Recent development in Large Language Models (LLMs) and Multi-modal Large Language Models (MLLMs) have achieved superior performance and generalization capabilities, covered extensive areas of traditional tasks. However, existing large model training frameworks support only a limited number of models and techniques, particularly lacking in support for new models, which makes fine-tuning LLMs challenging for most developers. Therefore, we develop SWIFT, a customizable one-stop infrastructure for large models. With support of over 350+ LLMs and 80+ MLLMs, SWIFT stands as the open-source framework that provide the most comprehensive support for fine-tuning large models. In particular, it is the first training framework that provides systematic support for MLLMs. Moreover, SWIFT integrates post-training processes such as inference, evaluation, and quantization, to facilitate fast adoptions of large models in various application scenarios, offering helpful utilities like benchmark comparisons among different training techniques.

Despite their remarkable image generation capabilities, text-to-image diffusion models inadvertently learn inappropriate concepts from vast and unfiltered training data, which leads to various ethical and business risks. Specifically, model-generated images may exhibit not safe for work (NSFW) content and style copyright infringements. The prompts that result in these problems often do not include explicit unsafe words; instead, they contain obscure and associative terms, which are referred to as *implicit unsafe prompts*. Existing approaches directly fine-tune models under textual guidance to alter the cognition of the diffusion model, thereby erasing inappropriate concepts. This not only requires concept-specific fine-tuning but may also incur catastrophic forgetting. To address these issues, we explore the representation of inappropriate concepts in the image space and guide them towards more suitable ones by injecting *growth inhibitors*, which are tailored based on the identified features related to inappropriate concepts during the diffusion process. Additionally, due to the varying degrees and scopes of inappropriate concepts, we train an adapter to infer the corresponding suppression scale during the injection process. Our method effectively captures the manifestation of subtle words at the image level, enabling direct and efficient erasure of target concepts without the need for fine-tuning. Through extensive experimentation, we demonstrate that our approach achieves superior erasure results with little effect on other normal concepts while preserving image quality and semantics.

Recently, advancements in video synthesis have attracted significant attention. Video synthesis models have demonstrated the practical applicability of diffusion models in creating dynamic visual content. Despite these advancements, the extension of video lengths remains constrained by computational resources. Most existing video synthesis models are limited to generating short video clips. In this paper, we propose a novel post-tuning methodology for video synthesis models, called ExVideo. This approach is designed to enhance the capability of current video synthesis models, allowing them to produce content over extended temporal durations while incurring lower training expenditures. In particular, we design extension strategies across common temporal model architectures respectively, including 3D convolution, temporal attention, and positional embedding. To evaluate the efficacy of our proposed post-tuning approach, we trained ExSVD, an extended model based on Stable Video Diffusion model. Our approach enhances the model's capacity to generate up to 5x its original number of frames, requiring only 1.5k GPU hours of training on a dataset comprising 40k videos. Importantly, the substantial increase in video length doesn't compromise the model's innate generalization capabilities, and the model showcases its advantages in generating videos of diverse styles and resolutions. We have released the source code and the enhanced model publicly.

Story visualization aims to create visual content, such as images and videos, that is consistent, coherent, and complete with a given story. Despite significant advances in the application of diffusion models for general text-to-image generation tasks, they still encounter difficulties when directly used to produce consistent visual content that accurately aligns with the narrative text. In this paper, we propose a novel training-free automated story visualization framework called AgentStory that can generate image illustrations based on a story synopsis provided by users. Specifically, the framework employs multiple agents empowered by Large Language Models (LLMs) to create detailed descriptions of each subject and scene in the entire story. Then, it integrates a masking mechanism with a fine-grained consistency refinement adapter to incorporate different subjects in a scene. Furthermore, it utilizes the visual understanding capabilities of multimodal LLMs to include detailed features of different subjects in the refinement adapter, thus improving the consistency of each subject across multiple scenes. Finally, we compare the AgentStory framework with state-of-the-art baselines for story visualization on the DS-500 dataset and demonstrate its superior performance in terms of subject consistency, text-image alignment, and aesthetic quality. Our code is publicly available at https://github.com/tc2000731/AgentStory.

