These 851 CVPR 2025 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every CVPR 2025 paper, free trial on arXivSub.
2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification
Jingwei Zhang (Stony Brook University), Mahdi S. Hosseini (Concordia University)
CodeClassificationRepresentation LearningImage
π― What it does: This paper proposes 2DMamba, which utilizes a 2D selective state space model for efficient modeling of large-scale images and achieves spatial continuity in multi-instance learning.
π― What it does: This study investigates how to extend 3D Gaussian Splatting to support arbitrary nonlinear camera projections and secondary rays, achieving efficient real-time rendering through the Unscented Transform, compatible with rolling shutter, distorted cameras, and phenomena such as reflection and refraction.
π― What it does: This paper proposes a Multi-Cognitive Visual Adapter (Mona) tuning method for parameter-efficient fine-tuning of visual tasks while maintaining the advantages of pre-trained models.
π― What it does: A systematic evaluation of the performance of visual models with different pre-training methods and data sources in robotic learning tasks (manipulation and perception) is conducted, and SlotMIM is proposed to learn more object-centric representations on non-single-object (NOC) data.
π― What it does: A new semantic segmentation anomaly detection dataset, ISSU, is proposed, and various existing methods are benchmarked on this dataset.
π― What it does: This paper proposes Flag Decomposition (FD), a matrix decomposition method that preserves hierarchical structures, capable of mapping hierarchical data to the flag manifold for reconstruction, clustering, and few-shot learning.
A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening
Jie Huang (University of Electronic Science and Technology of China), Liangjian Deng (University of Electronic Science and Technology of China)
CodeRestorationSuper ResolutionImage
π― What it does: An Adaptive Dual-layer Weighting Mechanism (ADWM) is proposed, which adjusts feature heterogeneity and redundancy through Covariance-weighted (CACW) to achieve remote sensing panchromatic fusion.
π― What it does: A new framework called hubREP is proposed for graph-based multi-view clustering, aimed at addressing the hubness problem in high-dimensional embeddings to improve clustering performance.
A Polarization-Aided Transformer for Image Deblurring via Motion Vector Decomposition
Duosheng Chen (Nankai University), Jufeng Yang (Nankai University)
CodeRestorationTransformerOptical FlowImage
π― What it does: A motion decomposition Transformer (MDT) based on polar coordinates is proposed, which achieves deblurring by separating the translational and rotational motions of image blur.
π― What it does: A self-supervised regularization strategy (EQ-Reg) is proposed, which can achieve rotation equivariance in ordinary CNNs while maintaining high representation accuracy.
π― What it does: A selective re-learning hyperspectral fusion network, SRLF-Net, is proposed for the fusion of low-resolution hyperspectral images and multispectral images.
π― What it does: In the semi-supervised medical image segmentation task, a Semantic Knowledge Complementary Decoupling Framework (SKCDF) is proposed, which trains the encoder, labeled decoder, and unlabeled decoder separately. It utilizes labeled data to guide pseudo-label generation and enriches labeled features with unlabeled data, while introducing an auxiliary balanced segmentation head to enhance the performance of minority classes.
π― What it does: A random distribution normalization method based on input-level data augmentation (FedRDN) is proposed in federated learning, which alleviates feature distribution shift and enhances model generalization by randomly injecting global statistical information into local samples.
A Stitch in Time Saves Nine: Small VLM is a Precise Guidance for Accelerating Large VLMs
Wangbo Zhao (National University of Singapore), Yang You (National University of Singapore)
CodeComputational EfficiencyKnowledge DistillationTransformerVision Language ModelMultimodality
π― What it does: A training-independent method called SGL is proposed, which utilizes the full-layer attention information of a small-scale VLM to guide the visual token pruning of a large-scale VLM, and enhances inference efficiency through early exit of the small VLM when necessary.
A Tale of Two Classes: Adapting Supervised Contrastive Learning to Binary Imbalanced Datasets
David Mildenberger (Technical University of Munich), Martin J. Menten (Technical University of Munich)
CodeClassificationRepresentation LearningContrastive LearningImageBiomedical Data
π― What it does: This study investigates the performance of supervised contrastive learning on imbalanced binary classification datasets and proposes two improvement methods.
