We propose a method to distill a complex multistep diffusion model into a single-step student model, dramatically accelerating inference while preserving image quality. Our approach interprets the diffusion distillation as a paired image-to-image translation task, using noise-to-image pairs of the diffusion model’s ODE trajectory. For efficient regression loss computation, we propose E-LatentLPIPS, a Perceptual loss in the latent space of the diffusion model with an ensemble of augmentations. Despite dataset construction costs, E-LatentLPIPS converges more efficiently than many existing distillation methods. Furthermore, we adapt a diffusion model to construct a multi-scale discriminator with a text alignment loss to build an effective conditional GAN-based formulation. We demonstrate that our one-step generator outperforms all published diffusion distillation models on the zero-shot COCO2014 benchmark.
论文检索
输入标题、作者或关键词，从 2,387 篇学术成果中精准定位
Enabling machines to understand structured visuals like slides and user interfaces is essential for making them accessible to people with disabilities. However, achieving such understanding computationally has required manual data collection and annotation, which is time-consuming and labor-intensive. To overcome this challenge, we present a method to generate synthetic, structured visuals with target labels using code generation. Our method allows people to create datasets with built-in labels and train models with a small number of human-annotated examples. We demonstrate performance improvements in three tasks for understanding slides and UIs: recognizing visual elements, describing visual content, and classifying visual content types.
In this paper, we introduce L-DiffER, a language-based diffusion model designed for the ill-posed single image reflection removal task. Although having shown impressive performance for image generation, existing language-based diffusion models struggle with precise control and faithfulness in image restoration. To overcome these limitations, we propose an iterative condition refinement strategy to resolve the problem of inaccurate control conditions. A multi-condition constraint mechanism is employed to ensure the recovery faithfulness of image color and structure while retaining the generation capability to handle low-transmitted reflections. We demonstrate the superiority of the proposed method through extensive experiments, showcasing both quantitative and qualitative improvements over existing methods.
We tackle the problem of source-free unsupervised domain adaptation (SFUDA) for 3D semantic segmentation. This challenging problem amounts to performing domain adaptation on an unlabeled target domain without any access to source data. The only available information is a model trained to achieve good performance on the source domain. Our first analysis reveals a pattern which commonly occurs with all SFUDA procedures: performance degrades after some training time, which is a by-product of an under-constrained and ill-posed problem. We discuss two strategies to alleviate this issue. First, we propose a sensible way to regularize the learning problem. Second, we introduce a novel criterion based on agreement with a reference model. It is used (1) to stop the training and (2) as validator to select hyperparameters. Our contributions are easy to implement and readily amenable for all SFUDA methods, ensuring stable improvements over all baselines. We validate our findings on various settings, achieving state-of-the-art performance.
Advancements in open-source pre-trained backbones make it relatively easy to fine-tune a model for new tasks. However, this lowered entry barrier poses potential risks, e.g., bad actors developing models for harmful applications. A question arises: "Is possible to develop a pre-trained model that is difficult to fine-tune for certain downstream tasks?" To begin studying this, we focus on few-shot classification (FSC). Specifically, we investigate methods to make FSC more challenging for a set of restricted classes while maintaining the performance of other classes. We propose to meta-learn over the pre-trained backbone in a manner that renders it a ``poor initialization''. Our proposed Learning to Obstruct (LTO) algorithm successfully obstructs four FSC methods across three datasets, including ImageNet and CIFAR100 for image classification, as well as CelebA for attribute classification.
Talking face generation aims to create a realistic video with accurate lip synchronization and high visual quality, using given audio and reference video, while preserving identity and visual characteristics. In this paper, we start by identifying several issues of existing synchronization learning methods. These involve unstable training, lip synchronization, and visual quality issues caused by lip-sync loss and SyncNet. We further tackle lip leaking problem from the identity reference and propose a silent-lip generator, aiming to prevent lip leaking by changing the lips of the identity reference. We then introduce stabilized synchronization loss and AVSyncNet to alleviate the problems caused by lip-sync loss and SyncNet. Finally, we present adaptive triplet loss to enhance visual quality and apply a post-processing technique to obtain high-quality videos. According to the experiments, our model outperforms state-of-the-art methods in both visual quality and lip synchronization. Comprehensive ablation studies further validate our individual contributions as well as their complementary effects.
Does our behavior change when we speak over video calls? Given two unpaired sets of videos of people, we seek to automatically find spatio-temporal patterns that are distinctive of each set. Existing methods use discriminative approaches and perform post-hoc explainability analysis. Such methods are insufficient as they are unable to provide insights beyond obvious dataset biases, and the explanations are useful only if humans themselves are good at the task. Instead, we tackle the problem through the lens of generative domain translation: our method generates a detailed report of learned, input-dependent spatio-temporal features and the extent to which they vary between the domains. We demonstrate that our method can discover behavioral differences between conversing face-to-face (F2F) and on video-calls (VCs). We also show the applicability of our method on discovering differences in presidential communication styles. Additionally, we are able to predict temporal change-points in videos that decouple expressions in an unsupervised way, and increase the interpretability and usefulness of our model. Finally, our method, being generative, can be used to transform a video call to appear as if it were recorded in a F2F setting. Experiments and visualizations show our approach is able to discover a range of behaviors, taking a step towards deeper understanding of human behaviors.
