Transformer-based large pre-trained models have shown remarkable generalization ability, and various parameter-efficient fine-tuning (PEFT) methods have been proposed to customize these models on downstream tasks with minimal computational and memory budgets. Previous PEFT methods are primarily designed from a tensor-decomposition perspective that tries to effectively tune the linear transformation by finding the smallest subset of parameters to train. Our study adopts an orthogonal view by representing the attention operation as a graph convolution and formulating the multi-head attention maps as a convolutional filter subspace, with each attention map as a subspace element. In this paper, we propose to tune the large pre-trained transformers by learning a small set of combination coefficients that construct a more expressive filter subspace from the original multi-head attention maps. We show analytically and experimentally that the tuned filter subspace can effectively expand the feature space of the multi-head attention and further enhance the capacity of transformers. We further stabilize the fine-tuning with a residual parameterization of the tunable subspace coefficients, and enhance the generalization with a regularization design by directly applying dropout on the tunable coefficient during training. The tunable coefficients take a tiny number of parameters and can be combined with previous PEFT methods in a plug-and-play manner. Extensive experiments show that our approach achieves superior performances than PEFT baselines with neglectable additional parameters.
论文检索
输入标题、作者或关键词，从 2,871 篇学术成果中精准定位
Generative models have made significant impacts across various domains, largely due to their ability to scale during training by increasing data, computational resources, and model size, a phenomenon characterized by the scaling laws. Recent research has begun to explore inference-time scaling behavior in Large Language Models (LLMs), revealing how performance can further improve with additional computation during inference. Unlike LLMs, diffusion models inherently possess the flexibility to adjust inference-time computation via the number of denoising steps, although the performance gains typically flatten after a few dozen. In this work, we explore the inference-time scaling behavior of diffusion models beyond increasing denoising steps and investigate how the generation performance can further improve with increased computation. Specifically, we consider a search problem aimed at identifying better noises for the diffusion sampling process. We structure the design space along two axes: the verifiers used to provide feedback, and the algorithms used to find better noise candidates. Through extensive experiments on class-conditioned and text-conditioned image generation benchmarks, our findings reveal that increasing inference-time compute leads to substantial improvements in the quality of samples generated by diffusion models, and with the complicated nature of images, combinations of the components in the framework can be specifically chosen to conform with different application scenario.
We present SpectroMotion, a novel approach that combines 3D Gaussian Splatting (3DGS) with physically-based rendering (PBR) and deformation fields to reconstruct dynamic specular scenes. Previous methods extending 3DGS to model dynamic scenes have struggled to represent specular surfaces accurately. Our method addresses this limitation by introducing a residual correction technique for accurate surface normal computation during deformation, complemented by a deformable environment map that adapts to time-varying lighting conditions. We implement a coarse-to-fine training strategy significantly enhancing scene geometry and specular color prediction. It is the only existing 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, outperforming state-of-the-art methods in rendering complex, dynamic, and specular scenes. Please see our project page at https://cdfan0627.github.io/spectromotion/.
Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens
PDF ↗Hallucinations in Large Vision-Language Models (LVLMs) significantly undermine their reliability, motivating researchers to explore the causes of hallucination. However, most studies primarily focus on the language aspect rather than the visual. In this paper, we address how LVLMs process visual information and whether this process causes hallucination. Firstly, we use the attention lens to identify the stages at which LVLMs handle visual data, discovering that the middle layers are crucial. Moreover, we find that these layers can be further divided into two stages: "visual information enrichment" and "semantic refinement" which respectively propagate visual data to object tokens and interpret it through text. By analyzing attention patterns during the visual information enrichment stage, we find that real tokens consistently receive higher attention weights than hallucinated ones, serving as a strong indicator of hallucination. Further examination of multi-head attention maps reveals that hallucination tokens often result from heads interacting with inconsistent objects. Based on these insights, we propose a simple inference-time method that adjusts visual attention by integrating information across various heads. Extensive experiments demonstrate that this approach effectively mitigates hallucinations in mainstream LVLMs without additional training costs.
