.2% with end-to-end training. Our code is available at https://github.com/Adamdad/DeRy

Sucheng Ren, Xingyi Yang, Songhua Liu et al.

Vision Transformer has demonstrated impressive success across various vision tasks. However, its heavy computation cost, which grows quadratically with respect to the token sequence length, largely limits its power in handling large feature maps. To alleviate the computation cost, previous works rely on either fine-grained self-attentions restricted to local small regions, or global self-attentions but to shorten the sequence length resulting in coarse granularity. In this paper, we propose a novel model, termed as Self-guided Transformer~(SG-Former), towards effective global self-attention with adaptive fine granularity. At the heart of our approach is to utilize a significance map, which is estimated through hybrid-scale self-attention and evolves itself during training, to reallocate tokens based on the significance of each region. Intuitively, we assign more tokens to the salient regions for achieving fine-grained attention, while allocating fewer tokens to the minor regions in exchange for efficiency and global receptive fields. The proposed SG-Former achieves performance superior to state of the art: our base size model achieves \textbf{84.7\%} Top-1 accuracy on ImageNet-1K, \textbf{51.2mAP} bbAP on CoCo, \textbf{52.7mIoU} on ADE20K surpassing the Swin Transformer by \textbf{+1.3\% / +2.7 mAP/ +3 mIoU}, with lower computation costs and fewer parameters. The code is available at \href{https://github.com/OliverRensu/SG-Former}{https://github.com/OliverRensu/SG-Former}

Songhua Liu, Jingwen Ye, Sucheng Ren et al.

One key challenge of exemplar-guided image generation lies in establishing fine-grained correspondences between input and guided images. Prior approaches, despite the promising results, have relied on either estimating dense attention to compute per-point matching, which is limited to only coarse scales due to the quadratic memory cost, or fixing the number of correspondences to achieve linear complexity, which lacks flexibility. In this paper, we propose a dynamic sparse attention based Transformer model, termed Dynamic Sparse Transformer (DynaST), to achieve fine-level matching with favorable efficiency. The heart of our approach is a novel dynamic-attention unit, dedicated to covering the variation on the optimal number of tokens one position should focus on. Specifically, DynaST leverages the multi-layer nature of Transformer structure, and performs the dynamic attention scheme in a cascaded manner to refine matching results and synthesize visually-pleasing outputs. In addition, we introduce a unified training objective for DynaST, making it a versatile reference-based image translation framework for both supervised and unsupervised scenarios. Extensive experiments on three applications, pose-guided person image generation, edge-based face synthesis, and undistorted image style transfer, demonstrate that DynaST achieves superior performance in local details, outperforming the state of the art while reducing the computational cost significantly. Our code is available at https://github.com/Huage001/DynaST

Songhua Liu, Jingwen Ye, Xinchao Wang

Style transfer aims to render the style of a given image for style reference to another given image for content reference, and has been widely adopted in artistic generation and image editing. Existing approaches either apply the holistic style of the style image in a global manner, or migrate local colors and textures of the style image to the content counterparts in a pre-defined way. In either case, only one result can be generated for a specific pair of content and style images, which therefore lacks flexibility and is hard to satisfy different users with different preferences. We propose here a novel strategy termed Any-to-Any Style Transfer to address this drawback, which enables users to interactively select styles of regions in the style image and apply them to the prescribed content regions. In this way, personalizable style transfer is achieved through human-computer interaction. At the heart of our approach lies in (1) a region segmentation module based on Segment Anything, which supports region selection with only some clicks or drawing on images and thus takes user inputs conveniently and flexibly; (2) and an attention fusion module, which converts inputs from users to controlling signals for the style transfer model. Experiments demonstrate the effectiveness for personalizable style transfer. Notably, our approach performs in a plug-and-play manner portable to any style transfer method and enhance the controllablity. Our code is available \href{https://github.com/Huage001/Transfer-Any-Style}{here}.

Jingwen Ye, Songhua Liu, Xinchao Wang

In this paper, we study a novel task that enables partial knowledge transfer from pre-trained models, which we term as Partial Network Cloning (PNC). Unlike prior methods that update all or at least part of the parameters in the target network throughout the knowledge transfer process, PNC conducts partial parametric "cloning" from a source network and then injects the cloned module to the target, without modifying its parameters. Thanks to the transferred module, the target network is expected to gain additional functionality, such as inference on new classes; whenever needed, the cloned module can be readily removed from the target, with its original parameters and competence kept intact. Specifically, we introduce an innovative learning scheme that allows us to identify simultaneously the component to be cloned from the source and the position to be inserted within the target network, so as to ensure the optimal performance. Experimental results on several datasets demonstrate that, our method yields a significant improvement of 5% in accuracy and 50% in locality when compared with parameter-tuning based methods. Our code is available at https://github.com/JngwenYe/PNCloning.

Ruonan Yu, Songhua Liu, Zigeng Chen et al.

Dataset distillation or condensation aims to condense a large-scale training dataset into a much smaller synthetic one such that the training performance of distilled and original sets on neural networks are similar. Although the number of training samples can be reduced substantially, current state-of-the-art methods heavily rely on enormous soft labels to achieve satisfactory performance. As a result, the required storage can be comparable even to original datasets, especially for large-scale ones. To solve this problem, instead of storing these heavy labels, we propose a novel label-lightening framework termed HeLlO aiming at effective image-to-label projectors, with which synthetic labels can be directly generated online from synthetic images. Specifically, to construct such projectors, we leverage prior knowledge in open-source foundation models, e.g., CLIP, and introduce a LoRA-like fine-tuning strategy to mitigate the gap between pre-trained and target distributions, so that original models for soft-label generation can be distilled into a group of low-rank matrices. Moreover, an effective image optimization method is proposed to further mitigate the potential error between the original and distilled label generators. Extensive experiments demonstrate that with only about 0.003% of the original storage required for a complete set of soft labels, we achieve comparable performance to current state-of-the-art dataset distillation methods on large-scale datasets. Our code will be available.

