Low-latency decoding for large language models (LLMs) is crucial for applications like chatbots and code assistants, yet generating long outputs remains slow in single-query settings. Prior work on speculative decoding (which combines a small draft model with a larger target model) and tensor parallelism has each accelerated decoding. However, conventional approaches fail to apply both simultaneously due to imbalanced compute requirements (between draft and target models), KV-cache inconsistencies, and communication overheads under small-batch tensor-parallelism. This paper introduces SwiftSpec, a system that targets ultra-low latency for LLM decoding. SwiftSpec redesigns the speculative decoding pipeline in an asynchronous and disaggregated manner, so that each component can be scaled flexibly and remove draft overhead from the critical path. To realize this design, SwiftSpec proposes parallel tree generation, tree-aware KV cache management, and fused, latency-optimized kernels to overcome the challenges listed above. Across 5 model families and 6 datasets, SwiftSpec achieves an average of 1.75x speedup over state-of-the-art speculative decoding systems and, as a highlight, serves Llama3-70B at 348 tokens/s on 8 Nvidia Hopper GPUs, making it the fastest known system for low-latency LLM serving at this scale.
Recent methods for visual question answering rely on large-scale annotated datasets. Manual annotation of questions and answers for videos, however, is tedious, expensive and prevents scalability. In this work, we propose to avoid manual annotation and generate a large-scale training dataset for video question answering making use of automatic cross-modal supervision. We leverage a question generation transformer trained on text data and use it to generate question-answer pairs from transcribed video narrations. Given narrated videos, we then automatically generate the HowToVQA69M dataset with 69M video-question-answer triplets. To handle the open vocabulary of diverse answers in this dataset, we propose a training procedure based on a contrastive loss between a video-question multi-modal transformer and an answer transformer. We introduce the zero-shot VideoQA task and show excellent results, in particular for rare answers. Furthermore, we demonstrate our method to significantly outperform the state of the art on MSRVTT-QA, MSVD-QA, ActivityNet-QA and How2QA. Finally, for a detailed evaluation we introduce iVQA, a new VideoQA dataset with reduced language biases and high-quality redundant manual annotations. Our code, datasets and trained models are available at https://antoyang.github.io/just-ask.html.
Upon the discovery of adversarial attacks, robust models have become obligatory for deep learning-based systems. Adversarial training with first-order attacks has been one of the most effective defenses against adversarial perturbations to this day. The majority of the adversarial training approaches focus on iteratively perturbing each pixel with the gradient of the loss function with respect to the input image. However, the adversarial training with gradient-based attacks lacks diversity and does not generalize well to natural images and various attacks. This study presents a robust training algorithm where the adversarial perturbations are automatically synthesized from a random vector using a generator network. The classifier is trained with cross-entropy loss regularized with the optimal transport distance between the representations of the natural and synthesized adversarial samples. Unlike prevailing generative defenses, the proposed one-step attack generation framework synthesizes diverse perturbations without utilizing gradient of the classifier's loss. Experimental results show that the proposed approach attains comparable robustness with various gradient-based and generative robust training techniques on CIFAR10, CIFAR100, and SVHN datasets. In addition, compared to the baselines, the proposed robust training framework generalizes well to the natural samples. Code and trained models will be made publicly available.
Though recent end-to-end neural models have shown promising progress on Conversational Recommender System (CRS), two key challenges still remain. First, the recommended items cannot be always incorporated into the generated replies precisely and appropriately. Second, only the items mentioned in the training corpus have a chance to be recommended in the conversation. To tackle these challenges, we introduce a novel framework called NTRD for recommender dialogue system that decouples the dialogue generation from the item recommendation. NTRD has two key components, i.e., response template generator and item selector. The former adopts an encoder-decoder model to generate a response template with slot locations tied to target items, while the latter fills in slot locations with the proper items using a sufficient attention mechanism. Our approach combines the strengths of both classical slot filling approaches (that are generally controllable) and modern neural NLG approaches (that are generally more natural and accurate). Extensive experiments on the benchmark ReDial show our NTRD significantly outperforms the previous state-of-the-art methods. Besides, our approach has the unique advantage to produce novel items that do not appear in the training set of dialogue corpus. The code is available at \url{https://github.com/jokieleung/NTRD}.