A variety of text-guided image editing models have been proposed recently. However, there is no widely-accepted standard evaluation method mainly due to the subjective nature of the task, letting researchers rely on manual user study. To address this, we introduce a novel Human-Aligned benchmark for Text-guided Image Editing (HATIE). Providing a large-scale benchmark set covering a wide range of editing tasks, it allows reliable evaluation, not limited to specific easy-to-evaluate cases. Also, HATIE provides a fully-automated and omnidirectional evaluation pipeline. Particularly, we combine multiple scores measuring various aspects of editing so as to align with human perception. We empirically verify that the evaluation of HATIE is indeed human-aligned in various aspects, and provide benchmark results on several state-of-the-art models to provide deeper insights on their performance.
Self-supervised learning (SSL) is pervasively exploited in training high-quality upstream encoders with a large amount of unlabeled data. However, it is found to be susceptible to backdoor attacks merely via polluting a small portion of training data. The victim encoders associate triggered inputs with target embeddings, e.g., mapping a triggered cat image to an airplane embedding, such that the downstream tasks inherit unintended behaviors when the trigger is activated. Emerging backdoor attacks have shown great threats across different SSL paradigms such as contrastive learning and CLIP, yet limited research is devoted to defending against such attacks, and existing defenses fall short in detecting advanced stealthy backdoors. To address the limitations, we propose a novel detection mechanism, DeDe, which detects the activation of backdoor mappings caused by triggered inputs on victim encoders. Specifically, DeDe trains a decoder for any given SSL encoder using an auxiliary dataset (which can be out-of-distribution or even slightly poisoned), so that for any triggered input that misleads the encoder into the target embedding, the decoder generates an output image significantly different from the input. DeDe leverages the discrepancy between the input and the decoded output to identify potential backdoor misbehavior during inference. We empirically evaluate DeDe on both contrastive learning and CLIP models against various types of backdoor attacks. Our results demonstrate promising detection effectiveness over various advanced attacks and superior performance compared over state-of-the-art detection methods.
Recently, pre-trained text-to-image (T2I) models have been extensively adopted for real-world image restoration because of their powerful generative prior. However, controlling these large models for image restoration usually requires a large number of high-quality images and immense computational resources for training, which is costly and not privacy-friendly. In this paper, we find that the well-trained large T2I model (i.e., Flux) is able to produce a variety of high-quality images aligned with real-world distributions, offering an unlimited supply of training samples to mitigate the above issue. Specifically, we proposed a training data construction pipeline for image restoration, namely FluxGen, which includes unconditional image generation, image selection, and degraded image simulation. A novel light-weighted adapter (FluxIR) with squeeze-and-excitation layers is also carefully designed to control the large Diffusion Transformer (DiT)-based T2I model so that reasonable details can be restored. Experiments demonstrate that our proposed method enables the Flux model to adapt effectively to real-world image restoration tasks, achieving superior scores and visual quality on both synthetic and real-world degradation datasets - at only about 8.5% of the training cost compared to current approaches.
Easy-editable Image Vectorization with Multi-layer Multi-scale Distributed Visual Feature Embedding
PDF ↗Current parameterized image representations embed visual information along the semantic boundaries and struggle to express the internal detailed texture structures of image components, leading to a lack of content consistency after image editing and driving. To address these challenges, this work proposes a novel parameterized representation based on hierarchical image proxy geometry, utilizing multi-layer hierarchically interrelated proxy geometric control points to embed multi-scale long-range structures and fine-grained texture details. The proposed representation enables smoother and more continuous interpolation during image rendering and ensures high-quality consistency within image components during image editing. Additionally, under the layer-wise representation strategy based on semantic-aware image layer decomposition, we enable decoupled image shape/texture editing of the targets of interest within the image. Extensive experimental results on image vectorization and editing tasks demonstrate that our proposed method achieves high rendering accuracy of general images, including natural images, with a significantly higher image parameter compression ratio, facilitating user-friendly editing of image semantic components.
Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for Medical Image Segmentation
PDF ↗Despite domain generalization (DG) has significantly addressed the performance degradation of pre-trained models caused by domain shifts, it often falls short in real-world deployment. Test-time adaptation (TTA), which adjusts a learned model using unlabeled test data, presents a promising solution. However, most existing TTA methods struggle to deliver strong performance in medical image segmentation, primarily because they overlook the crucial prior knowledge inherent to medical images. To address this challenge, we incorporate morphological information and propose a framework based on multi-graph matching. Specifically, we introduce learnable universe embeddings that integrate morphological priors during multi-source training, along with novel unsupervised test-time paradigms for domain adaptation. This approach guarantees cycle-consistency in multi-matching while enabling the model to more effectively capture the invariant priors of unseen data, significantly mitigating the effects of domain shifts. Extensive experiments demonstrate that our method outperforms other state-of-the-art approaches on two medical image segmentation benchmarks for both multi-source and single-source domain generalization tasks. The source code is available at https://github.com/Yore0/TTDG-MGM.
Monocular 3D object detection (Mono3D) holds noteworthy promise for autonomous driving applications owing to the cost-effectiveness and rich visual context of monocular camera sensors. However, depth ambiguity poses a significant challenge, as it requires extracting precise 3D scene geometry from a single image, resulting in suboptimal performance when transferring knowledge from a LiDAR-based teacher model to a camera-based student model. To facilitate effective distillation, we introduce Monocular Teaching Assistant Knowledge Distillation (MonoTAKD), which proposes a camera-based teaching assistant (TA) model to transfer robust 3D visual knowledge to the student model, leveraging the smaller feature representation gap. Additionally, we define 3D spatial cues as residual features that capture the differences between the teacher and the TA models. We then leverage these cues to improve the student model's 3D perception capabilities. Experimental results show that our MonoTAKD achieves state-of-the-art performance on the KITTI3D dataset. Furthermore, we evaluate the performance on nuScenes and KITTI raw datasets to demonstrate the generalization of our model to multi-view 3D and unsupervised data settings. Our code is available at https://github.com/hoiliu-0801/MonoTAKD.
Enhanced Contrastive Learning with Multi-view Longitudinal Data for Chest X-ray Report Generation
PDF ↗Automated radiology report generation offers an effective solution to alleviate radiologists' workload. However, most existing methods focus primarily on single or fixed-view images to model current disease conditions, which limits diagnostic accuracy and overlooks disease progression. Although some approaches utilize longitudinal data to track disease progression, they still rely on single images to analyze current visits. To address these issues, we propose enhanced contrastive learning with Multi-view Longitudinal data to facilitate chest X-ray Report Generation, named MLRG. Specifically, we introduce a multi-view longitudinal contrastive learning method that integrates spatial information from current multi-view images and temporal information from longitudinal data. This method also utilizes the inherent spatiotemporal information of radiology reports to supervise the pre-training of visual and textual representations. Subsequently, we present a tokenized absence encoding technique to flexibly handle missing patient-specific prior knowledge, allowing the model to produce more accurate radiology reports based on available prior knowledge. Extensive experiments on MIMIC-CXR, MIMIC-ABN, and Two-view CXR datasets demonstrate that our MLRG outperforms recent state-of-the-art methods, achieving a 2.3% BLEU-4 improvement on MIMIC-CXR, a 5.5% F1 score improvement on MIMIC-ABN, and a 2.7% F1 RadGraph improvement on Two-view CXR.
We investigate the generalization boundaries of current Large Multimodal Models (LMMs) via comprehensive evaluation under out-of-distribution scenarios and domain-specific tasks. We evaluate their zero-shot generalization across synthetic images, real-world distributional shifts, and specialized datasets like medical and molecular imagery. Empirical results indicate that LMMs struggle with generalization beyond common training domains, limiting their direct application without adaptation. To understand the cause of unreliable performance, we analyze three hypotheses: semantic misinterpretation, visual feature extraction insufficiency, and mapping deficiency. Results identify mapping deficiency as the primary hurdle. To address this problem, we show that in-context learning (ICL) can significantly enhance LMMs' generalization, opening new avenues for overcoming generalization barriers. We further explore the robustness of ICL under distribution shifts and show its vulnerability to domain shifts, label shifts, and spurious correlation shifts between in-context examples and test data.