A common practice in deep learning consists of training large neural networks on massive datasets to perform accurately for different domains and tasks. While this methodology may work well in numerous application areas, it only applies across modalities due to a larger distribution shift in data captured using different sensors. This paper focuses on the problem of adapting a large object detection model to one or multiple modalities while being efficient. To do so, we propose ModTr as an alternative to the common approach of fine-tuning large models. ModTr consists of adapting the input with a small transformation network trained to minimize the detection loss directly. The original model can therefore work on the translated inputs without any further change or fine-tuning to its parameters. Experimental results on translating from IR to RGB images on two well-known datasets show that this simple ModTr approach provides detectors that can perform comparably or better than the standard fine-tuning without forgetting the original knowledge. This opens the doors to a more flexible and efficient service-based detection pipeline in which, instead of using a different detector for each modality, a unique and unaltered server is constantly running, where multiple modalities with the corresponding translations can query it.
In the past few years, large-scale pre-trained vision-language models like CLIP have achieved tremendous success in various fields. Naturally, how to transfer the rich knowledge in such huge pre-trained models to downstream tasks and datasets becomes a hot topic. During downstream adaptation, the most challenging problems are overfitting and catastrophic forgetting, which can cause the model to overly focus on the current data and lose more crucial domain-general knowledge. Existing works use classic regularization techniques to solve the problems. As solutions become increasingly complex, the ever-growing storage and inference costs are also a significant problem that urgently needs to be addressed. While in this paper, we start from an observation that proper random noise can suppress overfitting and catastrophic forgetting. Then we regard quantization error as a kind of noise, and explore quantization for regularizing vision-language model, which is quite efficiency and effective. Furthermore, to improve the model's generalization capability while maintaining its specialization capacity at minimal cost, we deeply analyze the characteristics of the weight distribution in prompts, conclude several principles for quantization module design and follow such principles to create several competitive baselines. The proposed method is significantly efficient due to its inherent lightweight nature, making it possible to adapt on extremely resource-limited devices. Our method can be fruitfully integrated into many existing approaches like MaPLe, enhancing accuracy while reducing storage overhead, making it more powerful yet versatile. Extensive experiments on 11 datasets shows great superiority of our method sufficiently.
Temporal grounding, also known as video moment retrieval, aims at locating video segments corresponding to a given query sentence. The compositional nature of natural language enables the localization beyond predefined events, posing a certain challenge to the compositional generalizability of existing methods. Recent studies establish the correspondence between videos and queries through a decompose-reconstruct manner, utilizing contrastive learning to achieve compositional generalization. However, these approaches have significant drawbacks in constructing negative queries: (1) Only dominant verbs and nouns are considered, overlooking the impact of other primitives such as prepositions and adverbs. (2) These negative samples are formed through random sampling and recombination, resulting in numerous semantically implausible combinations and forcing the model to learn unrealistic semantic differences. To address these limitations, we first propose a large language model-driven method for negative query construction, utilizing GPT-3.5-Turbo to generate semantically plausible hard negative queries. Subsequently, we introduce a coarse-to-fine saliency ranking strategy, which encourages the model to learn the multi-granularity semantic relationships between videos and hierarchical negative queries to boost compositional generalization. Extensive experiments on two challenging benchmarks, Charades-CG and ActivityNet-CG, validate the effectiveness and generalizability of our proposed method. Our code will be available soon.
6-DoF grasp detection has been a fundamental and challenging problem in robotic vision. While previous works have focused on ensuring grasp stability, they often do not consider human intention conveyed through natural language, hindering effective collaboration between robots and users in complex 3D environments. In this paper, we present a new approach for language-driven 6-DoF grasp detection in cluttered point clouds. We first introduce Grasp-Anything-6D, a large-scale dataset for the language-driven 6-DoF grasp detection task with 1M point cloud scenes and more than 200M language-associated 3D grasp poses. We further introduce a novel diffusion model that incorporates a new negative prompt guidance learning strategy. The proposed negative prompt strategy directs the detection process toward the desired object while steering away from unwanted ones given the language input. Our method enables an end-to-end framework where humans can command the robot to grasp desired objects in a cluttered scene using natural language. Intensive experimental results show the effectiveness of our method in both benchmarking experiments and real-world scenarios, surpassing other baselines. In addition, we demonstrate the practicality of our approach in real-world robotic applications.