Magnetic Resonance (MR) image reconstruction from under-sampled acquisition promises faster scanning time. To this end, current State-of-The-Art (SoTA) approaches leverage deep neural networks and supervised training to learn a recovery model. While these approaches achieve impressive performances, the learned model can be fragile on unseen degradation, e.g. when given a different acceleration factor. These methods are also generally deterministic and provide a single solution to an ill-posed problem; as such, it can be difficult for practitioners to understand the reliability of the reconstruction. We introduce DiffuseRecon, a novel diffusion model-based MR reconstruction method. DiffuseRecon guides the generation process based on the observed signals and a pre-trained diffusion model, and does not require additional training on specific acceleration factors. DiffuseRecon is stochastic in nature and generates results from a distribution of fully-sampled MR images; as such, it allows us to explicitly visualize different potential reconstruction solutions. Lastly, DiffuseRecon proposes an accelerated, coarse-to-fine Monte-Carlo sampling scheme to approximate the most likely reconstruction candidate. The proposed DiffuseRecon achieves SoTA performances reconstructing from raw acquisition signals in fastMRI and SKM-TEA. Code will be open-sourced at www.github.com/cpeng93/DiffuseRecon.
Capitalizing on the recent advances in image generation models, existing controllable face image synthesis methods are able to generate high-fidelity images with some levels of controllability, e.g., controlling the shapes, expressions, textures, and poses of the generated face images. However, these methods focus on 2D image generative models, which are prone to producing inconsistent face images under large expression and pose changes. In this paper, we propose a new NeRF-based conditional 3D face synthesis framework, which enables 3D controllability over the generated face images by imposing explicit 3D conditions from 3D face priors. At its core is a conditional Generative Occupancy Field (cGOF) that effectively enforces the shape of the generated face to commit to a given 3D Morphable Model (3DMM) mesh. To achieve accurate control over fine-grained 3D face shapes of the synthesized image, we additionally incorporate a 3D landmark loss as well as a volume warping loss into our synthesis algorithm. Experiments validate the effectiveness of the proposed method, which is able to generate high-fidelity face images and shows more precise 3D controllability than state-of-the-art 2D-based controllable face synthesis methods. Find code and demo at https://keqiangsun.github.io/projects/cgof.
We present a novel bipartite graph reasoning Generative Adversarial Network (BiGraphGAN) for two challenging tasks: person pose and facial image synthesis. The proposed graph generator consists of two novel blocks that aim to model the pose-to-pose and pose-to-image relations, respectively. Specifically, the proposed bipartite graph reasoning (BGR) block aims to reason the long-range cross relations between the source and target pose in a bipartite graph, which mitigates some of the challenges caused by pose deformation. Moreover, we propose a new interaction-and-aggregation (IA) block to effectively update and enhance the feature representation capability of both a person's shape and appearance in an interactive way. To further capture the change in pose of each part more precisely, we propose a novel part-aware bipartite graph reasoning (PBGR) block to decompose the task of reasoning the global structure transformation with a bipartite graph into learning different local transformations for different semantic body/face parts. Experiments on two challenging generation tasks with three public datasets demonstrate the effectiveness of the proposed methods in terms of objective quantitative scores and subjective visual realness. The source code and trained models are available at https://github.com/Ha0Tang/BiGraphGAN.
The image captioning task is typically realized by an auto-regressive method that decodes the text tokens one by one. We present a diffusion-based captioning model, dubbed the name DDCap, to allow more decoding flexibility. Unlike image generation, where the output is continuous and redundant with a fixed length, texts in image captions are categorical and short with varied lengths. Therefore, naively applying the discrete diffusion model to text decoding does not work well, as shown in our experiments. To address the performance gap, we propose several key techniques including best-first inference, concentrated attention mask, text length prediction, and image-free training. On COCO without additional caption pre-training, it achieves a CIDEr score of 117.8, which is +5.0 higher than the auto-regressive baseline with the same architecture in the controlled setting. It also performs +26.8 higher CIDEr score than the auto-regressive baseline (230.3 v.s.203.5) on a caption infilling task. With 4M vision-language pre-training images and the base-sized model, we reach a CIDEr score of 125.1 on COCO, which is competitive to the best well-developed auto-regressive frameworks. The code is available at https://github.com/buxiangzhiren/DDCap.