Ruonan Yu, Songhua Liu, Xinchao Wang

Recent success of deep learning is largely attributed to the sheer amount of data used for training deep neural networks.Despite the unprecedented success, the massive data, unfortunately, significantly increases the burden on storage and transmission and further gives rise to a cumbersome model training process. Besides, relying on the raw data for training \emph{per se} yields concerns about privacy and copyright. To alleviate these shortcomings, dataset distillation~(DD), also known as dataset condensation (DC), was introduced and has recently attracted much research attention in the community. Given an original dataset, DD aims to derive a much smaller dataset containing synthetic samples, based on which the trained models yield performance comparable with those trained on the original dataset. In this paper, we give a comprehensive review and summary of recent advances in DD and its application. We first introduce the task formally and propose an overall algorithmic framework followed by all existing DD methods. Next, we provide a systematic taxonomy of current methodologies in this area, and discuss their theoretical interconnections. We also present current challenges in DD through extensive experiments and envision possible directions for future works.

Jingwen Ye, Yifang Fu, Jie Song et al.

Life-long learning aims at learning a sequence of tasks without forgetting the previously acquired knowledge. However, the involved training data may not be life-long legitimate due to privacy or copyright reasons. In practical scenarios, for instance, the model owner may wish to enable or disable the knowledge of specific tasks or specific samples from time to time. Such flexible control over knowledge transfer, unfortunately, has been largely overlooked in previous incremental or decremental learning methods, even at a problem-setup level. In this paper, we explore a novel learning scheme, termed as Learning wIth Recoverable Forgetting (LIRF), that explicitly handles the task- or sample-specific knowledge removal and recovery. Specifically, LIRF brings in two innovative schemes, namely knowledge deposit and withdrawal, which allow for isolating user-designated knowledge from a pre-trained network and injecting it back when necessary. During the knowledge deposit process, the specified knowledge is extracted from the target network and stored in a deposit module, while the insensitive or general knowledge of the target network is preserved and further augmented. During knowledge withdrawal, the taken-off knowledge is added back to the target network. The deposit and withdraw processes only demand for a few epochs of finetuning on the removal data, ensuring both data and time efficiency. We conduct experiments on several datasets, and demonstrate that the proposed LIRF strategy yields encouraging results with gratifying generalization capability.

Songhua Liu, Weihao Yu, Zhenxiong Tan et al.

Modern diffusion models, particularly those utilizing a Transformer-based UNet for denoising, rely heavily on self-attention operations to manage complex spatial relationships, thus achieving impressive generation performance. However, this existing paradigm faces significant challenges in generating high-resolution visual content due to its quadratic time and memory complexity with respect to the number of spatial tokens. To address this limitation, we aim at a novel linear attention mechanism as an alternative in this paper. Specifically, we begin our exploration from recently introduced models with linear complexity, e.g., Mamba2, RWKV6, Gated Linear Attention, etc, and identify two key features--attention normalization and non-causal inference--that enhance high-resolution visual generation performance. Building on these insights, we introduce a generalized linear attention paradigm, which serves as a low-rank approximation of a wide spectrum of popular linear token mixers. To save the training cost and better leverage pre-trained models, we initialize our models and distill the knowledge from pre-trained StableDiffusion (SD). We find that the distilled model, termed LinFusion, achieves performance on par with or superior to the original SD after only modest training, while significantly reducing time and memory complexity. Extensive experiments on SD-v1.5, SD-v2.1, and SD-XL demonstrate that LinFusion enables satisfactory and efficient zero-shot cross-resolution generation, accommodating ultra-resolution images like 16K on a single GPU. Moreover, it is highly compatible with pre-trained SD components and pipelines, such as ControlNet, IP-Adapter, DemoFusion, DistriFusion, etc, requiring no adaptation efforts. Codes are available at https://github.com/Huage001/LinFusion.

Runpeng Yu, Songhua Liu, Xingyi Yang et al.

Machine learning society has witnessed the emergence of a myriad of Out-of-Distribution (OoD) algorithms, which address the distribution shift between the training and the testing distribution by searching for a unified predictor or invariant feature representation. However, the task of directly mitigating the distribution shift in the unseen testing set is rarely investigated, due to the unavailability of the testing distribution during the training phase and thus the impossibility of training a distribution translator mapping between the training and testing distribution. In this paper, we explore how to bypass the requirement of testing distribution for distribution translator training and make the distribution translation useful for OoD prediction. We propose a portable Distribution Shift Inversion algorithm, in which, before being fed into the prediction model, the OoD testing samples are first linearly combined with additional Gaussian noise and then transferred back towards the training distribution using a diffusion model trained only on the source distribution. Theoretical analysis reveals the feasibility of our method. Experimental results, on both multiple-domain generalization datasets and single-domain generalization datasets, show that our method provides a general performance gain when plugged into a wide range of commonly used OoD algorithms.

Hao Tang, Songhua Liu, Tianwei Lin et al.

Transformer-based models achieve favorable performance in artistic style transfer recently thanks to its global receptive field and powerful multi-head/layer attention operations. Nevertheless, the over-paramerized multi-layer structure increases parameters significantly and thus presents a heavy burden for training. Moreover, for the task of style transfer, vanilla Transformer that fuses content and style features by residual connections is prone to content-wise distortion. In this paper, we devise a novel Transformer model termed as \emph{Master} specifically for style transfer. On the one hand, in the proposed model, different Transformer layers share a common group of parameters, which (1) reduces the total number of parameters, (2) leads to more robust training convergence, and (3) is readily to control the degree of stylization via tuning the number of stacked layers freely during inference. On the other hand, different from the vanilla version, we adopt a learnable scaling operation on content features before content-style feature interaction, which better preserves the original similarity between a pair of content features while ensuring the stylization quality. We also propose a novel meta learning scheme for the proposed model so that it can not only work in the typical setting of arbitrary style transfer, but also adaptable to the few-shot setting, by only fine-tuning the Transformer encoder layer in the few-shot stage for one specific style. Text-guided few-shot style transfer is firstly achieved with the proposed framework. Extensive experiments demonstrate the superiority of Master under both zero-shot and few-shot style transfer settings.

