Long-horizon tasks in robotic manipulation present significant challenges in reinforcement learning (RL) due to the difficulty of designing dense reward functions and effectively exploring the expansive state-action space. However, despite a lack of dense rewards, these tasks often have a multi-stage structure, which can be leveraged to decompose the overall objective into manageable subgoals. In this work, we propose DEMO3, a framework that exploits this structure for efficient learning from visual inputs. Specifically, our approach incorporates multi-stage dense reward learning, a bi-phasic training scheme, and world model learning into a carefully designed demonstration-augmented RL framework that strongly mitigates the challenge of exploration in long-horizon tasks. Our evaluations demonstrate that our method improves data-efficiency by an average of 40% and by 70% on particularly difficult tasks compared to state-of-the-art approaches. We validate this across 16 sparse-reward tasks spanning four domains, including challenging humanoid visual control tasks using as few as five demonstrations.
Visual Reinforcement Fine-Tuning (Visual-RFT)은 대규모 비전-언어 모델(Large Vision-Language Models, LVLMs)의 성능을 향상시키기 위해 설계된 혁신적인 방법론이다. 본 연구는 기존의 Reinforcement Fine-Tuning (RFT) 기법을 시각적 작업에 적용하여, 데이터가 부족한 상황에서도 효과적으로 모델을 파인튜닝할 수 있는 가능성을 제시한다. 기존의 RFT는 주로 언어 모델에 국한되어 있었으나, Visual-RFT는 멀티모달 도메인에서의 활용을 통해 그 적용 범위를 확장하였다.
Visual-RFT의 핵심은 LVLM을 사용하여 각 입력에 대해 여러 응답을 생성하고, 이를 기반으로 제안된 시각적 인식 검증 보상 함수를 통해 모델을 업데이트하는 것이다. 이 과정에서 Group Relative Policy Optimization (GRPO)와 같은 정책 최적화 알고리즘이 활용된다. 다양한 인식 작업에 맞춰 설계된 검증 가능한 보상 함수는 모델의 학습을 보다 효율적으로 이끌어준다. 예를 들어, 객체 탐지 작업에서는 Intersection over Union (IoU) 보상을 사용하여 성능을 극대화하였다.
실험 결과, Visual-RFT는 세밀한 이미지 분류, 몇 샷 객체 탐지, 추론 기반 작업 및 개방어휘 객체 탐지 벤치마크에서 Supervised Fine-Tuning (SFT) 기준선을 초과하는 뛰어난 성능을 보여주었다. 특히, 일회성 세밀한 이미지 분류에서는 기준선보다 24.3%의 정확도 향상을 달성하였으며, COCO 데이터셋의 두 샷 설정에서도 21.9 포인트, LVIS에서는 15.4 포인트의 성능 향상을 기록하였다. 이러한 성과는 Visual-RFT가 데이터 효율적이고 보상 기반의 접근 방식을 통해 도메인 특정 작업에 대한 추론과 적응력을 크게 향상시킬 수 있음을 시사한다.
본 연구는 LVLM의 파인튜닝에 있어 새로운 패러다임을 제시하며, 향후 다양한 멀티모달 작업에 대한 연구에 기여할 수 있는 기초 자료를 제공한다. 모든 훈련 코드와 데이터는 GitHub에 오픈 소스로 제공되어, 추가 연구와 발전을 촉진할 수 있는 기반을 마련하였다. Visual-RFT는 시각적 인식 작업에 대한 강화 학습의 가능성을 확장하며, 향후 연구에 있어 중요한 이정표가 될 것으로 기대된다.