π― What it does: A latent space diffusion model LASB based on linear Schrodinger Bridge is proposed for unsupervised anomaly detection and localization.
π― What it does: An auxiliary dual-modal cross-domain Transformer named ABC-Former is proposed to improve the white balance correction of sRGB images.
π― What it does: An adaptive cross-modal attack framework, ACAttack, is proposed for RGB-T multimodal trackers, capable of generating multimodal adversarial patches that can be deployed in both digital and physical domains, inducing tracker failure.
CodeCompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
π― What it does: This study investigates a prompt-aware visual token compression method that achieves inference acceleration by automatically searching for the optimal visual token reduction strategy within a multimodal large language model (MLLM);
ACL: Activating Capability of Linear Attention for Image Restoration
Yubin Gu (Xiamen University), Xiaoshuai Sun (National University of Singapore)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: A new image restoration model called ACL is proposed, which integrates linear attention and the Mamba structure to build an efficient encoder-decoder network.
π― What it does: This study focuses on 3D category incremental learning and proposes a framework called ILPC based on the concept of sparse activation components and task-level fusion.
π― What it does: A framework for efficient distillation of large multimodal models through Active Data Selection (ACID) is proposed, which is further combined with traditional Knowledge Distillation (KD) to obtain the ACED model, significantly improving the performance of small models in zero-shot classification and image-text retrieval.
π― What it does: This paper proposes a sample-free, class-incremental learning framework called ACMap, which achieves fixed inference time by merging task-specific adapters.
π― What it does: Transfer the pre-trained static 3D point cloud model to 4D point cloud videos, proposing Cross-frame Spatio-temporal Adaptation (CSA) to capture short-term and long-term spatio-temporal dynamics.
π― What it does: This paper proposes an adaptive method CLLS and a lightweight RNN model RNLS for trajectory prediction that aims to improve prediction accuracy under varying observation lengths.
π― What it does: This paper proposes Adaptive Dropout, a novel regularization method for blind image super-resolution (blind SR), which can adaptively incorporate Dropout in the intermediate layers of the network and enhance the model's generalization ability through a hierarchical annealing strategy.
Adaptive Keyframe Sampling for Long Video Understanding
Xi Tang (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
CodeOptimizationTransformerVision Language ModelVideoBenchmark
π― What it does: An Adaptive Keyframe Sampling (AKS) algorithm is designed as a pluggable preprocessing module to select the most informative frames as the visual context for MLLM in long video understanding.
Adaptive Unimodal Regulation for Balanced Multimodal Information Acquisition
Chengxiang Huang (Beijing University of Posts and Telecommunications), Di Hu (Renmin University of China)
CodeOptimizationVideoMultimodalityAudio
π― What it does: A method called Information Retrieval Regulation (InfoReg) is proposed to balance information retrieval in multimodal learning, particularly during the early learning phase (referred to as the primary learning window), by suppressing the information retrieval speed of information-rich modalities to promote information retrieval in information-scarce modalities.
π― What it does: A framework for attention data augmentation based on attribution, ADD/ADD+, and novel Calibrated Attribution Maps (CAM) is proposed to improve super-resolution training in low-level vision tasks.
π― What it does: Proposes the FUTURIST framework, which uses a unified visual sequence Transformer for future frame prediction in multimodal (semantic segmentation, depth maps).
π― What it does: This paper proposes a zero-shot self-supervised parallel imaging MRI reconstruction method called AeSPa, aimed at achieving fast and high-quality imaging without the need for fully sampled reference data.
π― What it does: A Stepwise Preference Optimization (SPO) method is proposed to enhance the aesthetic quality of images in text-to-image diffusion models.