The recent success of text-to-image generation diffusion models has also revolutionized semantic image editing, enabling the manipulation of images based on query/target texts. Despite these advancements, a significant challenge lies in the potential introduction of contextual prior bias in pre-trained models during image editing, e.g., making unexpected modifications to inappropriate regions. To address this issue, we present a novel approach called Dual-Cycle Diffusion, which generates an unbiased mask to guide image editing. The proposed model incorporates a Bias Elimination Cycle that consists of both a forward path and an inverted path, each featuring a Structural Consistency Cycle to ensure the preservation of image content during the editing process. The forward path utilizes the pre-trained model to produce the edited image, while the inverted path converts the result back to the source image. The unbiased mask is generated by comparing differences between the processed source image and the edited image to ensure that both conform to the same distribution. Our experiments demonstrate the effectiveness of the proposed method, as it significantly improves the D-CLIP score from 0.272 to 0.283. The code will be available at https://github.com/JohnDreamer/DualCycleDiffsion.
The state of the arts in vision-language pretraining (VLP) achieves exemplary performance but suffers from high training costs resulting from slow convergence and long training time, especially on large-scale web datasets. An essential obstacle to training efficiency lies in the entangled prediction rate (percentage of tokens for reconstruction) and corruption rate (percentage of corrupted tokens) in masked language modeling (MLM), that is, a proper corruption rate is achieved at the cost of a large portion of output tokens being excluded from prediction loss. To accelerate the convergence of VLP, we propose a new pretraining task, namely, free language modeling (FLM), that enables a 100% prediction rate with arbitrary corruption rates. FLM successfully frees the prediction rate from the tie-up with the corruption rate while allowing the corruption spans to be customized for each token to be predicted. FLM-trained models are encouraged to learn better and faster given the same GPU time by exploiting bidirectional contexts more flexibly. Extensive experiments show FLM could achieve an impressive 2.5x pretraining time reduction in comparison to the MLM-based methods, while keeping competitive performance on both vision-language understanding and generation tasks. Code will be public at https://github.com/TencentARC/FLM.
Text-to-image diffusion models are nothing but a revolution, allowing anyone, even without design skills, to create realistic images from simple text inputs. With powerful personalization tools like DreamBooth, they can generate images of a specific person just by learning from his/her few reference images. However, when misused, such a powerful and convenient tool can produce fake news or disturbing content targeting any individual victim, posing a severe negative social impact. In this paper, we explore a defense system called Anti-DreamBooth against such malicious use of DreamBooth. The system aims to add subtle noise perturbation to each user's image before publishing in order to disrupt the generation quality of any DreamBooth model trained on these perturbed images. We investigate a wide range of algorithms for perturbation optimization and extensively evaluate them on two facial datasets over various text-to-image model versions. Despite the complicated formulation of DreamBooth and Diffusion-based text-to-image models, our methods effectively defend users from the malicious use of those models. Their effectiveness withstands even adverse conditions, such as model or prompt/term mismatching between training and testing. Our code will be available at https://github.com/VinAIResearch/Anti-DreamBooth.git.
Fundus photography is an essential examination for clinical and differential diagnosis of fundus diseases. Recently, Ultra-Wide-angle Fundus (UWF) techniques, UWF Fluorescein Angiography (UWF-FA) and UWF Scanning Laser Ophthalmoscopy (UWF-SLO) have been gradually put into use. However, Fluorescein Angiography (FA) and UWF-FA require injecting sodium fluorescein which may have detrimental influences. To avoid negative impacts, cross-modality medical image generation algorithms have been proposed. Nevertheless, current methods in fundus imaging could not produce high-resolution images and are unable to capture tiny vascular lesion areas. This paper proposes a novel conditional generative adversarial network (UWAT-GAN) to synthesize UWF-FA from UWF-SLO. Using multi-scale generators and a fusion module patch to better extract global and local information, our model can generate high-resolution images. Moreover, an attention transmit module is proposed to help the decoder learn effectively. Besides, a supervised approach is used to train the network using multiple new weighted losses on different scales of data. Experiments on an in-house UWF image dataset demonstrate the superiority of the UWAT-GAN over the state-of-the-art methods. The source code is available at: https://github.com/Tinysqua/UWAT-GAN.