Text-to-video generative models convert textual prompts into dynamic visual content, offering wide-ranging applications in film production, gaming, and education. However, their real-world performance often falls short of user expectations. One key reason is that these models have not been trained on videos related to some topics users want to create. In this paper, we propose VideoUFO, the first Video dataset specifically curated to align with Users' FOcus in real-world scenarios. Beyond this, our VideoUFO also features: (1) minimal (0.29%) overlap with existing video datasets, and (2) videos searched exclusively via YouTube's official API under the Creative Commons license. These two attributes provide future researchers with greater freedom to broaden their training sources. The VideoUFO comprises over 1.09 million video clips, each paired with both a brief and a detailed caption (description). Specifically, through clustering, we first identify 1,291 user-focused topics from the million-scale real text-to-video prompt dataset, VidProM. Then, we use these topics to retrieve videos from YouTube, split the retrieved videos into clips, and generate both brief and detailed captions for each clip. After verifying the clips with specified topics, we are left with about 1.09 million video clips. Our experiments reveal that (1) current 16 text-to-video models do not achieve consistent performance across all user-focused topics; and (2) a simple model trained on VideoUFO outperforms others on worst-performing topics. The dataset and code are publicly available at https://huggingface.co/datasets/WenhaoWang/VideoUFO and https://github.com/WangWenhao0716/BenchUFO under the CC BY 4.0 License.
Recent training-free layout-to-image diffusion models have demonstrated remarkable performance in generating high-quality images with controllable layouts. These models follow a one-stage framework: Encouraging the model to focus the attention map of each concept on its corresponding region by defining attention map-based losses. However, these models still struggle to accurately follow layouts with significant overlap, often leading to issues like attribute leakage and missing entities. In this paper, we propose ToLo, a two-stage, training-free layout-to-image generation framework for high-overlap layouts. Our framework consists of two stages: the aggregation stage and the separation stage, each with its own loss function based on the attention map. To provide a more effective evaluation, we partition the HRS dataset based on the Intersection over Union (IoU) of the input layouts, creating a new dataset for layout-to-image generation with varying levels of overlap. Through extensive experiments on this dataset, we demonstrate that ToLo significantly enhances the performance of existing methods when dealing with high-overlap layouts. Our code and dataset are available here: https://github.com/misaka12435/ToLo.
DUSt3R introduced a novel paradigm in geometric computer vision by proposing a model that can provide dense and unconstrained Stereo 3D Reconstruction of arbitrary image collections with no prior information about camera calibration nor viewpoint poses. Under the hood, however, DUSt3R processes image pairs, regressing local 3D reconstructions that need to be aligned in a global coordinate system. The number of pairs, growing quadratically, is an inherent limitation that becomes especially concerning for robust and fast optimization in the case of large image collections. In this paper, we propose an extension of DUSt3R from pairs to multiple views, that addresses all aforementioned concerns. Indeed, we propose a Multi-view Network for Stereo 3D Reconstruction, or MUSt3R, that modifies the DUSt3R architecture by making it symmetric and extending it to directly predict 3D structure for all views in a common coordinate frame. Second, we entail the model with a multi-layer memory mechanism which allows to reduce the computational complexity and to scale the reconstruction to large collections, inferring thousands of 3D pointmaps at high frame-rates with limited added complexity. The framework is designed to perform 3D reconstruction both offline and online, and hence can be seamlessly applied to SfM and visual SLAM scenarios showing state-of-the-art performance on various 3D downstream tasks, including uncalibrated Visual Odometry, relative camera pose, scale and focal estimation, 3D reconstruction and multi-view depth estimation.
Multimodal representation learning has demonstrated remarkable potential in enabling models to process and integrate diverse data modalities, such as text and images, for improved understanding and performance. While the medical domain can benefit significantly from this paradigm, the scarcity of paired multimodal data and reliance on proprietary or pretrained encoders pose significant challenges. In this work, we present a shared encoder framework for multimodal representation learning tailored to the medical domain. Our approach employs a single set of encoder parameters shared across modalities, augmented with learnable modality features. Empirical results demonstrate that our shared encoder idea achieves superior performance compared to separate modality-specific encoders, demonstrating improved generalization in data-constrained settings. Notably, the performance gains are more pronounced with fewer training examples, underscoring the efficiency of our shared encoder framework for real-world medical applications with limited data. Our code and experiment setup are available at https://github.com/VectorInstitute/shared_encoder.
