Vision-based Reinforcement Learning (VRL) attempts to establish associations between visual inputs and optimal actions through interactions with the environment. Given the high-dimensional and complex nature of visual data, it becomes essential to learn policy upon high-quality state representation. To this end, existing VRL methods primarily rely on interaction-collected data, combined with self-supervised auxiliary tasks. However, two key challenges remain: limited data samples and a lack of task-relevant semantic constraints. To tackle this, we propose DGC, a method that distills guidance from Visual Language Models (VLMs) alongside self-supervised learning into a compact VRL agent. Notably, we leverage the state representation capabilities of VLMs, rather than their decision-making abilities. Within DGC, a novel prompting-reasoning pipeline is designed to convert historical observations and actions into usable supervision signals, enabling semantic understanding within the compact visual encoder. By leveraging these distilled semantic representations, the VRL agent achieves significant improvements in the sample efficiency. Extensive experiments on the Carla benchmark demonstrate our state-of-the-art performance.
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
ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge
PDF ↗Edge-based computer vision models running on compact, resource-limited devices benefit greatly from using unprocessed, detail-rich RAW sensor data instead of processed RGB images. Training these models, however, necessitates large labeled RAW datasets, which are costly and often impractical to obtain. Thus, converting existing labeled RGB datasets into sensor-specific RAW images becomes crucial for effective model training. In this paper, we introduce ReRAW, an RGB-to-RAW conversion model that achieves state-of-the-art reconstruction performance across five diverse RAW datasets. This is accomplished through ReRAW's novel multi-head architecture predicting RAW image candidates in gamma space. The performance is further boosted by a stratified sampling-based training data selection heuristic, which helps the model better reconstruct brighter RAW pixels. We finally demonstrate that pretraining compact models on a combination of high-quality synthetic RAW datasets (such as generated by ReRAW) and ground-truth RAW images for downstream tasks like object detection, outperforms both standard RGB pipelines, and RAW fine-tuning of RGB-pretrained models for the same task.
Multi-modal large language models (MLLMs) extend large language models (LLMs) to process multi-modal information, enabling them to generate responses to image-text inputs. MLLMs have been incorporated into diverse multi-modal applications, such as autonomous driving and medical diagnosis, via plug-and-play without fine-tuning. This deployment paradigm increases the vulnerability of MLLMs to backdoor attacks. However, existing backdoor attacks against MLLMs achieve limited effectiveness and stealthiness. In this work, we propose BadToken, the first token-level backdoor attack to MLLMs. BadToken introduces two novel backdoor behaviors: Token-substitution and Token-addition, which enable flexible and stealthy attacks by making token-level modifications to the original output for backdoored inputs. We formulate a general optimization problem that considers the two backdoor behaviors to maximize the attack effectiveness. We evaluate BadToken on two open-source MLLMs and various tasks. Our results show that our attack maintains the model's utility while achieving high attack success rates and stealthiness. We also show the real-world threats of BadToken in two scenarios, i.e., autonomous driving and medical diagnosis. Furthermore, we consider defenses including fine-tuning and input purification. Our results highlight the threat of our attack.
In-context learning (ICL), a predominant trend in instruction learning, aims at enhancing the performance of large language models by providing clear task guidance and examples, improving their capability in task understanding and execution. This paper investigates ICL on Large Vision-Language Models (LVLMs) and explores the policies of multi-modal demonstration selection. Existing research efforts in ICL face significant challenges: First, they rely on pre-defined demonstrations or heuristic selecting strategies based on human intuition, which are usually inadequate for covering diverse task requirements, leading to sub-optimal solutions; Second, individually selecting each demonstration fails in modeling the interactions between them, resulting in information redundancy. Unlike these prevailing efforts, we propose a new exploration-exploitation reinforcement learning framework, which explores policies to fuse multi-modal information and adaptively select adequate demonstrations as an integrated whole. The framework allows LVLMs to optimize themselves by continually refining their demonstrations through self-exploration, enabling the ability to autonomously identify and generate the most effective selection policies for in-context learning. Experimental results verify that our approach achieves significant performance improvements on four Visual Question-Answering (VQA) datasets, demonstrating its effectiveness in enhancing the generalization capability of few-shot LVLMs.