Multi-view crowd localization predicts the ground locations of all people in the scene. Typical methods usually estimate the crowd density maps on the ground plane first, and then obtain the crowd locations. However, existing methods' performance is limited by the ambiguity of the density maps in crowded areas, where local peaks can be smoothed away. To mitigate the weakness of density map supervision, optimal transport-based point supervision methods have been proposed in the single-image crowd localization tasks, but have not been explored for multi-view crowd localization yet. Thus, in this paper, we propose a novel Mahalanobis distance-based multi-view optimal transport (M-MVOT) loss specifically designed for multi-view crowd localization. First, we replace the Euclidean-based transport cost with the Mahalanobis distance, which defines elliptical iso-contours in the cost function whose long-axis and short-axis directions are guided by the view ray direction. Second, the object-to-camera distance in each view is used to adjust the optimal transport cost of each location further, where the wrong predictions far away from the camera are more heavily penalized. Finally, we propose a strategy to consider all the input camera views in the model loss (M-MVOT) by computing the optimal transport cost for each ground-truth point based on its closest camera. Experiments demonstrate the advantage of the proposed method over density map-based or common Euclidean distance-based optimal transport loss on several multi-view crowd localization datasets.
We present WorldPose, a novel dataset for advancing research in multi-person, global pose estimation in the wild, featuring footage from the 2022 FIFA World Cup. While previous datasets have primarily focused on local poses, often limited to a single person or in constrained, indoor settings, the infrastructure deployed for this sporting event allows access to multiple fixed and mobile cameras in different stadiums. We exploit the static multi-view setup of HD cameras to recover the 3D player poses and motions with unprecedented accuracy given capture areas of more than 1.75 acres (7k sqm). We then leverage the captured players' motions and field markings to calibrate a moving broadcasting camera. The resulting dataset comprises 88 sequences with more than 2.5 million 3D poses and a total traveling distance of over 120 km. Subsequently, we conduct an in-depth analysis of the SOTA methods for global pose estimation. Our experiments demonstrate that WorldPose challenges existing multi-person techniques, supporting the potential for new research in this area and others, such as sports analysis. All pose annotations (in SMPL format), broadcasting camera parameters and footage will be released for academic research purposes.
3D Gaussian Splatting (3DGS) has demonstrated impressive novel view synthesis results while advancing real-time rendering performance. However, it relies heavily on the quality of the initial point cloud, resulting in blurring and needle-like artifacts in areas with insufficient initializing points. This is mainly attributed to the point cloud growth condition in 3DGS that only considers the average gradient magnitude of points from observable views, thereby failing to grow enough points in these areas. To this end, we propose a novel method, named Pixel-GS, to take into account the number of pixels covered by the Gaussian in each view during the computation of the growth condition. We regard the covered pixel numbers as the weights to dynamically average the gradients from different views. Doing so can mitigate the issue of large Gaussians participating in calculations across too many viewpoints, where a significant number of these viewpoints involve calculations for only a few pixels in boundary regions, leading to smaller gradients and consequently lowering the average gradient magnitude. As a result, points within the areas with insufficient initializing points can be grown more effectively, leading to a more accurate and detailed reconstruction. Besides, we also propose a simple yet effective strategy to scale the gradient field according to the distance to the camera, to suppress the growth of floaters near the camera. Extensive experiments both qualitatively and quantitatively demonstrate that our method achieves state-of-the-art rendering quality while maintaining real-time rendering speed, on the challenging Mip-NeRF 360 and Tanks & Temples datasets. The source code will be publicly available.
Recently, the scale of transformers has grown rapidly, which introduces considerable challenges in terms of training overhead and inference efficiency in the scope of task adaptation. Existing works, namely Parameter-Efficient Fine-Tuning (PEFT) and model compression, have separately investigated the challenges. However, PEFT cannot guarantee the inference efficiency of the original backbone, especially for large-scale models. Model compression requires significant training costs for structure searching and re-training. Consequently, a simple combination of them cannot guarantee accomplishing both training efficiency and inference efficiency with minimal costs. In this paper, we propose a novel Parallel Yielding Re-Activation (PYRA) method for such a challenge of training-inference efficient task adaptation. PYRA first utilizes parallel yielding adaptive weights to comprehensively perceive the data distribution in downstream tasks. A re-activation strategy for token modulation is then applied for tokens to be merged, leading to calibrated token features. Extensive experiments demonstrate that PYRA outperforms all competing methods under both low compression rate and high compression rate, demonstrating its effectiveness and superiority in maintaining both training efficiency and inference efficiency for large-scale foundation models. Our code will be released to the public.