The increasing prevalence of lumbar spinal canal stenosis has resulted in a surge of MRI (Magnetic Resonance Imaging), leading to labor-intensive interpretation and significant inter-reader variability, even among expert radiologists. This paper introduces a novel and efficient deep-learning framework that fully automates the grading of lumbar spinal canal stenosis. We demonstrate state-of-the-art performance in grading spinal canal stenosis on a dataset of 1,975 unique studies, each containing three distinct types of 3D cross-sectional spine images: Axial T2, Sagittal T1, and Sagittal T2/STIR. Employing a distinctive training strategy, our proposed multistage approach effectively integrates sagittal and axial images. This strategy employs a multi-view model with a sequence-based architecture, optimizing feature extraction and cross-view alignment to achieve an AUROC (Area Under the Receiver Operating Characteristic Curve) of 0.971 in spinal canal stenosis grading surpassing other state-of-the-art methods.
Soybean and cotton are major drivers of many countries' agricultural sectors, offering substantial economic returns but also facing persistent challenges from volunteer plants and weeds that hamper sustainable management. Effectively controlling volunteer plants and weeds demands advanced recognition strategies that can identify these amidst complex crop canopies. While deep learning methods have demonstrated promising results for leaf-level detection and segmentation, existing datasets often fail to capture the complexity of real-world agricultural fields. To address this, we collected 640 high-resolution images from a commercial farm spanning multiple growth stages, weed pressures, and lighting variations. Each image is annotated at the leaf-instance level, with 7,221 soybean and 5,190 cotton leaves labeled via bounding boxes and segmentation masks, capturing overlapping foliage, small leaf size, and morphological similarities. We validate this dataset using YOLOv11, demonstrating state-of-the-art performance in accurately identifying and segmenting overlapping foliage. Our publicly available dataset supports advanced applications such as selective herbicide spraying and pest monitoring and can foster more robust, data-driven strategies for soybean-cotton management.
Objective:This study introduces a residual error-shifting mechanism that drastically reduces sampling steps while preserving critical anatomical details, thus accelerating MRI reconstruction. Approach:We propose a novel diffusion-based SR framework called Res-SRDiff, which integrates residual error shifting into the forward diffusion process. This enables efficient HR image reconstruction by aligning the degraded HR and LR distributions.We evaluated Res-SRDiff on ultra-high-field brain T1 MP2RAGE maps and T2-weighted prostate images, comparing it with Bicubic, Pix2pix, CycleGAN, and a conventional denoising diffusion probabilistic model with vision transformer backbone (TM-DDPM), using quantitative metrics such as peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), gradient magnitude similarity deviation (GMSD), and learned perceptual image patch similarity (LPIPS). Main results: Res-SRDiff significantly outperformed all comparative methods in terms of PSNR, SSIM, and GMSD across both datasets, with statistically significant improvements (p-values<<0.05). The model achieved high-fidelity image restoration with only four sampling steps, drastically reducing computational time to under one second per slice, which is substantially faster than conventional TM-DDPM with around 20 seconds per slice. Qualitative analyses further demonstrated that Res-SRDiff effectively preserved fine anatomical details and lesion morphology in both brain and pelvic MRI images. Significance: Our findings show that Res-SRDiff is an efficient and accurate MRI SR method, markedly improving computational efficiency and image quality. Integrating residual error shifting into the diffusion process allows for rapid and robust HR image reconstruction, enhancing clinical MRI workflows and advancing medical imaging research. The source at:https://github.com/mosaf/Res-SRDiff
In recent years, the increasing popularity of Hi-DPI screens has driven a rising demand for high-resolution images. However, the limited computational power of edge devices poses a challenge in deploying complex super-resolution neural networks, highlighting the need for efficient methods. While prior works have made significant progress, they have not fully exploited pixel-level information. Moreover, their reliance on fixed sampling patterns limits both accuracy and the ability to capture fine details in low-resolution images. To address these challenges, we introduce two plug-and-play modules designed to capture and leverage pixel information effectively in Look-Up Table (LUT) based super-resolution networks. Our method introduces Automatic Sampling (AutoSample), a flexible LUT sampling approach where sampling weights are automatically learned during training to adapt to pixel variations and expand the receptive field without added inference cost. We also incorporate Adaptive Residual Learning (AdaRL) to enhance inter-layer connections, enabling detailed information flow and improving the network's ability to reconstruct fine details. Our method achieves significant performance improvements on both MuLUT and SPF-LUT while maintaining similar storage sizes. Specifically, for MuLUT, we achieve a PSNR improvement of approximately +0.20 dB improvement on average across five datasets. For SPF-LUT, with more than a 50% reduction in storage space and about a 2/3 reduction in inference time, our method still maintains performance comparable to the original. The code is available at https://github.com/SuperKenVery/AutoLUT.