Letian Zhou, Songhua Liu, Xinchao Wang

Prevailing Dataset Distillation (DD) methods leveraging generative models confront two fundamental limitations. First, despite pioneering the use of diffusion models in DD and delivering impressive performance, the vast majority of approaches paradoxically require a diffusion model pre-trained on the full target dataset, undermining the very purpose of DD and incurring prohibitive training costs. Second, although some methods turn to general text-to-image models without relying on such target-specific training, they suffer from a significant distributional mismatch, as the web-scale priors encapsulated in these foundation models fail to faithfully capture the target-specific semantics, leading to suboptimal performance. To tackle these challenges, we propose Core Distribution Alignment (CoDA), a framework that enables effective DD using only an off-the-shelf text-to-image model. Our key idea is to first identify the "intrinsic core distribution" of the target dataset using a robust density-based discovery mechanism. We then steer the generative process to align the generated samples with this core distribution. By doing so, CoDA effectively bridges the gap between general-purpose generative priors and target semantics, yielding highly representative distilled datasets. Extensive experiments suggest that, without relying on a generative model specifically trained on the target dataset, CoDA achieves performance on par with or even superior to previous methods with such reliance across all benchmarks, including ImageNet-1K and its subsets. Notably, it establishes a new state-of-the-art accuracy of 60.4% at the 50-images-per-class (IPC) setup on ImageNet-1K. Our code is available on the project webpage: https://github.com/zzzlt422/CoDA

Ruonan Yu, Zhenxiong Tan, Zigeng Chen et al.

Diffusion Transformers (DiTs) have demonstrated remarkable scalability and quality in image and video generation, prompting growing interest in extending them to controllable generation and editing tasks. However, compared to the image counterparts, progress in video control and editing remains limited, mainly due to the scarcity of paired video data and the high computational cost of training video diffusion models. To address this issue, in this paper, we propose a video-free tuning framework termed ViFeEdit for video diffusion transformers. Without requiring any forms of video training data, ViFeEdit achieves versatile video generation and editing, adapted solely with 2D images. At the core of our approach is an architectural reparameterization that decouples spatial independence from the full 3D attention in modern video diffusion transformers, which enables visually faithful editing while maintaining temporal consistency with only minimal additional parameters. Moreover, this design operates in a dual-path pipeline with separate timestep embeddings for noise scheduling, exhibiting strong adaptability to diverse conditioning signals. Extensive experiments demonstrate that our method delivers promising results of controllable video generation and editing with only minimal training on 2D image data. Codes are available https://github.com/Lexie-YU/ViFeEdit.

Songhua Liu, Zhenxiong Tan, Xinchao Wang

Diffusion Transformers (DiT) have become a leading architecture in image generation. However, the quadratic complexity of attention mechanisms, which are responsible for modeling token-wise relationships, results in significant latency when generating high-resolution images. To address this issue, we aim at a linear attention mechanism in this paper that reduces the complexity of pre-trained DiTs to linear. We begin our exploration with a comprehensive summary of existing efficient attention mechanisms and identify four key factors crucial for successful linearization of pre-trained DiTs: locality, formulation consistency, high-rank attention maps, and feature integrity. Based on these insights, we introduce a convolution-like local attention strategy termed CLEAR, which limits feature interactions to a local window around each query token, and thus achieves linear complexity. Our experiments indicate that, by fine-tuning the attention layer on merely 10K self-generated samples for 10K iterations, we can effectively transfer knowledge from a pre-trained DiT to a student model with linear complexity, yielding results comparable to the teacher model. Simultaneously, it reduces attention computations by 99.5% and accelerates generation by 6.3 times for generating 8K-resolution images. Furthermore, we investigate favorable properties in the distilled attention layers, such as zero-shot generalization cross various models and plugins, and improved support for multi-GPU parallel inference. Models and codes are available here: https://github.com/Huage001/CLEAR.

Yingcheng Hu, Haowen Gong, Chuanguang Yang et al.

Pose-guided human image animation aims to synthesize realistic videos of a reference character driven by a sequence of poses. While diffusion-based methods have achieved remarkable success, most existing approaches are limited to single-character animation. We observe that naively extending these methods to multi-character scenarios often leads to identity confusion and implausible occlusions between characters. To address these challenges, in this paper, we propose an extensible multi-character image animation framework built upon modern Diffusion Transformers (DiTs) for video generation. At its core, our framework introduces two novel components-Identifier Assigner and Identifier Adapter - which collaboratively capture per-person positional cues and inter-person spatial relationships. This mask-driven scheme, along with a scalable training strategy, not only enhances flexibility but also enables generalization to scenarios with more characters than those seen during training. Remarkably, trained on only a two-character dataset, our model generalizes to multi-character animation while maintaining compatibility with single-character cases. Extensive experiments demonstrate that our approach achieves state-of-the-art performance in multi-character image animation, surpassing existing diffusion-based baselines.

Wei Chow, Linfeng Li, Lingdong Kong et al.

Recent advances in diffusion models (DMs) have achieved exceptional visual quality in image editing tasks. However, the global denoising dynamics of DMs inherently conflate local editing targets with the full-image context, leading to unintended modifications in non-target regions. In this paper, we shift our attention beyond DMs and turn to Masked Generative Transformers (MGTs) as an alternative approach to tackle this challenge. By predicting multiple masked tokens rather than holistic refinement, MGTs exhibit a localized decoding paradigm that endows them with the inherent capacity to explicitly preserve non-relevant regions during the editing process. Building upon this insight, we introduce the first MGT-based image editing framework, termed EditMGT. We first demonstrate that MGT's cross-attention maps provide informative localization signals for localizing edit-relevant regions and devise a multi-layer attention consolidation scheme that refines these maps to achieve fine-grained and precise localization. On top of these adaptive localization results, we introduce region-hold sampling, which restricts token flipping within low-attention areas to suppress spurious edits, thereby confining modifications to the intended target regions and preserving the integrity of surrounding non-target areas. To train EditMGT, we construct CrispEdit-2M, a high-resolution dataset spanning seven diverse editing categories. Without introducing additional parameters, we adapt a pre-trained text-to-image MGT into an image editing model through attention injection. Extensive experiments across four standard benchmarks demonstrate that, with fewer than 1B parameters, our model achieves similarity performance while enabling 6 times faster editing. Moreover, it delivers comparable or superior editing quality, with improvements of 3.6% and 17.6% on style change and style transfer tasks, respectively.

Zhibin Qin, Zhenxiong Tan, Zeqing Wang et al.