Given unlabelled datasets containing both old and new categories, generalized category discovery (GCD) aims to accurately discover new classes while correctly classifying old classes. Current GCD methods only use a single visual modality of information, resulting in poor classification of visually similar classes. As a different modality, text information can provide complementary discriminative information, which motivates us to introduce it into the GCD task. However, the lack of class names for unlabelled data makes it impractical to utilize text information. To tackle this challenging problem, in this paper, we propose a Text Embedding Synthesizer (TES) to generate pseudo text embeddings for unlabelled samples. Specifically, our TES leverages the property that CLIP can generate aligned vision-language features, converting visual embeddings into tokens of the CLIP's text encoder to generate pseudo text embeddings. Besides, we employ a dual-branch framework, through the joint learning and instance consistency of different modality branches, visual and semantic information mutually enhance each other, promoting the interaction and fusion of visual and text knowledge. Our method unlocks the multi-modal potentials of CLIP and outperforms the baseline methods by a large margin on all GCD benchmarks, achieving new state-of-the-art.
Semantic future prediction is important for autonomous systems navigating dynamic environments. This paper introduces FUTURIST, a method for multimodal future semantic prediction that uses a unified and efficient visual sequence transformer architecture. Our approach incorporates a multimodal masked visual modeling objective and a novel masking mechanism designed for multimodal training. This allows the model to effectively integrate visible information from various modalities, improving prediction accuracy. Additionally, we propose a VAE-free hierarchical tokenization process, which reduces computational complexity, streamlines the training pipeline, and enables end-to-end training with high-resolution, multimodal inputs. We validate FUTURIST on the Cityscapes dataset, demonstrating state-of-the-art performance in future semantic segmentation for both short- and mid-term forecasting. We provide the implementation code and model weights at https://github.com/Sta8is/FUTURIST.
The growing popularity of Contrastive Language-Image Pretraining (CLIP) has led to its widespread application in various visual downstream tasks. To enhance CLIP's effectiveness and versatility, efficient few-shot adaptation techniques have been widely adopted. Among these approaches, training-free methods, particularly caching methods exemplified by Tip-Adapter, have gained attention for their lightweight adaptation without the need for additional fine-tuning. In this paper, we revisit Tip-Adapter from a kernel perspective, showing that caching methods function as local adapters and are connected to a well-established kernel literature. Drawing on this insight, we offer a theoretical understanding of how these methods operate and suggest multiple avenues for enhancing the Tip-Adapter baseline. Notably, our analysis shows the importance of incorporating global information in local adapters. Therefore, we subsequently propose a global method that learns a proximal regularizer in a reproducing kernel Hilbert space (RKHS) using CLIP as a base learner. Our method, which we call ProKeR (Proximal Kernel ridge Regression), has a closed form solution and achieves state-of-the-art performances across 11 datasets in the standard few-shot adaptation benchmark. Code is available at https://ybendou.github.io/ProKeR/.
Prompts play a critical role in unleashing the power of language and vision foundation models for specific tasks. For the first time, we introduce prompting into depth foundation models, creating a new paradigm for metric depth estimation termed Prompt Depth Anything. Specifically, we use a low-cost LiDAR as the prompt to guide the Depth Anything model for accurate metric depth output, achieving up to 4K resolution. Our approach centers on a concise prompt fusion design that integrates the LiDAR at multiple scales within the depth decoder. To address training challenges posed by limited datasets containing both LiDAR depth and precise GT depth, we propose a scalable data pipeline that includes synthetic data LiDAR simulation and real data pseudo GT depth generation. Our approach sets new state-of-the-arts on the ARKitScenes and ScanNet++ datasets. Furthermore, we demonstrate that it benefits several downstream applications, including 3D reconstruction and generalized robotic grasping. Code will be released.