Reconstructing complex reflections in real-world scenes from 2D images is essential for achieving photorealistic novel view synthesis. Existing methods that utilize environment maps to model reflections from distant lighting often struggle with high-frequency reflection details and fail to account for near-field reflections. In this work, we introduce EnvGS, a novel approach that employs a set of Gaussian primitives as an explicit 3D representation for capturing reflections of environments. These environment Gaussian primitives are incorporated with base Gaussian primitives to model the appearance of the whole scene. To efficiently render these environment Gaussian primitives, we developed a ray-tracing-based renderer that leverages the GPU's RT core for fast rendering. This allows us to jointly optimize our model for high-quality reconstruction while maintaining real-time rendering speeds. Results from multiple real-world and synthetic datasets demonstrate that our method produces significantly more detailed reflections, achieving the best rendering quality in real-time novel view synthesis. The code is available at https://zju3dv.github.io/envgs.
Diffusion models can generate realistic and diverse images, potentially facilitating data availability for data-intensive perception tasks. However, leveraging these models to boost performance on downstream tasks with synthetic data poses several challenges, including aligning with real data distribution, scaling synthetic sample volumes, and ensuring their quality. To bridge these gaps, we present Auto Cherry-Picker (ACP), a novel framework that generates high-quality cross-modality training samples at scale to augment perception and multi-modal training. ACP first uses LLMs to sample descriptions and layouts based on object combinations from real data priors, eliminating the need for ground truth image captions or annotations. Next, we use an off-the-shelf controllable diffusion model to generate multiple images. Then, the generated data are refined using a comprehensively designed metric, Composite Layout and Image Score (CLIS), to ensure quality. Our customized synthetic high-quality samples boost performance in various scenarios, especially in addressing challenges associated with long-tailed distribution and imbalanced datasets. Experiment results on downstream tasks demonstrate that ACP can significantly improve the performance of existing models. In addition, we find a positive correlation between CLIS and performance gains in downstream tasks. This finding shows the potential for evaluation metrics as the role for various visual perception and MLLM tasks.
Multi-view representation learning integrates multiple observable views of an entity into a unified representation to facilitate downstream tasks. Current methods predominantly focus on distinguishing compatible components across views, followed by a single-step parallel fusion process. However, this parallel fusion is static in essence, overlooking potential conflicts among views and compromising representation ability. To address this issue, this paper proposes a novel Sequential fusion framework for Multi-view Representation Learning, termed SeqMvRL. Specifically, we model multi-view fusion as a sequential decision-making problem and construct a pairwise integrator (PI) and a next-view selector (NVS), which represent the environment and agent in reinforcement learning, respectively. PI merges the current fused feature with the selected view, while NVS is introduced to determine which view to fuse subsequently. By adaptively selecting the next optimal view for fusion based on the current fusion state, SeqMvRL thereby effectively reduces conflicts and enhances unified representation quality. Additionally, an elaborate novel reward function encourages the model to prioritize views that enhance the discriminability of the fused features. Experimental results demonstrate that SeqMvRL outperforms parallel fusion approaches in classification and clustering tasks.
Manga, a popular form of multimodal artwork, has traditionally been overlooked in deep learning advancements due to the absence of a robust dataset and comprehensive annotation. Manga segmentation is the key to the digital migration of manga. There exists a significant domain gap between the manga and the natural images, that fails most existing learning-based methods. To address this gap, we introduce an augmented segmentation annotation for the Manga109 dataset, a collection of 109 manga volumes, that offers intricate artworks in a rich variety of styles. We introduce a detailed annotation that extends beyond the original simple bounding boxes to the segmentation masks with pixel-level precision. It provides object category, location, and instance information that can be used for semantic segmentation and instance segmentation. We also provide a comprehensive analysis of our annotation dataset from various aspects. We further measure the improvement of the state-of-the-art segmentation model after training it with our augmented dataset. The benefits of this augmented dataset are profound, with the potential to significantly enhance manga analysis algorithms and catalyze the novel development in digital art processing and cultural analytics. This annotation, named MangaSeg, is publicly available at https://huggingface.co/datasets/MS92/MangaSegmentation.