Diffusion Transformer models have significantly advanced image editing by encoding conditional images and integrating them into transformer layers. However, most edits involve modifying only small regions, while current methods uniformly process and denoise all tokens at every timestep, causing redundant computation and potentially degrading unchanged areas. This raises a fundamental question: Is it truly necessary to regenerate every region during editing? To address this, we propose SpotEdit, a training-free diffusion editing framework that selectively updates only the modified regions. SpotEdit comprises two key components: SpotSelector identifies stable regions via perceptual similarity and skips their computation by reusing conditional image features; SpotFusion adaptively blends these features with edited tokens through a dynamic fusion mechanism, preserving contextual coherence and editing quality. By reducing unnecessary computation and maintaining high fidelity in unmodified areas, SpotEdit achieves efficient and precise image editing.

Lingao Xiao, Songhua Liu, Yang He et al.

Dataset distillation and dataset pruning are two prominent techniques for compressing datasets to improve computational and storage efficiency. Despite their overlapping objectives, these approaches are rarely compared directly. Even within each field, the evaluation protocols are inconsistent across various methods, which complicates fair comparisons and hinders reproducibility. Considering these limitations, we introduce in this paper a benchmark that equitably evaluates methodologies across both distillation and pruning literatures. Notably, our benchmark reveals that in the mainstream dataset distillation setting for large-scale datasets, which heavily rely on soft labels from pre-trained models, even randomly selected subsets can achieve surprisingly competitive performance. This finding suggests that an overemphasis on soft labels may be diverting attention from the intrinsic value of the image data, while also imposing additional burdens in terms of generation, storage, and application. To address these issues, we propose a new framework for dataset compression, termed Prune, Combine, and Augment (PCA), which focuses on leveraging image data exclusively, relies solely on hard labels for evaluation, and achieves state-of-the-art performance in this setup. By shifting the emphasis back to the images, our benchmark and PCA framework pave the way for more balanced and accessible techniques in dataset compression research. Our code is available at: https://github.com/ArmandXiao/Rethinking-Dataset-Compression

Junyuan Zhang, Songhua Liu, Xinchao Wang

One-shot Federated learning (FL) is a powerful technology facilitating collaborative training of machine learning models in a single round of communication. While its superiority lies in communication efficiency and privacy preservation compared to iterative FL, one-shot FL often compromises model performance. Prior research has primarily focused on employing data-free knowledge distillation to optimize data generators and ensemble models for better aggregating local knowledge into the server model. However, these methods typically struggle with data heterogeneity, where inconsistent local data distributions can cause teachers to provide misleading knowledge. Additionally, they may encounter scalability issues with complex datasets due to inherent two-step information loss: first, during local training (from data to model), and second, when transferring knowledge to the server model (from model to inversed data). In this paper, we propose FedSD2C, a novel and practical one-shot FL framework designed to address these challenges. FedSD2C introduces a distiller to synthesize informative distillates directly from local data to reduce information loss and proposes sharing synthetic distillates instead of inconsistent local models to tackle data heterogeneity. Our empirical results demonstrate that FedSD2C consistently outperforms other one-shot FL methods with more complex and real datasets, achieving up to 2.6 the performance of the best baseline. Code: https://github.com/Carkham/FedSD2C

Ruonan Yu, Songhua Liu, Zhenxiong Tan et al.

Text-to-image diffusion models have achieved remarkable progress in recent years. However, training models for high-resolution image generation remains challenging, particularly when training data and computational resources are limited. In this paper, we explore this practical problem from two key perspectives: data and parameter efficiency, and propose a set of key guidelines for ultra-resolution adaptation termed \emph{URAE}. For data efficiency, we theoretically and empirically demonstrate that synthetic data generated by some teacher models can significantly promote training convergence. For parameter efficiency, we find that tuning minor components of the weight matrices outperforms widely-used low-rank adapters when synthetic data are unavailable, offering substantial performance gains while maintaining efficiency. Additionally, for models leveraging guidance distillation, such as FLUX, we show that disabling classifier-free guidance, \textit{i.e.}, setting the guidance scale to 1 during adaptation, is crucial for satisfactory performance. Extensive experiments validate that URAE achieves comparable 2K-generation performance to state-of-the-art closed-source models like FLUX1.1 [Pro] Ultra with only 3K samples and 2K iterations, while setting new benchmarks for 4K-resolution generation. Codes are available \href{https://github.com/Huage001/URAE}{here}.

Yuhe Liu, Zhenxiong Tan, Yujia Hu et al.

Recent advances in diffusion-based controllable visual generation have led to remarkable improvements in image quality. However, these powerful models are typically deployed on cloud servers due to their large computational demands, raising serious concerns about user data privacy. To enable secure and efficient on-device generation, we explore in this paper controllable diffusion models built upon linear attention architectures, which offer superior scalability and efficiency, even on edge devices. Yet, our experiments reveal that existing controllable generation frameworks, such as ControlNet and OminiControl, either lack the flexibility to support multiple heterogeneous condition types or suffer from slow convergence on such linear-attention models. To address these limitations, we propose a novel controllable diffusion framework tailored for linear attention backbones like SANA. The core of our method lies in a unified gated conditioning module working in a dual-path pipeline, which effectively integrates multi-type conditional inputs, such as spatially aligned and non-aligned cues. Extensive experiments on multiple tasks and benchmarks demonstrate that our approach achieves state-of-the-art controllable generation performance based on linear-attention models, surpassing existing methods in terms of fidelity and controllability.

Yujia Hu, Songhua Liu, Xingyi Yang et al.

Existing text-to-3D and image-to-3D models often struggle with complex scenes involving multiple objects and intricate interactions. Although some recent attempts have explored such compositional scenarios, they still require an extensive process of optimizing the entire layout, which is highly cumbersome if not infeasible at all. To overcome these challenges, we propose Flash Sculptor in this paper, a simple yet effective framework for compositional 3D scene/object reconstruction from a single image. At the heart of Flash Sculptor lies a divide-and-conquer strategy, which decouples compositional scene reconstruction into a sequence of sub-tasks, including handling the appearance, rotation, scale, and translation of each individual instance. Specifically, for rotation, we introduce a coarse-to-fine scheme that brings the best of both worlds--efficiency and accuracy--while for translation, we develop an outlier-removal-based algorithm that ensures robust and precise parameters in a single step, without any iterative optimization. Extensive experiments demonstrate that Flash Sculptor achieves at least a 3 times speedup over existing compositional 3D methods, while setting new benchmarks in compositional 3D reconstruction performance. Codes are available at https://github.com/YujiaHu1109/Flash-Sculptor.