Federated Learning (FL) is becoming a popular paradigm for leveraging distributed data and preserving data privacy. However, due to the distributed characteristic, FL systems are vulnerable to Byzantine attacks that compromised clients attack the global model by uploading malicious model updates. With the development of layer-level and parameter-level fine-grained attacks, the attacks' stealthiness and effectiveness have been significantly improved. The existing defense mechanisms solely analyze the model-level statistics of individual model updates uploaded by clients to mitigate Byzantine attacks, which are ineffective against the fine-grained attacks due to unawareness or overreaction. To address this problem, we propose SkyMask, a new attack-agnostic robust FL system that firstly leverages fine-grained learnable masks to identify malicious model updates at the parameter-level. Specifically, the FL server freezes and multiplies the model updates uploaded by clients with the parameter-level masks, and trains the masks over a small clean dataset (i.e., root dataset) to learn the subtle difference between benign and malicious model updates in a high-dimension space. Our extensive experiments involve different models on three public datasets under state-of-the-art (SOTA) attacks, where the results show that SkyMask achieves up to 14% higher testing accuracy compared with SOTA defense strategies under the same attacks and successfully defends against attacks with malicious clients of a high fraction up to 80%. SkyMask will be open-sourced upon acceptance.
Dance, as an art form, fundamentally hinges on the precise synchronization with musical beats. However, achieving aesthetically pleasing dance sequences from music is challenging, with existing methods often falling short in controllability and beat alignment. To address these shortcomings, this paper introduces Beat-It, a novel framework for beat-specific, key pose-guided dance generation. Unlike prior approaches, Beat-It uniquely integrates explicit beat awareness and key pose guidance, effectively resolving two main issues: the misalignment of generated dance motions with musical beats, and the inability to map key poses to specific beats, critical for practical choreography. Our approach disentangles beat conditions from music using a nearest beat distance representation and employs a hierarchical multi-condition fusion mechanism. This mechanism seamlessly integrates key poses, beats, and music features, mitigating condition conflicts and offering rich, multi-conditioned guidance for dance generation. Additionally, a specially designed beat alignment loss ensures the generated dance movements remain in sync with the designated beats. Extensive experiments confirm Beat-It's superiority over existing state-of-the-art methods in terms of beat alignment and motion controllability. Qualitative results of our method can be found on our anonymous website.
The correct insertion of virtual objects in images of real-world scenes requires a deep understanding of the scene's lighting, geometry and materials, as well as the image formation process. While recent large-scale diffusion models have shown strong generative and inpainting capabilities, we find that current models do not sufficiently ``understand'' the scene shown in a single picture to generate consistent lighting effects (shadows, bright reflections, etc.) while preserving the identity and details of the composited object. We propose using a personalized large diffusion model as guidance to a physically based inverse rendering process. Our method recovers scene lighting and tone-mapping parameters, allowing the photorealistic composition of arbitrary virtual objects in single frames or videos of indoor or outdoor scenes. Our physically based pipeline further enables automatic materials and tone-mapping refinement.
We introduce Videoshop, a training-free video editing algorithm for localized semantic edits. Videoshop allows users to use any editing software, including Photoshop and generative inpainting, to modify the first frame; it automatically propagates those changes, with semantic, spatial, and temporally consistent motion, to the remaining frames. Unlike existing methods that enable edits only through imprecise textual instructions, Videoshop allows users to add or remove objects, semantically change objects, insert stock photos into videos, etc. with fine-grained control over locations and appearance. We achieve this through image-based video editing by inverting latents with noise extrapolation, from which we generate videos conditioned on the edited image. Videoshop produces higher quality edits against 6 baselines on 2 editing benchmarks using 10 evaluation metrics.
Learn to Preserve and Diversify: Parameter-Efficient Group with Orthogonal Regularization for Domain Generalization
PDF ↗Domain generalization (DG) aims to avoid the performance degradation of the model when the distribution shift between the limited training data and unseen test data occurs. Recently, foundation models with enormous parameters have been pre-trained with huge datasets, demonstrating strong generalization ability and showing promising direction for solving the DG problem. However, fully Fine-Tuning (FT) the foundation models results in unsatisfactory out-of-distribution accuracy due to the destroyed pre-trained generalized features. Recently, Parameter-Efficient Fine-Tuning (PEFT) alleviates the above problem by fine-tuning a small portion of the model parameters while keeping the rest frozen, which achieves better generalization performance compared to FT. Nevertheless, PEFT still suffers from the issue of overfitting to the training domains. To address the above issue, we propose Parameter-Efficient Group with Orthogonal regularization (PEGO) for vision transformers, which effectively preserves the generalization ability of the pre-trained network and learns more diverse knowledge compared with conventional PEFT. Specifically, we inject a group of trainable Low-Rank Adaptation (LoRA) modules into the pre-trained model and propose an orthogonal regularization loss to enhance the generalization ability of the model. Our framework achieves SOTA performance on five DG benchmarks, while only requiring training a small number of parameters without adding additional testing cost.