Image generation today can produce somewhat realistic images from text prompts. However, if one asks the generator to synthesize a specific camera setting such as creating different fields of view using a 24mm lens versus a 70mm lens, the generator will not be able to interpret and generate scene-consistent images. This limitation not only hinders the adoption of generative tools in professional photography but also highlights the broader challenge of aligning data-driven models with real-world physical settings. In this paper, we introduce Generative Photography, a framework that allows controlling camera intrinsic settings during content generation. The core innovation of this work are the concepts of Dimensionality Lifting and Differential Camera Intrinsics Learning, enabling smooth and consistent transitions across different camera settings. Experimental results show that our method produces significantly more scene-consistent photorealistic images than state-of-the-art models such as Stable Diffusion 3 and FLUX. Our code and additional results are available at https://generative-photography.github.io/project.
Hallo3: Highly Dynamic and Realistic Portrait Image Animation with Video Diffusion Transformer
PDF ↗Existing methodologies for animating portrait images face significant challenges, particularly in handling non-frontal perspectives, rendering dynamic objects around the portrait, and generating immersive, realistic backgrounds. In this paper, we introduce the first application of a pretrained transformer-based video generative model that demonstrates strong generalization capabilities and generates highly dynamic, realistic videos for portrait animation, effectively addressing these challenges. The adoption of a new video backbone model makes previous U-Net-based methods for identity maintenance, audio conditioning, and video extrapolation inapplicable. To address this limitation, we design an identity reference network consisting of a causal 3D VAE combined with a stacked series of transformer layers, ensuring consistent facial identity across video sequences. Additionally, we investigate various speech audio conditioning and motion frame mechanisms to enable the generation of continuous video driven by speech audio. Our method is validated through experiments on benchmark and newly proposed wild datasets, demonstrating substantial improvements over prior methods in generating realistic portraits characterized by diverse orientations within dynamic and immersive scenes.
High-resolution image inputs allow Large Vision-Language Models (LVLMs) to capture finer visual details, improving comprehension. However, the increased training and computational costs associated with such inputs pose significant challenges. A common approach to mitigate these costs involves slicing the input into uniform patches using sliding windows, each aligned with the vision encoder's input size. While efficient, this method fragments the input, disrupting the continuity of context, which negatively impacts cross-patch perception tasks. To address these limitations, we propose HiRes-LLaVA, a novel framework designed to efficiently process high-resolution inputs of any size without altering the original contextual and geometric information. HiRes-LLaVA introduces two key components: (i) a SliceRestore Adapter (SRA) that reconstructs sliced patches into their original form, enabling efficient extraction of both global and local features through down-up-sampling and convolutional layers, and (ii) a Self-Mining Sampler (SMS) that compresses visual tokens based on internal relationships, preserving original context and positional information while reducing training overhead. To assess the ability of handling context fragmentation, we construct a new benchmark, EntityGrid-QA, consisting of edge-related tasks. Extensive experiments demonstrate the superiority of HiRes-LLaVA on both existing public benchmarks and EntityGrid-QA. For example, with SRA, our method achieves a performance improvement of 12% over state-of-the-art LVLMs in addressing fragmentation issues. Additionally, our SMS outperforms other visual token downsamplers, while offering high data efficiency.
Recent advances in diffusion models have positioned them as powerful generative frameworks for high-resolution image synthesis across diverse domains. The emerging h-space within these models, defined by bottleneck activations in the denoiser, offers promising pathways for semantic image editing similar to GAN latent spaces. However, as demand grows for content erasure and concept removal, privacy concerns highlight the need for identity disentanglement in the latent space of diffusion models. The high dimensional latent space poses challenges for identity removal, as traversing with random or orthogonal directions often leads to semantically unvalidated regions, resulting in unrealistic outputs.To address these issues, we propose Black Hole Driven Identity Absorption (BIA), a novel approach for identity erasure within the latent space of diffusion models. BIA uses a black hole metaphor, where the latent region representing a specified identity acts as an attractor, drawing in nearby latent points of surrounding identities to wrap the black hole. Instead of relying on random traversals for optimization, BIA employs an identity absorption mechanism by attracting and wrapping nearby validated latent points associated with other identities to achieve a vanishing effect for specified identity. Our method effectively prevents the generation of a specified identity while preserving other attributes, as validated by improved scores on identity similarity SID, FID metrics, qualitative evaluations, and user studies as compared to SOTA.