Zhenxiong Tan, Songhua Liu, Xingyi Yang et al.

We present OminiControl, a novel approach that rethinks how image conditions are integrated into Diffusion Transformer (DiT) architectures. Current image conditioning methods either introduce substantial parameter overhead or handle only specific control tasks effectively, limiting their practical versatility. OminiControl addresses these limitations through three key innovations: (1) a minimal architectural design that leverages the DiT's own VAE encoder and transformer blocks, requiring just 0.1% additional parameters; (2) a unified sequence processing strategy that combines condition tokens with image tokens for flexible token interactions; and (3) a dynamic position encoding mechanism that adapts to both spatially-aligned and non-aligned control tasks. Our extensive experiments show that this streamlined approach not only matches but surpasses the performance of specialized methods across multiple conditioning tasks. To overcome data limitations in subject-driven generation, we also introduce Subjects200K, a large-scale dataset of identity-consistent image pairs synthesized using DiT models themselves. This work demonstrates that effective image control can be achieved without architectural complexity, opening new possibilities for efficient and versatile image generation systems.

Yunfeng Wu, Hongying Cheng, Zihao He et al.

Transformer-based video diffusion models rely on 3D attention over spatial and temporal tokens, which incurs quadratic time and memory complexity and makes end-to-end training for ultra-high-resolution videos prohibitively expensive. To overcome this bottleneck, we propose a pure image adaptation framework that upgrades a video Diffusion Transformer pre-trained at its native scale to synthesize higher-resolution videos. Unfortunately, naively fine-tuning with high-resolution images alone often introduces noticeable noise due to the image-video modality gap. To address this, we decouple the learning objective to separately handle modality alignment and spatial extrapolation. At the core of our approach is Relay LoRA, a two-stage adaptation strategy. In the first stage, the video diffusion model is adapted to the image domain using low-resolution images to bridge the modality gap. In the second stage, the model is further adapted with high-resolution images to acquire spatial extrapolation capability. During inference, only the high-resolution adaptation is retained to preserve the video generation modality while enabling high-resolution video synthesis. To enhance fine-grained detail synthesis, we further propose a High-Frequency-Awareness-Training-Objective, which explicitly encourages the model to recover high-frequency components from degraded latent representations via a dedicated reconstruction loss. Extensive experiments demonstrate that our method produces ultra-high-resolution videos with rich visual details without requiring any video training data, even outperforming previous state-of-the-art models trained on high-resolution videos by 0.8 on the VBench benchmark. Code will be available at https://github.com/WillWu111/ViBe.

Yunfeng Wu, Jiayi Song, Zhenxiong Tan et al.

The quadratic time and memory complexity of the attention mechanism in modern Transformer based video generators makes end-to-end training for ultra high resolution videos prohibitively expensive. Motivated by this limitation, we introduce a training-free approach that leverages video Diffusion Transformers pretrained at their native scale to synthesize higher resolution videos without any additional training or adaptation. At the core of our method lies an inward sliding window attention mechanism, which originates from a key observation: maintaining each query token's training scale receptive field is crucial for preserving visual fidelity and detail. However, naive local window attention, unfortunately, often leads to repetitive content and exhibits a lack of global coherence in the generated results. To overcome this challenge, we devise a dual-path pipeline that backs up window attention with a novel cross-attention override strategy, enabling the semantic content produced by local attention to be guided by another branch with a full receptive field and, therefore, ensuring holistic consistency. Furthermore, to improve efficiency, we incorporate a cross-attention caching strategy for this branch to avoid the frequent computation of full 3D attention. Extensive experiments demonstrate that our method delivers ultra-high-resolution videos with fine-grained visual details and high efficiency in a training-free paradigm. Meanwhile, it achieves superior performance on VBench, even compared to training-based alternatives, with competitive or improved efficiency. Codes are available at: https://github.com/WillWu111/FreeSwim

Wei Chow, Linfeng Li, Xian Sun et al.

Diffusion models dominate image editing, yet their global denoising mechanism entangles edited regions with surrounding context, causing modifications to propagate into areas that should remain intact. We propose a fundamentally different approach by leveraging Masked Generative Transformers (MGTs), whose localized token-prediction paradigm naturally confines changes to intended regions. We present EditMGT, an MGT-based editing framework that is the first of its kind. Our approach employs multi-layer attention consolidation to aggregate cross-attention maps into precise edit localization signals, and region-hold sampling to explicitly prevent token flipping in non-target areas. To support training, we construct CrispEdit-2M, a 2M-sample high-resolution (>1024) editing dataset spanning seven categories. With only 960M parameters, EditMGT achieves state-of-the-art image similarity on multiple benchmarks while delivering 6x faster editing, demonstrating that MGTs offer a compelling alternative to diffusion-based editing.

Jiajun Li, Yue Ma, Xinyu Zhang et al.

Recent studies on Visual Autoregressive (VAR) models have highlighted that high-frequency components, or later steps, in the generation process contribute disproportionately to inference latency. However, the underlying computational redundancy involved in these steps has yet to be thoroughly investigated. In this paper, we conduct an in-depth analysis of the VAR inference process and identify two primary sources of inefficiency: step redundancy and unconditional branch redundancy. To address step redundancy, we propose an automatic step-skipping strategy that selectively omits unnecessary generation steps to improve efficiency. For unconditional branch redundancy, we observe that the information gap between the conditional and unconditional branches is minimal. Leveraging this insight, we introduce unconditional branch replacement, a technique that bypasses the unconditional branch to reduce computational cost. Notably, we observe that the effectiveness of acceleration strategies varies significantly across different samples. Motivated by this, we propose SkipVAR, a sample-adaptive framework that leverages frequency information to dynamically select the most suitable acceleration strategy for each instance. To evaluate the role of high-frequency information, we introduce high-variation benchmark datasets that test model sensitivity to fine details. Extensive experiments show SkipVAR achieves over 0.88 average SSIM with up to 1.81x overall acceleration and 2.62x speedup on the GenEval benchmark, maintaining model quality. These results confirm the effectiveness of frequency-aware, training-free adaptive acceleration for scalable autoregressive image generation. Our code is available at https://github.com/fakerone-li/SkipVAR and has been publicly released.

Yujia Hu, Songhua Liu, Zhenxiong Tan et al.