Alignment, Mining and Fusion: Representation Alignment with Hard Negative Mining and Selective Knowledge Fusion for Medical Visual Question Answering
Yuanhao Zou (University of Michigan), Zhaozheng Yin (Stony Brook University)
CodeRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodalityBiomedical DataElectronic Health Records
π― What it does: A unified multimodal alignment framework AMiF is proposed, combining hard negative sample mining and selective knowledge fusion, specifically designed for pre-training and fine-tuning in the medical visual question answering (Med-VQA) task.
AlphaPre: Amplitude-Phase Disentanglement Model for Precipitation Nowcasting
Kenghong Lin (Harbin Institute of Technology), Yunming Ye (Harbin Institute of Technology)
CodeGenerationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImageTime Series
π― What it does: This paper proposes the AlphaPre model, which utilizes frequency domain amplitude-phase separation to predict changes in rainfall location through a phase network and changes in rainfall intensity through an amplitude network, and integrates them through AlphaMixer to achieve more refined rainfall forecasting.
π― What it does: A neural CFD solving pipeline that combines Adaptive Mesh Refinement (AMR) with Transformers is proposed, which can efficiently capture long-range dependencies and fine-grained structures in fluid dynamics.
π― What it does: An end-to-end point cloud semantic segmentation network called CDSegNet is developed based on the Conditional-Noise Framework (CNF), utilizing the noise system of DDPM to achieve single-step inference and enhance robustness against noise and sparse data.
π― What it does: This study investigates the synthetic-to-real domain gap in 3D hand pose estimation and proposes a high-quality hand data synthesis pipeline.
Anatomical Consistency and Adaptive Prior-informed Transformation for Multi-contrast MR Image Synthesis via Diffusion Model
Yejee Shin (Yonsei University), Dosik Hwang (Korea Institute of Science and Technology)
CodeGenerationData SynthesisVision Language ModelDiffusion modelImageBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
π― What it does: This paper proposes the APT model, which achieves the synthesis of multi-contrast MR images without modal loss through multi-contrast information fusion and anatomical consistency loss.
π― What it does: In 3D reconstruction based on monocular video, an adaptive and anisotropic sampling strategy called AniGrad is introduced. It utilizes local basis functions to represent TSDF and quickly determines the sampling density of each voxel by combining gradient upper bounds, achieving high-quality and low-latency mesh extraction.
π― What it does: Generate multi-view frontal images from a single portrait image and reconstruct an animatable 3D portrait model using 4D Gaussian projection.
π― What it does: A multi-class industrial anomaly classification framework called AnomalyNCD is proposed, which is compatible with existing anomaly detection methods.
π― What it does: This paper presents AnyMap, a differentiable structure from motion (SfM) framework that can simultaneously estimate dense 3D geometry, camera poses, and a general camera model implemented by a learnable inverse neural network (including radial and tangential distortion) while achieving motion regularization in dynamic scenes.
π― What it does: This paper presents AnySat, a self-supervised learning model for Earth observation that can simultaneously handle various resolutions, scales, and sensors.
Apply Hierarchical-Chain-of-Generation to Complex Attributes Text-to-3D Generation
Yiming Qin (Peking University), Yang Liu (Peking University)
CodeGenerationData SynthesisLarge Language ModelGaussian SplattingText
π― What it does: An automated 3D Gaussian Splatting generation framework called HCoG is proposed, which can automatically chunk and generate and refine 3D assets according to complex attributes and occlusion relationships in the text in an internal-to-external order.
ASAP: Advancing Semantic Alignment Promotes Multi-Modal Manipulation Detecting and Grounding
Zhenxing Zhang (Hefei University of Technology), Meng Wang (Hefei Comprehensive National Science Center)
CodeClassificationRecognitionObject DetectionTransformerLarge Language ModelContrastive LearningImageTextMultimodality
π― What it does: The ASAP framework is proposed for detecting and locating multimodal media forgery, utilizing subtitles and explanatory texts generated by multimodal large language models to enhance the semantic alignment between images and text.
π― What it does: An Asynchronous Collaborative Graph Representation (ACGR) is proposed, which unifies the modeling of frames and events to achieve high-performance, low-latency visual task inference.