Audio-Visual Segmentation (AVS) aims to precisely outline audible objects in a visual scene at the pixel level. Existing AVS methods require fine-grained annotations of audio-mask pairs in supervised learning fashion. This limits their scalability since it is time consuming and tedious to acquire such cross-modality pixel level labels. To overcome this obstacle, in this work we introduce unsupervised audio-visual segmentation with no need for task-specific data annotations and model training. For tackling this newly proposed problem, we formulate a novel Cross-Modality Semantic Filtering (CMSF) approach to accurately associate the underlying audio-mask pairs by leveraging the off-the-shelf multi-modal foundation models (e.g., detection [1], open-world segmentation [2] and multi-modal alignment [3]). Guiding the proposal generation by either audio or visual cues, we design two training-free variants: AT-GDINO-SAM and OWOD-BIND. Extensive experiments on the AVS-Bench dataset show that our unsupervised approach can perform well in comparison to prior art supervised counterparts across complex scenarios with multiple auditory objects. Particularly, in situations where existing supervised AVS methods struggle with overlapping foreground objects, our models still excel in accurately segmenting overlapped auditory objects. Our code will be publicly released.
Diffusion models build a new milestone for image generation yet raising public concerns, for they can be fine-tuned on unauthorized images for customization. Protection based on adversarial attacks rises to encounter this unauthorized diffusion customization, by adding protective watermarks to images and poisoning diffusion models. However, current protection, leveraging untargeted attacks, does not appear to be effective enough. In this paper, we propose a simple yet effective improvement for the protection against unauthorized diffusion customization by introducing targeted attacks. We show that by carefully selecting the target, targeted attacks significantly outperform untargeted attacks in poisoning diffusion models and degrading the customization image quality. Extensive experiments validate the superiority of our method on two mainstream customization methods of diffusion models, compared to existing protections. To explain the surprising success of targeted attacks, we delve into the mechanism of attack-based protections and propose a hypothesis based on our observation, which enhances the comprehension of attack-based protections. To the best of our knowledge, we are the first to both reveal the vulnerability of diffusion models to targeted attacks and leverage targeted attacks to enhance protection against unauthorized diffusion customization. Our code is available on GitHub: \url{https://github.com/psyker-team/mist-v2}.
The practical deployment of diffusion models still suffers from the high memory and time overhead. While quantization paves a way for compression and acceleration, existing methods unfortunately fail when the models are quantized to low-bits. In this paper, we empirically unravel three properties in quantized diffusion models that compromise the efficacy of current methods: imbalanced activation distributions, imprecise temporal information, and vulnerability to perturbations of specific modules. To alleviate the intensified low-bit quantization difficulty stemming from the distribution imbalance, we propose finetuning the quantized model to better adapt to the activation distribution. Building on this idea, we identify two critical types of quantized layers: those holding vital temporal information and those sensitive to reduced bit-width, and finetune them to mitigate performance degradation with efficiency. We empirically verify that our approach modifies the activation distribution and provides meaningful temporal information, facilitating easier and more accurate quantization. Our method is evaluated over three high-resolution image generation tasks and achieves state-of-the-art performance under various bit-width settings, as well as being the first method to generate readable images on full 4-bit (i.e. W4A4) Stable Diffusion. Code is available \href{https://github.com/hatchetProject/QuEST}{here}.
The Audio Description (AD) task aims to generate descriptions of visual elements for visually impaired individuals to help them access long-form video content, like movies. With video feature, text, character bank and context information as inputs, the generated ADs are able to correspond to the characters by name and provide reasonable, contextual descriptions to help audience understand the storyline of movie. To achieve this goal, we propose to leverage pre-trained foundation models through a simple and unified framework to generate ADs with interleaved multimodal sequence as input, termed as Uni-AD. To enhance the alignment of features across various modalities with finer granularity, we introduce a simple and lightweight module that maps video features into the textual feature space. Moreover, we also propose a character-refinement module to provide more precise information by identifying the main characters who play more significant roles in the video context. With these unique designs, we further incorporate contextual information and a contrastive loss into our architecture to generate smoother and more contextually appropriate ADs. Experiments on multiple AD datasets show that Uni-AD performs well on AD generation, which demonstrates the effectiveness of our approach. Our code is available at: https://github.com/ant-research/UniAD.