While diffusion models have achieved remarkable success in text-to-image generation, they encounter significant challenges with instruction-driven image editing. Our research highlights a key challenge: these models particularly struggle with structurally inconsistent edits that involve substantial layout changes. To mitigate this gap, we introduce Image Editing As Programs (IEAP), a unified image editing framework built upon the Diffusion Transformer (DiT) architecture. At its core, IEAP approaches instructional editing through a reductionist lens, decomposing complex editing instructions into sequences of atomic operations. Each operation is implemented via a lightweight adapter sharing the same DiT backbone and is specialized for a specific type of edit. Programmed by a vision-language model (VLM)-based agent, these operations collaboratively support arbitrary and structurally inconsistent transformations. By modularizing and sequencing edits in this way, IEAP generalizes robustly across a wide range of editing tasks, from simple adjustments to substantial structural changes. Extensive experiments demonstrate that IEAP significantly outperforms state-of-the-art methods on standard benchmarks across various editing scenarios. In these evaluations, our framework delivers superior accuracy and semantic fidelity, particularly for complex, multi-step instructions. Codes are available at https://github.com/YujiaHu1109/IEAP.

Songhua Liu, Xin Jin, Xingyi Yang et al.

Single domain generalization (single DG) aims at learning a robust model generalizable to unseen domains from only one training domain, making it a highly ambitious and challenging task. State-of-the-art approaches have mostly relied on data augmentations, such as adversarial perturbation and style enhancement, to synthesize new data and thus increase robustness. Nevertheless, they have largely overlooked the underlying coherence between the augmented domains, which in turn leads to inferior results in real-world scenarios. In this paper, we propose a simple yet effective scheme, termed as \emph{StyDeSty}, to explicitly account for the alignment of the source and pseudo domains in the process of data augmentation, enabling them to interact with each other in a self-consistent manner and further giving rise to a latent domain with strong generalization power. The heart of StyDeSty lies in the interaction between a \emph{stylization} module for generating novel stylized samples using the source domain, and a \emph{destylization} module for transferring stylized and source samples to a latent domain to learn content-invariant features. The stylization and destylization modules work adversarially and reinforce each other. During inference, the destylization module transforms the input sample with an arbitrary style shift to the latent domain, in which the downstream tasks are carried out. Specifically, the location of the destylization layer within the backbone network is determined by a dedicated neural architecture search (NAS) strategy. We evaluate StyDeSty on multiple benchmarks and demonstrate that it yields encouraging results, outperforming the state of the art by up to {13.44%} on classification accuracy. Codes are available here: https://github.com/Huage001/StyDeSty.

Shizun Wang, Songhua Liu, Zhenxiong Tan et al.

Brain decoding, a pivotal field in neuroscience, aims to reconstruct stimuli from acquired brain signals, primarily utilizing functional magnetic resonance imaging (fMRI). Currently, brain decoding is confined to a per-subject-per-model paradigm, limiting its applicability to the same individual for whom the decoding model is trained. This constraint stems from three key challenges: 1) the inherent variability in input dimensions across subjects due to differences in brain size; 2) the unique intrinsic neural patterns, influencing how different individuals perceive and process sensory information; 3) limited data availability for new subjects in real-world scenarios hampers the performance of decoding models. In this paper, we present a novel approach, MindBridge, that achieves cross-subject brain decoding by employing only one model. Our proposed framework establishes a generic paradigm capable of addressing these challenges by introducing biological-inspired aggregation function and novel cyclic fMRI reconstruction mechanism for subject-invariant representation learning. Notably, by cycle reconstruction of fMRI, MindBridge can enable novel fMRI synthesis, which also can serve as pseudo data augmentation. Within the framework, we also devise a novel reset-tuning method for adapting a pretrained model to a new subject. Experimental results demonstrate MindBridge's ability to reconstruct images for multiple subjects, which is competitive with dedicated subject-specific models. Furthermore, with limited data for a new subject, we achieve a high level of decoding accuracy, surpassing that of subject-specific models. This advancement in cross-subject brain decoding suggests promising directions for wider applications in neuroscience and indicates potential for more efficient utilization of limited fMRI data in real-world scenarios. Project page: https://littlepure2333.github.io/MindBridge

Yijun Lu, Rui Ye, Yuwen Du et al.

Long-horizon search agents must manage a rapidly growing working context as they reason, call tools, and observe information. Naively accumulating all intermediate content can overwhelm the agent, increasing costs and the risk of errors. We propose that effective context management should be adaptive: parts of the agent's trajectory are maintained at different levels of detail depending on their current relevance to the task. To operationalize this principle, we introduce Context-ReAct, a general agentic paradigm for elastic context orchestration that integrates reasoning, context management, and tool use in a unified loop. Context-ReAct provides five atomic operations: Skip, Compress, Rollback, Snippet and Delete, which allow the agent to dynamically reshape its working context, preserving important evidence, summarizing resolved information, discarding unhelpful branches, and controlling context size. We prove that the Compress operator is expressively complete, while the other specialized operators provide efficiency and fidelity guarantees that reduce generation cost and hallucination risk. Building on this paradigm, we develop LongSeeker, a long-horizon search agent fine-tuned from Qwen3-30B-A3B on 10k synthesized trajectories. Across four representative search benchmarks, LongSeeker achieves 61.5% on BrowseComp and 62.5% on BrowseComp-ZH, substantially outperforming Tongyi DeepResearch (43.2% and 46.7%) and AgentFold (36.2% and 47.3%). These results highlight the potential of adaptive context management, showing that agents can achieve more reliable and efficient long-horizon reasoning by actively shaping their working memory.

Songping Wang, Rufan Qian, Yueming Lyu et al.

Image-to-Video (I2V) generation synthesizes dynamic visual content from image and text inputs, providing significant creative control. However, the security of such multimodal systems, particularly their vulnerability to jailbreak attacks, remains critically underexplored. To bridge this gap, we propose RunawayEvil, the first multimodal jailbreak framework for I2V models with dynamic evolutionary capability. Built on a "Strategy-Tactic-Action" paradigm, our framework exhibits self-amplifying attack through three core components: (1) Strategy-Aware Command Unit that enables the attack to self-evolve its strategies through reinforcement learning-driven strategy customization and LLM-based strategy exploration; (2) Multimodal Tactical Planning Unit that generates coordinated text jailbreak instructions and image tampering guidelines based on the selected strategies; (3) Tactical Action Unit that executes and evaluates the multimodal coordinated attacks. This self-evolving architecture allows the framework to continuously adapt and intensify its attack strategies without human intervention. Extensive experiments demonstrate RunawayEvil achieves state-of-the-art attack success rates on commercial I2V models, such as Open-Sora 2.0 and CogVideoX. Specifically, RunawayEvil outperforms existing methods by 58.5 to 79 percent on COCO2017. This work provides a critical tool for vulnerability analysis of I2V models, thereby laying a foundation for more robust video generation systems.