π― What it does: A post-training acceleration method called SDTM is proposed, which dynamically merges useless visual tokens in the diffusion transformer (DiT) using a structure-detail denoising prior.
Attribute-formed Class-specific Concept Space: Endowing Language Bottleneck Model with Better Interpretability and Scalability
Jianyang Zhang (University of Electronic Science and Technology of China), Fengmao Lv (Southwest Jiaotong University)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImage
π― What it does: This paper proposes the Attribute-formed Language Bottleneck Model (ALBM), which achieves interpretable image classification by constructing an attribute-based class-specific concept space and combining Visual Attribute Prompt Learning (VAPL) with LLM for automatic concept set generation (DSS).
π― What it does: This paper proposes the Audio-Visual Instance Segmentation (AVIS) task and implements a strong baseline model (AVISM) for the classification, segmentation, and tracking of sound-emitting objects.
Augmented Deep Contexts for Spatially Embedded Video Coding
Yifan Bian (University of Science and Technology of China), Dong Liu (University of Science and Technology of China)
CodeCompressionTransformerVideo
π― What it does: This paper proposes a Spatially Embedded Video Codec (SEVC), which first compresses low-resolution videos to obtain spatial references, and then uses these spatial references along with temporal references to enhance motion vectors and features, ultimately generating a mixed spatial-temporal context and improved latent priors.
Automated Proof of Polynomial Inequalities via Reinforcement Learning
Banglong Liu (East China Normal University), Zhengfeng Yang (East China Normal University)
CodeOptimizationReinforcement Learning
π― What it does: This paper proposes a method for automated proof of polynomial inequalities based on reinforcement learning, transforming the inequality proof into a linear programming problem using the Krivine basis representation, constructing an RL environment, and training an agent with DQN to gradually select basis elements, ultimately obtaining a non-negative representation.
π― What it does: The GETA framework is proposed to achieve joint automated training of structured pruning and quantization, supporting any deep network model.
π― What it does: AutoSSVH compresses unlabeled videos into high-quality hash codes for efficient video retrieval through automated adversarial frame sampling and contrastive learning.
π― What it does: A continuous learning benchmark for audio-visual question answering, AVQACL, is proposed, along with the development of two datasets, Split-AVQA and Split-MUSIC-AVQA. A continuous learning method is introduced that combines question-guided cross-modal fusion (QCIF), task-specific knowledge distillation with spatiotemporal constraints (TKD-STFC), and question semantic consistency constraints (QSCC).
Balanced Direction from Multifarious Choices: Arithmetic Meta-Learning for Domain Generalization
Xiran Wang (Nanjing University), Yinghuan Shi (Nanjing University)
CodeDomain AdaptationMeta LearningImageBenchmark
π― What it does: An arithmetic meta-learning (Arith) framework is proposed, which achieves gradient matching between source domains through a gradient weighting approach and obtains more balanced model parameters by approximating the centroid of the optimal parameters of the source domains.
π― What it does: This work constructs a large-scale basketball video dataset called BASKET, aimed at evaluating players' levels in 20 fine-grained basketball skills (such as three-point shooting, rebounding, passing, etc.) and assessing existing long video recognition models based on this task.
Be More Specific: Evaluating Object-centric Realism in Synthetic Images
Anqi Liang, Aleix Martinez
CodeObject DetectionData SynthesisVision Language ModelImage
π― What it does: This study investigates the assessment of realism in synthetic images from an object-oriented perspective, constructing an object-level realism (OcR) dataset and proposing a corresponding evaluation framework and model.
π― What it does: This paper proposes three benchmark datasets specifically designed to evaluate the generalization performance of satellite image object detection under real distribution shifts (especially spatial distribution shifts) (RWDS-CZ, RWDS-FR, RWDS-HE), and systematically assesses the performance of various mainstream object detection models under single-source and multi-source training settings.
beta-FFT: Nonlinear Interpolation and Differentiated Training Strategies for Semi-Supervised Medical Image Segmentation
Ming Hu (Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences), Quan Wang (Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences)
π― What it does: This paper proposes a FFT-based nonlinear interpolation and differentiation training strategy (Ξ²-FFT) to address the homogenization problem caused by co-training in semi-supervised medical image segmentation.