Despite the impressive capability of large language models (LLMs), knowing when to trust their generations remains an open challenge. The recent literature on uncertainty quantification of natural language generation (NLG) utilises a conventional natural language inference (NLI) classifier to measure the semantic dispersion of LLMs responses. These studies employ logits of NLI classifier for semantic clustering to estimate uncertainty. However, logits represent the probability of the predicted class and barely contain feature information for potential clustering. Alternatively, CLIP (Contrastive Language-Image Pre-training) performs impressively in extracting image-text pair features and measuring their similarity. To extend its usability, we propose Contrastive Semantic Similarity, the CLIP-based feature extraction module to obtain similarity features for measuring uncertainty for text pairs. We apply this method to selective NLG, which detects and rejects unreliable generations for better trustworthiness of LLMs. We conduct extensive experiments with three LLMs on several benchmark question-answering datasets with comprehensive evaluation metrics. Results show that our proposed method performs better in estimating reliable responses of LLMs than comparable baselines. Results show that our proposed method performs better in estimating reliable responses of LLMs than comparable baselines. The code are available at \url{https://github.com/AoShuang92/css_uq_llms}.
Generating realistic human grasps is a crucial yet challenging task for applications involving object manipulation in computer graphics and robotics. Existing methods often struggle with generating fine-grained realistic human grasps that ensure all fingers effectively interact with objects, as they focus on encoding hand with the whole representation and then estimating both hand posture and position in a single step. In this paper, we propose a novel Decomposed Vector-Quantized Variational Autoencoder (DVQ-VAE) to address this limitation by decomposing hand into several distinct parts and encoding them separately. This part-aware decomposed architecture facilitates more precise management of the interaction between each component of hand and object, enhancing the overall reality of generated human grasps. Furthermore, we design a newly dual-stage decoding strategy, by first determining the type of grasping under skeletal physical constraints, and then identifying the location of the grasp, which can greatly improve the verisimilitude as well as adaptability of the model to unseen hand-object interaction. In experiments, our model achieved about 14.1% relative improvement in the quality index compared to the state-of-the-art methods in four widely-adopted benchmarks. Our source code is available at https://github.com/florasion/D-VQVAE.
The application of the Multi-modal Large Language Models (MLLMs) in medical clinical scenarios remains underexplored. Previous benchmarks only focus on the capacity of the MLLMs in medical visual question-answering (VQA) or report generation and fail to assess the performance of the MLLMs on complex clinical multi-modal tasks. In this paper, we propose a novel Medical Personalized Multi-modal Consultation (Med-PMC) paradigm to evaluate the clinical capacity of the MLLMs. Med-PMC builds a simulated clinical environment where the MLLMs are required to interact with a patient simulator to complete the multi-modal information-gathering and decision-making task. Specifically, the patient simulator is decorated with personalized actors to simulate diverse patients in real scenarios. We conduct extensive experiments to access 12 types of MLLMs, providing a comprehensive view of the MLLMs' clinical performance. We found that current MLLMs fail to gather multimodal information and show potential bias in the decision-making task when consulted with the personalized patient simulators. Further analysis demonstrates the effectiveness of Med-PMC, showing the potential to guide the development of robust and reliable clinical MLLMs. Code and data are available at https://github.com/LiuHC0428/Med-PMC.
Accurate classification of haematological cells is critical for diagnosing blood disorders, but presents significant challenges for machine automation owing to the complexity of cell morphology, heterogeneities of biological, pathological, and imaging characteristics, and the imbalance of cell type frequencies. We introduce CytoDiffusion, a diffusion-based classifier that effectively models blood cell morphology, combining accurate classification with robust anomaly detection, resistance to distributional shifts, interpretability, data efficiency, and superhuman uncertainty quantification. Our approach outperforms state-of-the-art discriminative models in anomaly detection (AUC 0.990 vs. 0.918), resistance to domain shifts (85.85% vs. 74.38% balanced accuracy), and performance in low-data regimes (95.88% vs. 94.95% balanced accuracy). Notably, our model generates synthetic blood cell images that are nearly indistinguishable from real images, as demonstrated by an authenticity test in which expert haematologists achieved only 52.3% accuracy (95% CI: [50.5%, 54.2%]) in distinguishing real from generated images. Furthermore, we enhance model explainability through the generation of directly interpretable counterfactual heatmaps. Our comprehensive evaluation framework, encompassing these multiple performance dimensions, establishes a new benchmark for medical image analysis in haematology, ultimately enabling improved diagnostic accuracy in clinical settings. Our code is available at https://github.com/CambridgeCIA/CytoDiffusion.