Zhenxiong Tan, Qiaochu Xue, Xingyi Yang et al.

Fine-grained control of text-to-image diffusion transformer models (DiT) remains a critical challenge for practical deployment. While recent advances such as OminiControl and others have enabled a controllable generation of diverse control signals, these methods face significant computational inefficiency when handling long conditional inputs. We present OminiControl2, an efficient framework that achieves efficient image-conditional image generation. OminiControl2 introduces two key innovations: (1) a dynamic compression strategy that streamlines conditional inputs by preserving only the most semantically relevant tokens during generation, and (2) a conditional feature reuse mechanism that computes condition token features only once and reuses them across denoising steps. These architectural improvements preserve the original framework's parameter efficiency and multi-modal versatility while dramatically reducing computational costs. Our experiments demonstrate that OminiControl2 reduces conditional processing overhead by over 90% compared to its predecessor, achieving an overall 5.9$\times$ speedup in multi-conditional generation scenarios. This efficiency enables the practical implementation of complex, multi-modal control for high-quality image synthesis with DiT models.

Haoyu Chen, Xiaojie Xu, Wenbo Li et al.

Poster design is a critical medium for visual communication. Prior work has explored automatic poster design using deep learning techniques, but these approaches lack text accuracy, user customization, and aesthetic appeal, limiting their applicability in artistic domains such as movies and exhibitions, where both clear content delivery and visual impact are essential. To address these limitations, we present POSTA: a modular framework powered by diffusion models and multimodal large language models (MLLMs) for customized artistic poster generation. The framework consists of three modules. Background Diffusion creates a themed background based on user input. Design MLLM then generates layout and typography elements that align with and complement the background style. Finally, to enhance the poster's aesthetic appeal, ArtText Diffusion applies additional stylization to key text elements. The final result is a visually cohesive and appealing poster, with a fully modular process that allows for complete customization. To train our models, we develop the PosterArt dataset, comprising high-quality artistic posters annotated with layout, typography, and pixel-level stylized text segmentation. Our comprehensive experimental analysis demonstrates POSTA's exceptional controllability and design diversity, outperforming existing models in both text accuracy and aesthetic quality.

Jingwen Ye, Ruonan Yu, Songhua Liu et al.

Adversarial attacks constitute a notable threat to machine learning systems, given their potential to induce erroneous predictions and classifications. However, within real-world contexts, the essential specifics of the deployed model are frequently treated as a black box, consequently mitigating the vulnerability to such attacks. Thus, enhancing the transferability of the adversarial samples has become a crucial area of research, which heavily relies on selecting appropriate surrogate models. To address this challenge, we propose a novel approach that generates adversarial attacks in a mutual-modality optimization scheme. Our approach is accomplished by leveraging the pre-trained CLIP model. Firstly, we conduct a visual attack on the clean image that causes semantic perturbations on the aligned embedding space with the other textual modality. Then, we apply the corresponding defense on the textual modality by updating the prompts, which forces the re-matching on the perturbed embedding space. Finally, to enhance the attack transferability, we utilize the iterative training strategy on the visual attack and the textual defense, where the two processes optimize from each other. We evaluate our approach on several benchmark datasets and demonstrate that our mutual-modal attack strategy can effectively produce high-transferable attacks, which are stable regardless of the target networks. Our approach outperforms state-of-the-art attack methods and can be readily deployed as a plug-and-play solution.

Jingwen Ye, Ruonan Yu, Songhua Liu et al.

The proliferation of large-scale AI models trained on extensive datasets has revolutionized machine learning. With these models taking on increasingly central roles in various applications, the need to understand their behavior and enhance interpretability has become paramount. To investigate the impact of changes in training data on a pre-trained model, a common approach is leave-one-out retraining. This entails systematically altering the training dataset by removing specific samples to observe resulting changes within the model. However, retraining the model for each altered dataset presents a significant computational challenge, given the need to perform this operation for every dataset variation. In this paper, we introduce an efficient framework for assessing data impact, comprising offline training and online evaluation stages. During the offline training phase, we approximate the influence of training data on the target model through a distilled synset, formulated as a reversed gradient matching problem. For online evaluation, we expedite the leave-one-out process using the synset, which is then utilized to compute the attribution matrix based on the evaluation objective. Experimental evaluations, including training data attribution and assessments of data quality, demonstrate that our proposed method achieves comparable model behavior evaluation while significantly speeding up the process compared to the direct retraining method.

Bonan Li, Yinhan Hu, Songhua Liu et al.

Layout-to-Image generation aims to create complex scenes with precise control over the placement and arrangement of subjects. Existing works have demonstrated that pre-trained Text-to-Image diffusion models can achieve this goal without training on any specific data; however, they often face challenges with imprecise localization and unrealistic artifacts. Focusing on these drawbacks, we propose a novel training-free method, WinWinLay. At its core, WinWinLay presents two key strategies, Non-local Attention Energy Function and Adaptive Update, that collaboratively enhance control precision and realism. On one hand, we theoretically demonstrate that the commonly used attention energy function introduces inherent spatial distribution biases, hindering objects from being uniformly aligned with layout instructions. To overcome this issue, non-local attention prior is explored to redistribute attention scores, facilitating objects to better conform to the specified spatial conditions. On the other hand, we identify that the vanilla backpropagation update rule can cause deviations from the pre-trained domain, leading to out-of-distribution artifacts. We accordingly introduce a Langevin dynamics-based adaptive update scheme as a remedy that promotes in-domain updating while respecting layout constraints. Extensive experiments demonstrate that WinWinLay excels in controlling element placement and achieving photorealistic visual fidelity, outperforming the current state-of-the-art methods.

Bonan li, Zicheng Zhang, Songhua Liu et al.