Diffusion models have achieved great success in generating 2D images. However, the quality and generalizability of 3D content generation remain limited. State-of-the-art methods often require large-scale 3D assets for training, which are challenging to collect. In this work, we introduce Kiss3DGen (Keep It Simple and Straightforward in 3D Generation), an efficient framework for generating, editing, and enhancing 3D objects by repurposing a well-trained 2D image diffusion model for 3D generation. Specifically, we fine-tune a diffusion model to generate ''3D Bundle Image'', a tiled representation composed of multi-view images and their corresponding normal maps. The normal maps are then used to reconstruct a 3D mesh, and the multi-view images provide texture mapping, resulting in a complete 3D model. This simple method effectively transforms the 3D generation problem into a 2D image generation task, maximizing the utilization of knowledge in pretrained diffusion models. Furthermore, we demonstrate that our Kiss3DGen model is compatible with various diffusion model techniques, enabling advanced features such as 3D editing, mesh and texture enhancement, etc. Through extensive experiments, we demonstrate the effectiveness of our approach, showcasing its ability to produce high-quality 3D models efficiently.
Contrast-enhanced magnetic resonance imaging (CE-MRI) is crucial for tumor detection and diagnosis, but the use of gadolinium-based contrast agents (GBCAs) in clinical settings raises safety concerns due to potential health risks. To circumvent these issues while preserving diagnostic accuracy, we propose a novel Transformer with Localization Prompts (TLP) framework for synthesizing CE-MRI from non-contrast MR images. Our architecture introduces three key innovations: a hierarchical backbone that uses efficient Transformer to process multi-scale features; a multi-stage fusion system consisting of Local and Global Fusion modules that hierarchically integrate complementary information via spatial attention operations and cross-attention mechanisms, respectively; and a Fuzzy Prompt Generation (FPG) module that enhances the TLP model's generalization by emulating radiologists' manual annotation through stochastic feature perturbation. The framework uniquely enables interactive clinical integration by allowing radiologists to input diagnostic prompts during inference, synergizing artificial intelligence with medical expertise. This research establishes a new paradigm for contrast-free MRI synthesis while addressing critical clinical needs for safer diagnostic procedures. Codes are available at https://github.com/ChanghuiSu/TLP.
Diffusion models have emerged as powerful generative frameworks by progressively adding noise to data through a forward process and then reversing this process to generate realistic samples. While these models have achieved strong performance across various tasks and modalities, their application to temporal predictive learning remains underexplored. Existing approaches treat predictive learning as a conditional generation problem, but often fail to fully exploit the temporal dynamics inherent in the data, leading to challenges in generating temporally coherent sequences. To address this, we introduce Dynamical Diffusion (DyDiff), a theoretically sound framework that incorporates temporally aware forward and reverse processes. Dynamical Diffusion explicitly models temporal transitions at each diffusion step, establishing dependencies on preceding states to better capture temporal dynamics. Through the reparameterization trick, Dynamical Diffusion achieves efficient training and inference similar to any standard diffusion model. Extensive experiments across scientific spatiotemporal forecasting, video prediction, and time series forecasting demonstrate that Dynamical Diffusion consistently improves performance in temporal predictive tasks, filling a crucial gap in existing methodologies. Code is available at this repository: https://github.com/thuml/dynamical-diffusion.