Recent multi-teacher distillation methods have unified the encoders of multiple foundation models into a single encoder, achieving competitive performance on core vision tasks like classification, segmentation, and depth estimation. This led us to ask: Could similar success be achieved when the pool of teachers also includes vision models specialized in diverse tasks across both 2D and 3D perception? In this paper, we define and investigate the problem of heterogeneous teacher distillation, or co-distillation, a challenging multi-teacher distillation scenario where teacher models vary significantly in both (a) their design objectives and (b) the data they were trained on. We explore data-sharing strategies and teacher-specific encoding, and introduce DUNE, a single encoder excelling in 2D vision, 3D understanding, and 3D human perception. Our model achieves performance comparable to that of its larger teachers, sometimes even outperforming them, on their respective tasks. Notably, DUNE surpasses MASt3R in Map-free Visual Relocalization with a much smaller encoder.
STiL: Semi-supervised Tabular-Image Learning for Comprehensive Task-Relevant Information Exploration in Multimodal Classification
PDF ↗Multimodal image-tabular learning is gaining attention, yet it faces challenges due to limited labeled data. While earlier work has applied self-supervised learning (SSL) to unlabeled data, its task-agnostic nature often results in learning suboptimal features for downstream tasks. Semi-supervised learning (SemiSL), which combines labeled and unlabeled data, offers a promising solution. However, existing multimodal SemiSL methods typically focus on unimodal or modality-shared features, ignoring valuable task-relevant modality-specific information, leading to a Modality Information Gap. In this paper, we propose STiL, a novel SemiSL tabular-image framework that addresses this gap by comprehensively exploring task-relevant information. STiL features a new disentangled contrastive consistency module to learn cross-modal invariant representations of shared information while retaining modality-specific information via disentanglement. We also propose a novel consensus-guided pseudo-labeling strategy to generate reliable pseudo-labels based on classifier consensus, along with a new prototype-guided label smoothing technique to refine pseudo-label quality with prototype embeddings, thereby enhancing task-relevant information learning in unlabeled data. Experiments on natural and medical image datasets show that STiL outperforms the state-of-the-art supervised/SSL/SemiSL image/multimodal approaches. Our code is available at https://github.com/siyi-wind/STiL.
DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation
PDF ↗Semi-supervised learning in medical image segmentation leverages unlabeled data to reduce annotation burdens through consistency learning. However, current methods struggle with class imbalance and high uncertainty from pathology variations, leading to inaccurate segmentation in 3D medical images. To address these challenges, we present DyCON, a Dynamic Uncertainty-aware Consistency and Contrastive Learning framework that enhances the generalization of consistency methods with two complementary losses: Uncertainty-aware Consistency Loss (UnCL) and Focal Entropy-aware Contrastive Loss (FeCL). UnCL enforces global consistency by dynamically weighting the contribution of each voxel to the consistency loss based on its uncertainty, preserving high-uncertainty regions instead of filtering them out. Initially, UnCL prioritizes learning from uncertain voxels with lower penalties, encouraging the model to explore challenging regions. As training progress, the penalty shift towards confident voxels to refine predictions and ensure global consistency. Meanwhile, FeCL enhances local feature discrimination in imbalanced regions by introducing dual focal mechanisms and adaptive confidence adjustments into the contrastive principle. These mechanisms jointly prioritizes hard positives and negatives while focusing on uncertain sample pairs, effectively capturing subtle lesion variations under class imbalance. Extensive evaluations on four diverse medical image segmentation datasets (ISLES'22, BraTS'19, LA, Pancreas) show DyCON's superior performance against SOTA methods(\href https://dycon25.github.io/) Project Page: https://dycon25.github.io/ .
Light Transport-aware Diffusion Posterior Sampling for Single-View Reconstruction of 3D Volumes
PDF ↗We introduce a single-view reconstruction technique of volumetric fields in which multiple light scattering effects are omnipresent, such as in clouds. We model the unknown distribution of volumetric fields using an unconditional diffusion model trained on a novel benchmark dataset comprising 1,000 synthetically simulated volumetric density fields. The neural diffusion model is trained on the latent codes of a novel, diffusion-friendly, monoplanar representation. The generative model is used to incorporate a tailored parametric diffusion posterior sampling technique into different reconstruction tasks. A physically-based differentiable volume renderer is employed to provide gradients with respect to light transport in the latent space. This stands in contrast to classic NeRF approaches and makes the reconstructions better aligned with observed data. Through various experiments, we demonstrate single-view reconstruction of volumetric clouds at a previously unattainable quality.