Visual instruction tuning aims to enable large language models to comprehend the visual world, with a pivotal challenge lying in establishing an effective vision-to-language projection. However, existing methods often grapple with the intractable trade-off between accuracy and efficiency. In this paper, we present LLaVA-Meteor, a novel approach designed to break this deadlock, equipped with a novel Top-Down Compression paradigm that strategically compresses visual tokens without compromising core information. Specifically, we construct a trainable Flash Global Fusion module based on efficient selective state space operators, which aligns the feature space while enabling each token to perceive holistic visual context and instruction preference at low cost. Furthermore, a local-to-single scanning manner is employed to effectively capture local dependencies, thereby enhancing the model's capability in vision modeling. To alleviate computational overhead, we explore a Visual-Native Selection mechanism that independently assesses token significance by both the visual and native experts, followed by aggregation to retain the most critical subset. Extensive experiments show that our approach reduces visual tokens by 75--95% while achieving comparable or superior performance across 12 benchmarks, significantly improving efficiency.

Deyu Bo, Songhua Liu, Xinchao Wang

Dataset distillation (DD) has emerged as a promising approach to compress datasets and speed up model training. However, the underlying connections among various DD methods remain largely unexplored. In this paper, we introduce UniDD, a spectral filtering framework that unifies diverse DD objectives. UniDD interprets each DD objective as a specific filter function that affects the eigenvalues of the feature-feature correlation (FFC) matrix and modulates the frequency components of the feature-label correlation (FLC) matrix. In this way, UniDD reveals that the essence of DD fundamentally lies in matching frequency-specific features. Moreover, according to the filter behaviors, we classify existing methods into low-frequency matching and high-frequency matching, encoding global texture and local details, respectively. However, existing methods rely on fixed filter functions throughout distillation, which cannot capture the low- and high-frequency information simultaneously. To address this limitation, we further propose Curriculum Frequency Matching (CFM), which gradually adjusts the filter parameter to cover both low- and high-frequency information of the FFC and FLC matrices. Extensive experiments on small-scale datasets, such as CIFAR-10/100, and large-scale datasets, including ImageNet-1K, demonstrate the superior performance of CFM over existing baselines and validate the practicality of UniDD.

Zhenxiong Tan, Zeqing Wang, Xingyi Yang et al.

We introduce Vision Bridge Transformer (ViBT), a large-scale instantiation of Brownian Bridge Models designed for conditional generation. Unlike traditional diffusion models that transform noise into data, Bridge Models directly model the trajectory between inputs and outputs, creating an efficient data-to-data translation paradigm. By scaling these models to 20B and 1.3B parameters, we demonstrate their effectiveness for image and video translation tasks. To support this scale, we adopt a Transformer architecture and propose a variance-stabilized velocity-matching objective for robust training. Together, these advances highlight the power of scaling Bridge Models for instruction-based image editing and complex video translation.

Zhenxiong Tan, Xingyi Yang, Songhua Liu et al.

Diffusion models have recently achieved remarkable results for video generation. Despite the encouraging performances, the generated videos are typically constrained to a small number of frames, resulting in clips lasting merely a few seconds. The primary challenges in producing longer videos include the substantial memory requirements and the extended processing time required on a single GPU. A straightforward solution would be to split the workload across multiple GPUs, which, however, leads to two issues: (1) ensuring all GPUs communicate effectively to share timing and context information, and (2) modifying existing video diffusion models, which are usually trained on short sequences, to create longer videos without additional training. To tackle these, in this paper we introduce Video-Infinity, a distributed inference pipeline that enables parallel processing across multiple GPUs for long-form video generation. Specifically, we propose two coherent mechanisms: Clip parallelism and Dual-scope attention. Clip parallelism optimizes the gathering and sharing of context information across GPUs which minimizes communication overhead, while Dual-scope attention modulates the temporal self-attention to balance local and global contexts efficiently across the devices. Together, the two mechanisms join forces to distribute the workload and enable the fast generation of long videos. Under an 8 x Nvidia 6000 Ada GPU (48G) setup, our method generates videos up to 2,300 frames in approximately 5 minutes, enabling long video generation at a speed 100 times faster than the prior methods.

Songhua Liu, Tianwei Lin, Dongliang He et al.

Neural painting refers to the procedure of producing a series of strokes for a given image and non-photo-realistically recreating it using neural networks. While reinforcement learning (RL) based agents can generate a stroke sequence step by step for this task, it is not easy to train a stable RL agent. On the other hand, stroke optimization methods search for a set of stroke parameters iteratively in a large search space; such low efficiency significantly limits their prevalence and practicality. Different from previous methods, in this paper, we formulate the task as a set prediction problem and propose a novel Transformer-based framework, dubbed Paint Transformer, to predict the parameters of a stroke set with a feed forward network. This way, our model can generate a set of strokes in parallel and obtain the final painting of size 512 * 512 in near real time. More importantly, since there is no dataset available for training the Paint Transformer, we devise a self-training pipeline such that it can be trained without any off-the-shelf dataset while still achieving excellent generalization capability. Experiments demonstrate that our method achieves better painting performance than previous ones with cheaper training and inference costs. Codes and models are available.

Songhua Liu, Tianwei Lin, Dongliang He et al.

Fast arbitrary neural style transfer has attracted widespread attention from academic, industrial and art communities due to its flexibility in enabling various applications. Existing solutions either attentively fuse deep style feature into deep content feature without considering feature distributions, or adaptively normalize deep content feature according to the style such that their global statistics are matched. Although effective, leaving shallow feature unexplored and without locally considering feature statistics, they are prone to unnatural output with unpleasing local distortions. To alleviate this problem, in this paper, we propose a novel attention and normalization module, named Adaptive Attention Normalization (AdaAttN), to adaptively perform attentive normalization on per-point basis. Specifically, spatial attention score is learnt from both shallow and deep features of content and style images. Then per-point weighted statistics are calculated by regarding a style feature point as a distribution of attention-weighted output of all style feature points. Finally, the content feature is normalized so that they demonstrate the same local feature statistics as the calculated per-point weighted style feature statistics. Besides, a novel local feature loss is derived based on AdaAttN to enhance local visual quality. We also extend AdaAttN to be ready for video style transfer with slight modifications. Experiments demonstrate that our method achieves state-of-the-art arbitrary image/video style transfer. Codes and models are available.