Xiangyan Liu, Bo Lan, Zhiyuan Hu, Yang Liu, Zhicheng Zhang, Fei Wang, Michael Qizhe Shieh, Wenmeng Zhou. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

The remarkable ability of diffusion models to generate high-fidelity images has led to their widespread adoption. However, concerns have also arisen regarding their potential to produce Not Safe for Work (NSFW) content and exhibit social biases, hindering their practical use in real-world applications. In response to this challenge, prior work has focused on employing security filters to identify and exclude toxic text, or alternatively, fine-tuning pre-trained diffusion models to erase sensitive concepts. Unfortunately, existing methods struggle to achieve satisfactory performance in the sense that they can have a significant impact on the normal model output while still failing to prevent the generation of harmful content in some cases. In this paper, we propose a novel self-discovery approach to identifying a semantic direction vector in the embedding space to restrict text embedding within a safe region. Our method circumvents the need for correcting individual words within the input text and steers the entire text prompt towards a safe region in the embedding space, thereby enhancing model robustness against all possibly unsafe prompts. In addition, we employ Low-Rank Adaptation (LoRA) for semantic direction vector initialization to reduce the impact on the model performance for other semantics. Furthermore, our method can also be integrated with existing methods to improve their social responsibility. Extensive experiments on benchmark datasets demonstrate that our method can effectively reduce NSFW content and mitigate social bias generated by diffusion models compared to several state-of-the-art baselines. WARNING:This paper contains model-generated images that may be potentially offensive.

Textual adversarial attack in black-box scenarios is a challenging task, as only the predicted label is available, and the text space is discrete and non-differentiable. Current research in this area is still in its infancy and mostly focuses on untargeted attack, lacking the capability to control the labels of the generated adversarial examples. Meanwhile, existing textual adversarial attack methods primarily rely on word substitution operations to maintain semantic similarity between the adversarial and original examples, which greatly limits the search space for adversarial examples. To address these issues, we propose a novel Lexical-Syntactic Targeted Adversarial Attack method tailored for the black-box settings, referred to as LST2A. Our approach involves adversarial perturbations at different levels of granularities, i.e., word-level with word substitution operations and syntactic-level through rewriting the syntax of the examples. Specifically, we first embed the entire text into the embedding layer of a masked language model, and then optimize perturbations at the word level within the hidden state to generate adversarial examples with the target label. For examples that are difficult to attack successfully with only word-level perturbations at higher semantic similarity thresholds, we leverage Large Language Model (LLM) to introduce syntactic-level perturbations to these examples, making them more vulnerable to the decision boundary of the victim model. Subsequently, we re-optimize the word-level perturbations for these vulnerable examples. Extensive experiments and human evaluations demonstrate that our proposed method consistently outperforms the state-of-the-art baselines, crafting smoother, more grammatically correct adversarial examples.

A prevailing belief in attack and defense community is that the higher flatness of adversarial examples enables their better cross-model transferability, leading to a growing interest in employing sharpness-aware minimization and its variants. However, the theoretical relationship between the transferability of adversarial examples and their flatness has not been well established, making the belief questionable. To bridge this gap, we embark on a theoretical investigation and, for the first time, derive a theoretical bound for the transferability of adversarial examples with few practical assumptions. Our analysis challenges this belief by demonstrating that the increased flatness of adversarial examples does not necessarily guarantee improved transferability. Moreover, building upon the theoretical analysis, we propose TPA, a Theoretically Provable Attack that optimizes a surrogate of the derived bound to craft adversarial examples. Extensive experiments across widely used benchmark datasets and various real-world applications show that TPA can craft more transferable adversarial examples compared to state-of-the-art baselines. We hope that these results can recalibrate preconceived impressions within the community and facilitate the development of stronger adversarial attack and defense mechanisms.

Federated learning is a privacy-focused learning paradigm, which trains a global model with gradients uploaded from multiple participants, circumventing explicit exposure of private data. However, previous research of gradient leakage attacks suggests that gradients alone are sufficient to reconstruct private data, rendering the privacy protection mechanism of federated learning unreliable. Existing defenses commonly craft transformed gradients based on ground-truth gradients to obfuscate the attacks, but often are less capable of maintaining good model performance together with satisfactory privacy protection. In this paper, we propose a novel yet effective defense framework named guarding against gradient leakage (Guardian) that produces transformed gradients by jointly optimizing two theoretically-derived metrics associated with gradients for performance maintenance and privacy protection. In this way, the transformed gradients produced via Guardian can achieve minimal privacy leakage in theory with the given performance maintenance level. Moreover, we design an ingenious initialization strategy for faster generation of transformed gradients to enhance the practicality of Guardian in real-world applications, while demonstrating theoretical convergence of Guardian to the performance of the global model. Extensive experiments on various tasks show that, without sacrificing much accuracy, Guardian can effectively defend state-of-the-art gradient leakage attacks, compared with the slight effects of baseline defense approaches.