Leveraging the large generative prior of the flow transformer for tuning-free image editing requires authentic inversion to project the image into the model's domain and a flexible invariance control mechanism to preserve non-target contents. However, the prevailing diffusion inversion performs deficiently in flow-based models, and the invariance control cannot reconcile diverse rigid and non-rigid editing tasks. To address these, we systematically analyze the inversion and invariance control based on the flow transformer. Specifically, we unveil that the Euler inversion shares a similar structure to DDIM yet is more susceptible to the approximation error. Thus, we propose a two-stage inversion to first refine the velocity estimation and then compensate for the leftover error, which pivots closely to the model prior and benefits editing. Meanwhile, we propose the invariance control that manipulates the text features within the adaptive layer normalization, connecting the changes in the text prompt to image semantics. This mechanism can simultaneously preserve the non-target contents while allowing rigid and non-rigid manipulation, enabling a wide range of editing types. Experiments on various scenarios demonstrate that our framework achieves flexible and accurate editing, unlocking the potential of the flow transformer for versatile image editing.
Active learning (AL) and semi-supervised learning (SSL) both aim to reduce annotation costs: AL selectively annotates high-value samples from the unlabeled data, while SSL leverages abundant unlabeled data to improve model performance. Although these two appear intuitively compatible, directly combining them remains challenging due to fundamental differences in their frameworks. Current semi-supervised active learning (SSAL) methods often lack theoretical foundations and often design AL strategies tailored to a specific SSL algorithm rather than genuinely integrating the two fields.In this paper, we incorporate AL objectives into the overall risk formulation within the mainstream pseudo-label-based SSL framework, clarifying key differences between SSAL and traditional AL scenarios. To bridge these gaps, we propose a feature re-alignment module that aligns the features of unlabeled data under different augmentations by leveraging clustering and consistency constraints. Experimental results demonstrate that our module enables flexible combinations of SOTA methods from both AL and SSL, yielding more efficient algorithm performance.
UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation
PDF ↗Recently, text-to-image generation models have achieved remarkable advancements, particularly with diffusion models facilitating high-quality image synthesis from textual descriptions. However, these models often struggle with achieving precise control over pixel-level layouts, object appearances, and global styles when using text prompts alone. To mitigate this issue, previous works introduce conditional images as auxiliary inputs for image generation, enhancing control but typically necessitating specialized models tailored to different types of reference inputs. In this paper, we explore a new approach to unify controllable generation within a single framework. Specifically, we propose the unified image-instruction adapter (UNIC-Adapter) built on the Multi-Modal-Diffusion Transformer architecture, to enable flexible and controllable generation across diverse conditions without the need for multiple specialized models. Our UNIC-Adapter effectively extracts multi-modal instruction information by incorporating both conditional images and task instructions, injecting this information into the image generation process through a cross-attention mechanism enhanced by Rotary Position Embedding. Experimental results across a variety of tasks, including pixel-level spatial control, subject-driven image generation, and style-image-based image synthesis, demonstrate the effectiveness of our UNIC-Adapter in unified controllable image generation.
In this work, we propose an extreme compression technique for Large Multimodal Models (LMMs). While previous studies have explored quantization as an efficient post-training compression method for Large Language Models (LLMs), low-bit compression for multimodal models remains under-explored. The redundant nature of inputs in multimodal models results in a highly sparse attention matrix. We theoretically and experimentally demonstrate that the attention matrix's sparsity bounds the compression error of the Query and Key weight matrices. Based on this, we introduce CASP, a model compression technique for LMMs. Our approach performs a data-aware low-rank decomposition on the Query and Key weight matrix, followed by quantization across all layers based on an optimal bit allocation process. CASP is compatible with any quantization technique and enhances state-of-the-art 2-bit quantization methods (AQLM and QuIP#) by an average of 21% on image- and video-language benchmarks. The code is provided in the supplementary materials.