Reducing radiation doses benefits patients, however, the resultant low-dose computed tomography (LDCT) images often suffer from clinically unacceptable noise and artifacts. While deep learning (DL) shows promise in LDCT reconstruction, it requires large-scale data collection from multiple clients, raising privacy concerns. Federated learning (FL) has been introduced to address these privacy concerns; however, current methods are typically tailored to specific scanning protocols, which limits their generalizability and makes them less effective for unseen protocols. To address these issues, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Since the noise distribution in LDCT data is closely tied to scanning protocols and anatomical structures being scanned, we design a dual-level physics-informed way to address these challenges. Specifically, we incorporate physical and anatomical prompts into our physics-informed hypernetworks to capture scanning- and anatomy-specific information, enabling dual-level physics-driven personalization of imaging features. These prompts are derived from the scanning protocol and the radiology report generated by a medical large language model (MLLM), respectively. Subsequently, client-specific decoders project these dual-level personalized imaging features back into the image domain. Besides, to tackle the challenge of unseen data, we introduce a novel protocol vector-quantization strategy (PVQS), which ensures consistent performance across new clients by quantifying the unseen scanning code as one of the codes in the scanning codebook. Extensive experimental results demonstrate the superior performance of SCAN-PhysFed on public datasets.
End-to-end vision-based imitation learning has demonstrated promising results in autonomous driving by learning control commands directly from expert demonstrations. However, traditional approaches rely on either regressionbased models, which provide precise control but lack confidence estimation, or classification-based models, which offer confidence scores but suffer from reduced precision due to discretization. This limitation makes it challenging to quantify the reliability of predicted actions and apply corrections when necessary. In this work, we introduce a dual-head neural network architecture that integrates both regression and classification heads to improve decision reliability in imitation learning. The regression head predicts continuous driving actions, while the classification head estimates confidence, enabling a correction mechanism that adjusts actions in low-confidence scenarios, enhancing driving stability. We evaluate our approach in a closed-loop setting within the CARLA simulator, demonstrating its ability to detect uncertain actions, estimate confidence, and apply real-time corrections. Experimental results show that our method reduces lane deviation and improves trajectory accuracy by up to 50%, outperforming conventional regression-only models. These findings highlight the potential of classification-guided confidence estimation in enhancing the robustness of vision-based imitation learning for autonomous driving. The source code is available at https://github.com/ElaheDlv/Confidence_Aware_IL.
Diffusion models transform an unknown data distribution into a Gaussian prior by progressively adding noise until the data become indistinguishable from pure noise. This stochastic process traces a path in probability space, evolving from the original data distribution (considered as a Gaussian with near-zero variance) to an isotropic Gaussian. The denoiser then learns to reverse this process, generating high-quality samples from random Gaussian noise. However, standard diffusion models, such as the Denoising Diffusion Probabilistic Model (DDPM), do not ensure a geodesic (i.e., shortest) path in probability space. This inefficiency necessitates the use of many intermediate time steps, leading to high computational costs in training and sampling. To address this limitation, we propose the Geodesic Diffusion Model (GDM), which defines a geodesic path under the Fisher-Rao metric with a variance-exploding noise scheduler. This formulation transforms the data distribution into a Gaussian prior with minimal energy, significantly improving the efficiency of diffusion models. We trained GDM by continuously sampling time steps from 0 to 1 and using as few as 15 evenly spaced time steps for model sampling. We evaluated GDM on two medical image-to-image generation tasks: CT image denoising and MRI image super-resolution. Experimental results show that GDM achieved state-of-the-art performance while reducing training time by a 50-fold compared to DDPM and 10-fold compared to Fast-DDPM, with 66 times faster sampling than DDPM and a similar sampling speed to Fast-DDPM. These efficiency gains enable rapid model exploration and real-time clinical applications. Our code is publicly available at: https://github.com/mirthAI/GDM-VE.