π― What it does: An Enhanced Instance Replay (EIR) framework is proposed to address the issues of background shift and catastrophic forgetting in continual semantic segmentation through instance-level storage and fusion.
CodeObject DetectionSegmentationTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
π― What it does: The MonoMulti-3DVG task is proposed, and a large-scale multi-target 3D visual alignment (3DVG) dataset called MonoMulti3D-ROPE is constructed from monocular RGB images. The CyclopsNet network is designed to achieve multi-modal semantic alignment and fusion.
π― What it does: This paper proposes a cross-modal anomaly detection method that can generalize to unknown modalities after training on known modalities.
BFANet: Revisiting 3D Semantic Segmentation with Boundary Feature Analysis
Weiguang Zhao (University of Liverpool), Kaizhu Huang (Duke Kunshan University)
CodeSegmentationTransformerPoint Cloud
π― What it does: This paper addresses four types of errors in 3D point cloud semantic segmentation (false positives, merging errors, displacement errors, and region classification errors), proposes corresponding evaluation metrics, and designs the BFANet network to enhance segmentation performance through boundary feature analysis.
π― What it does: This paper proposes a binarizable hybrid visual Transformer (BHViT) aimed at significantly reducing computational load and energy consumption while maintaining high accuracy, making it suitable for deployment on edge devices.
BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning
Hao Zhu (Data61 CSIRO), Piotr Koniusz (Data61 CSIRO)
CodeClassificationDomain AdaptationImage
π― What it does: Proposes BiLoRA - a parameter-efficient continual learning method achieved through bilinear rewriting and fixed orthogonal bases (Fourier).
BiomedCoOp: Learning to Prompt for Biomedical Vision-Language Models
Taha Koleilat (Concordia University), Yiming Xiao (Concordia University)
CodeClassificationAnomaly DetectionKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
π― What it does: For the task of medical image classification with limited annotations, we propose BiomedCoOp, a prompt learning-based framework that utilizes the BiomedCLIP vision-language model and diverse medical prompts generated by GPT-4. By combining semantic consistency and knowledge distillation mechanisms, it achieves efficient and transferable few-shot learning.
CodeRetrievalTransformerVision Language ModelContrastive LearningImageTextBiomedical DataBenchmark
π― What it does: The BIOMEDICA framework was constructed, collecting and organizing 24 million image-text pairs from 6 million PubMed Central Open Access papers, and providing 27 metadata fields; subsequently, a variant of CLIP (BMC-CLIP) was continuously pre-trained on this large-scale dataset.
π― What it does: A black-box forgery and removal attack against semantic watermarks is proposed, demonstrating that a high success rate of watermark transfer can be achieved between different models with just one reference watermark image.
CodeDepth EstimationTransformerImageBiomedical Data
π― What it does: This paper proposes an end-to-end deep learning framework called OCTA-Flow, which can directly estimate blood flow velocity from Optical Coherence Tomography Angiography (OCTA) images, replacing the traditional Optical Doppler Tomography (ODT) that requires expensive hardware and complex signal processing.
BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding
Shuming Liu (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
CodeRetrievalTransformerVision Language ModelVideoMultimodality
π― What it does: The BOLT method is proposed, which enhances the performance of large-scale visual language models in long video understanding tasks through frame selection strategies such as inverse transformation sampling, without the need for additional training.
π― What it does: This paper proposes a semi-supervised polyp segmentation framework RD-Net based on Mean Teacher, which utilizes an auxiliary student network to train depth images and achieves high-precision segmentation during inference using only RGB through depth-guided cross-modal mutual learning.
π― What it does: A model space augmentation-based adversarial attack method called OPS is proposed, which constructs neighborhood models in the hypothesis space using random input transformation operators and perturbations, and solves random optimization, significantly enhancing the transferability of adversarial samples.