Neural rendering techniques have made substantial progress in generating photo-realistic 3D scenes. The latest 3D Gaussian Splatting technique has achieved high quality novel view synthesis as well as fast rendering speed. However, 3D Gaussians lack proficiency in defining accurate 3D geometric structures despite their explicit primitive representations. This is due to the fact that Gaussian's attributes are primarily tailored and fine-tuned for rendering diverse 2D images by their anisotropic nature. To pave the way for efficient 3D reconstruction, we present Spherical Gaussians, a simple and effective representation for 3D geometric boundaries, from which we can directly reconstruct 3D feature curves from a set of calibrated multi-view images. Spherical Gaussians is optimized from grid initialization with a view-based rendering loss, where a 2D edge map is rendered at a specific view and then compared to the ground-truth edge map extracted from the corresponding image, without the need for any 3D guidance or supervision. Given Spherical Gaussians serve as intermedia for the robust edge representation, we further introduce a novel optimization-based algorithm called SGCR to directly extract accurate parametric curves from aligned Spherical Gaussians. We demonstrate that SGCR outperforms existing state-of-the-art methods in 3D edge reconstruction while enjoying great efficiency.
Subject-driven text-to-image generation has demonstrated remarkable advancements in its ability to learn and capture characteristics of a subject using only a limited number of images. However, existing methods commonly rely on high-quality images for training and often struggle to generate reasonable images when the input images are blemished by artifacts. This is primarily attributed to the inadequate capability of current techniques in distinguishing subject-related features from disruptive artifacts. In this paper, we introduce ArtiFade to tackle this issue and successfully generate high-quality artifact-free images from blemished datasets. Specifically, ArtiFade exploits fine-tuning of a pre-trained text-to-image model, aiming to remove artifacts. The elimination of artifacts is achieved by utilizing a specialized dataset that encompasses both unblemished images and their corresponding blemished counterparts during fine-tuning. ArtiFade also ensures the preservation of the original generative capabilities inherent within the diffusion model, thereby enhancing the overall performance of subject-driven methods in generating high-quality and artifact-free images. We further devise evaluation benchmarks tailored for this task. Through extensive qualitative and quantitative experiments, we demonstrate the generalizability of ArtiFade in effective artifact removal under both in-distribution and out-of-distribution scenarios.
Large language models (LLMs) excel at retrieving information from lengthy text, but their vision-language counterparts (VLMs) face difficulties with hour-long videos, especially for temporal grounding. Specifically, these VLMs are constrained by frame limitations, often losing essential temporal details needed for accurate event localization in extended video content. We propose ReVisionLLM, a recursive vision-language model designed to locate events in hour-long videos. Inspired by human search strategies, our model initially targets broad segments of interest, progressively revising its focus to pinpoint exact temporal boundaries. Our model can seamlessly handle videos of vastly different lengths--from minutes to hours. We also introduce a hierarchical training strategy that starts with short clips to capture distinct events and progressively extends to longer videos. To our knowledge, ReVisionLLM is the first VLM capable of temporal grounding in hour-long videos, outperforming previous state-of-the-art methods across multiple datasets by a significant margin (e.g., +2.6% R1@0.1 on MAD). The code is available in the supplementary and will be released.
Human-motion video generation has been a challenging task, primarily due to the difficulty inherent in learning human body movements. While some approaches have attempted to drive human-centric video generation explicitly through pose control, these methods typically rely on poses derived from existing videos, thereby lacking flexibility. To address this, we propose HumanDreamer, a decoupled human video generation framework that first generates diverse poses from text prompts and then leverages these poses to generate human-motion videos. Specifically, we propose MotionVid, the largest dataset for human-motion pose generation. Based on the dataset, we present MotionDiT, which is trained to generate structured human-motion poses from text prompts. Besides, a novel LAMA loss is introduced, which together contribute to a significant improvement in FID by 62.4%, along with respective enhancements in R-precision for top1, top2, and top3 by 41.8%, 26.3%, and 18.3%, thereby advancing both the Text-to-Pose control accuracy and FID metrics. Our experiments across various Pose-to-Video baselines demonstrate that the poses generated by our method can produce diverse and high-quality human-motion videos. Furthermore, our model can facilitate other downstream tasks, such as pose sequence prediction and 2D-3D motion lifting.