Split learning enables collaborative deep learning model training while preserving data privacy and model security by avoiding direct sharing of raw data and model details (i.e., server and clients only hold partial sub-networks and exchange intermediate computations). However, existing research has mainly focused on examining its reliability for privacy protection, with little investigation into model security. Specifically, by exploring full models, attackers can launch adversarial attacks, and split learning can mitigate this severe threat by only disclosing part of models to untrusted servers. This paper aims to evaluate the robustness of split learning against adversarial attacks, particularly in the most challenging setting where untrusted servers only have access to the intermediate layers of the model. Existing adversarial attacks mostly focus on the centralized setting instead of the collaborative setting, thus, to better evaluate the robustness of split learning, we develop a tailored attack called SLADV, which comprises two stages: 1) shadow model training that addresses the issue of lacking part of the model and 2) local adversarial attack that produces adversarial examples to evaluate. The first stage only requires a few unlabeled non-IID data, and, in the second stage, SLADV perturbs the intermediate output of natural samples to craft the adversarial ones. The overall cost of the proposed attack process is relatively low, yet the empirical attack effectiveness is significantly high, demonstrating the surprising vulnerability of split learning to adversarial attacks.

With the explosive growth of video and text data on the web, text-video retrieval has become a vital task for online video platforms. Recently, text-video retrieval methods based on pre-trained models have attracted a lot of attention. However, existing methods cannot effectively capture the fine-grained information in videos, and typically suffer from the hubness problem where a collection of similar videos are retrieved by a large number of different queries. In this paper, we propose Match4Match, a new text-video retrieval method based on CLIP (Contrastive Language-Image Pretraining) and graph optimization theories. To balance calculation efficiency and model accuracy, Match4Match seamlessly supports three inference modes for different application scenarios. In fast vector retrieval mode, we embed texts and videos in the same space and employ a vector retrieval engine to obtain the top K videos. In fine-grained alignment mode, our method fully utilizes the pre-trained knowledge of the CLIP model to align words with corresponding video frames, and uses the fine-grained information to compute text-video similarity more accurately. In flow-style matching mode, to alleviate the detrimental impact of the hubness problem, we model the retrieval problem as a combinatorial optimization problem and solve it using maximum flow with minimum cost algorithm. To demonstrate the effectiveness of our method, we conduct experiments on five public text-video datasets. The overall performance of our proposed method outperforms state-of-the-art methods. Additionally, we evaluate the computational efficiency of Match4Match. Benefiting from the three flexible inference modes, Match4Match can respond to a large number of query requests with low latency or achieve high recall with acceptable time consumption.

Scene text recognition has attracted increasing attention in computer vision due to its various applications. Most of the existing scene text recognition methods are under the encoder-decoder framework. In order to improve text feature learning of these methods, Generative Adversarial Networks (GANs) are recently integrated to generate clean text images without distorted letters. However, the existing GANs assume the input images are spatially aligned, while the words in natural images are often in irregular shapes. The misalignment brings a big problem for both image generation and text recognition. In this paper, we present a novel text feature alignment network to solve this problem. Our method can handle both horizontal and vertical images with irregular texts. Our proposed framework is end-to-end trainable, and extensive experiments on several public benchmarks demonstrate its superiority in terms of both effectiveness and efficiency.

Incidental scene text detection, especially for multi-oriented text regions, is one of the most challenging tasks in many computer vision applications.Different from the common object detection task, scene text often suffers from a large variance of aspect ratio, scale, and orientation. To solve this problem, we propose a novel end-to-end scene text detector IncepText from an instance-aware segmentation perspective. We design a novel Inception-Text module and introduce deformable PSROI pooling to deal with multi-oriented text detection. Extensive experiments on ICDAR2015, RCTW-17, and MSRA-TD500 datasets demonstrate our method's superiority in terms of both effectiveness and efficiency. Our proposed method achieves 1st place result on ICDAR2015 challenge and the state-of-the-art performance on other datasets. Moreover, we have released our implementation as an OCR product which is available for public access.