Boosting the Dual-Stream Architecture in Ultra-High Resolution Segmentation with Resolution-Biased Uncertainty Estimation
Rong Qin (Nankai University), Jufeng Yang (Nankai University)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: A dual-stream ultra-high resolution (UHR) semantic segmentation framework is proposed, which simultaneously achieves three main objectives: feature fusion, important region amplification, and detail supplementation by estimating resolution deviation uncertainty.
π― What it does: A framework is proposed that utilizes a self-supervised masking strategy (self-view mask and cross-view mask) for representation learning on unpaired and asynchronous first-person and third-person videos.
π― What it does: This study investigates the low-rank bottleneck of linear attention, proposing Rank-Augmented Linear Attention (RALA) and constructing the Rank-Augmented Vision Linear Transformer (RAVLT) based on it to achieve efficient visual modeling.
Bridging Past and Future: End-to-End Autonomous Driving with Historical Prediction and Planning
Bozhou Zhang (Fudan University), Li Zhang (Fudan University)
CodeAutonomous DrivingTransformerPoint Cloud
π― What it does: Proposes the BridgeAD framework, which integrates historical prediction and planning information into an end-to-end autonomous driving system.
π― What it does: This paper proposes a semantic correspondence method that combines 3D geometric alignment with deformation and sparse semantic matching, utilizing DUSt3R trained on synthetic cross-instance perspective data to achieve efficient and robust correspondences.
Zhaozhi Wang (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
CodeClassificationRestorationObject DetectionSegmentationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImagePhysics Related
π― What it does: A visual representation model vHeat based on the physical principles of thermal diffusion is proposed, and a Heat Conduction Operator (HCO) is designed to achieve global information propagation using DCT/IDCT;
π― What it does: Construct an end-to-end Transformer architecture for reconstructing 3D wireframe models of buildings from aerial LiDAR point clouds.
π― What it does: An improved method called ByTheWay has been developed, which is training-independent, requires no additional parameters, and incurs no sampling costs. It aims to enhance the structural rationality, temporal consistency, and motion amplitude of text-to-video (T2V) generation models.
π― What it does: A post-hoc OOD detection method based on category-related relative feature error, CARef, is proposed, and further enhanced to a more powerful CADRef through feature decomposition and error scaling.
Camera Resection from Known Line Pencils and a Radially Distorted Scanline
Juan C. Dibene (Stevens Institute of Technology), Enrique Dunn (Stevens Institute of Technology)
CodePose EstimationSimultaneous Localization and MappingImage
π― What it does: This paper proposes a framework for absolute camera pose estimation based on a single radial distortion scanning line, including 6-point minimal solutions, 7-point unique solutions, and 8+ linear solutions;
π― What it does: A hardware-adaptive background attention adversarial training defense method against latency attacks on NMS-based object detectors for edge devices is proposed.
π― What it does: This paper presents CarPlanner, a consistency autoregressive trajectory planner that utilizes reinforcement learning to generate multimodal trajectories, addressing the training efficiency and performance issues of large-scale real driving tasks.
CodeCompressionTransformerLarge Language ModelImageVideoTextMultimodality
π― What it does: A very low-bit compression method CASP based on attention sparsity is proposed, which compresses large multimodal models using low-rank decomposition and quantization.
CATANet: Efficient Content-Aware Token Aggregation for Lightweight Image Super-Resolution
Xin Liu (Nanjing University), Gangshan Wu (Nanjing University)
CodeRestorationSuper ResolutionTransformerImage
π― What it does: A lightweight super-resolution network called CATANet is proposed, which achieves long-distance information interaction through content-aware token aggregation.
Causal Composition Diffusion Model for Closed-loop Traffic Generation
Haohong Lin (Carnegie Mellon University), Hongge Chen (Cruise LLC)
CodeGenerationAutonomous DrivingDiffusion modelTime Series
π― What it does: This paper studies a causal combination diffusion model called CCDiff, designed for generating closed-loop traffic scenarios while balancing controllability and realism.