Fully-supervised lesion recognition methods in medical imaging face challenges due to the reliance on large annotated datasets, which are expensive and difficult to collect. To address this, synthetic lesion generation has become a promising approach. However, existing models struggle with scalability, fine-grained control over lesion attributes, and the generation of complex structures. We propose LesionDiffusion, a text-controllable lesion synthesis framework for 3D CT imaging that generates both lesions and corresponding masks. By utilizing a structured lesion report template, our model provides greater control over lesion attributes and supports a wider variety of lesion types. We introduce a dataset of 1,505 annotated CT scans with paired lesion masks and structured reports, covering 14 lesion types across 8 organs. LesionDiffusion consists of two components: a lesion mask synthesis network (LMNet) and a lesion inpainting network (LINet), both guided by lesion attributes and image features. Extensive experiments demonstrate that LesionDiffusion significantly improves segmentation performance, with strong generalization to unseen lesion types and organs, outperforming current state-of-the-art models. Code is available at https://github.com/HengruiTianSJTU/LesionDiffusion.
Multimodal instruction tuning has proven to be an effective strategy for achieving zero-shot generalization by fine-tuning pre-trained Large Multimodal Models (LMMs) with instruction-following data. However, as the scale of LMMs continues to grow, fully fine-tuning these models has become highly parameter-intensive. Although Parameter-Efficient Fine-Tuning (PEFT) methods have been introduced to reduce the number of tunable parameters, a significant performance gap remains compared to full fine-tuning. Furthermore, existing PEFT approaches are often highly parameterized, making them difficult to interpret and control. In light of this, we introduce Multimodal Representation Tuning (MRT), a novel approach that focuses on directly editing semantically rich multimodal representations to achieve strong performance and provide intuitive control over LMMs. Empirical results show that our method surpasses current state-of-the-art baselines with significant performance gains (e.g., 1580.40 MME score) while requiring substantially fewer tunable parameters (e.g., 0.03% parameters). Additionally, we conduct experiments on editing instrumental tokens within multimodal representations, demonstrating that direct manipulation of these representations enables simple yet effective control over network behavior.
Recently, diffusion- and flow-based generative models of protein structures have emerged as a powerful tool for de novo protein design. Here, we develop Proteina, a new large-scale flow-based protein backbone generator that utilizes hierarchical fold class labels for conditioning and relies on a tailored scalable transformer architecture with up to 5x as many parameters as previous models. To meaningfully quantify performance, we introduce a new set of metrics that directly measure the distributional similarity of generated proteins with reference sets, complementing existing metrics. We further explore scaling training data to millions of synthetic protein structures and explore improved training and sampling recipes adapted to protein backbone generation. This includes fine-tuning strategies like LoRA for protein backbones, new guidance methods like classifier-free guidance and autoguidance for protein backbones, and new adjusted training objectives. Proteina achieves state-of-the-art performance on de novo protein backbone design and produces diverse and designable proteins at unprecedented length, up to 800 residues. The hierarchical conditioning offers novel control, enabling high-level secondary-structure guidance as well as low-level fold-specific generation.
Transformer-based deep learning models have achieved state-of-the-art performance across numerous language and vision tasks. While the self-attention mechanism, a core component of transformers, has proven capable of handling complex data patterns, it has been observed that the representational capacity of the attention matrix degrades significantly across transformer layers, thereby hurting its overall performance. In this work, we leverage the connection between self-attention computations and low-pass non-local means (NLM) smoothing filters and propose the Twicing Attention, a novel attention mechanism that uses kernel twicing procedure in nonparametric regression to alleviate the low-pass behavior of associated NLM smoothing with compelling theoretical guarantees and enhanced adversarial robustness. This approach enables the extraction and reuse of meaningful information retained in the residuals following the imperfect smoothing operation at each layer. Our proposed method offers two key advantages over standard self-attention: 1) a provably slower decay of representational capacity and 2) improved robustness and accuracy across various data modalities and tasks. We empirically demonstrate the performance gains of our model over baseline transformers on multiple tasks and benchmarks, including image classification and language modeling, on both clean and corrupted data.