Pub Date : 2026-02-01Epub Date: 2025-11-24DOI: 10.1016/j.cag.2025.104493
Mengyi Wang, Beiqi Chen, Shengfang Pan, Niansheng Liu, Jinhe Su
Efficient and high-quality reconstruction of large-scale 3D scenes remains a key challenge for novel view synthesis. Recent advances in 3D Gaussian Splatting (3DGS) have achieved photorealistic rendering and real-time performance, but scaling 3DGS to city-scale environments typically relies on block-based training. This divide-and-conquer approach suffers from two major limitations: (1) the Gaussian properties of overlapping regions of adjacent blocks are inconsistent, resulting in noticeable visual artifacts after merging; (2) the sparse Gaussian distribution near block boundaries causes cracks or holes. To address these challenges, we propose a novel framework that regularizes the Gaussian properties of overlapping regions and enhances the Gaussian density near block edges, thus ensuring smooth transitions and seamless rendering. In addition, we introduce appearance decouple to further adapt to viewpoint-dependent appearance variations in urban scenes and adopt a multi-scale densification strategy to balance details and efficiency at different scene scales. Experimental results show that in large-scale urban scenes with densely partitioned blocks, our method achieves consistently better reconstruction quality, with an average PSNR improvement of 0.25 dB over strong baselines on both aerial and street datasets.
{"title":"Consistency-preserving Gaussian splatting for block-based large-scale scene reconstruction","authors":"Mengyi Wang, Beiqi Chen, Shengfang Pan, Niansheng Liu, Jinhe Su","doi":"10.1016/j.cag.2025.104493","DOIUrl":"10.1016/j.cag.2025.104493","url":null,"abstract":"<div><div>Efficient and high-quality reconstruction of large-scale 3D scenes remains a key challenge for novel view synthesis. Recent advances in 3D Gaussian Splatting (3DGS) have achieved photorealistic rendering and real-time performance, but scaling 3DGS to city-scale environments typically relies on block-based training. This divide-and-conquer approach suffers from two major limitations: (1) the Gaussian properties of overlapping regions of adjacent blocks are inconsistent, resulting in noticeable visual artifacts after merging; (2) the sparse Gaussian distribution near block boundaries causes cracks or holes. To address these challenges, we propose a novel framework that regularizes the Gaussian properties of overlapping regions and enhances the Gaussian density near block edges, thus ensuring smooth transitions and seamless rendering. In addition, we introduce appearance decouple to further adapt to viewpoint-dependent appearance variations in urban scenes and adopt a multi-scale densification strategy to balance details and efficiency at different scene scales. Experimental results show that in large-scale urban scenes with densely partitioned blocks, our method achieves consistently better reconstruction quality, with an average PSNR improvement of 0.25 dB over strong baselines on both aerial and street datasets.</div></div>","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104493"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145625036","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2025-11-25DOI: 10.1016/j.cag.2025.104489
Germain Garcia-Zanabria , Daniel A. Gutierrez-Pachas , Jorge Poco , Erick Gomez-Nieto
Student dropout is a major concern for universities, leading them to invest heavily in strategies to lower attrition rates. Analytical tools are crucial for predicting dropout risks and informing policies on academic and social support. However, many of these tools depend solely on automate tu d predictions, ignoring valuable insights from professors, mentors, and specialists. These experts can help identify the root causes of dropout and develop effective interventions. This paper introduces CSDA-Vis, a visualization system designed to analyze the influence of individual, institutional, and socioeconomic factors on student dropout rates. CSDA-Vis facilitates the identification of actionable strategies to mitigate dropout by integrating counterfactual and survival analysis methods. Unlike traditional approaches, our tool enables decision-makers to incorporate their expertise into the evaluation of different dropout scenarios. Developed in collaboration with domain experts, CSDA-Vis builds upon previous visualization tools and was validated through a case study using real datasets from a Latin American university. Additionally, we conducted an expert evaluation with professionals specializing in dropout analysis, further demonstrating the tool’s practical value and effectiveness.
{"title":"CSDA-Vis: A (What-If-and-When) visual system for early dropout detection using counterfactual and survival analysis interactions","authors":"Germain Garcia-Zanabria , Daniel A. Gutierrez-Pachas , Jorge Poco , Erick Gomez-Nieto","doi":"10.1016/j.cag.2025.104489","DOIUrl":"10.1016/j.cag.2025.104489","url":null,"abstract":"<div><div>Student dropout is a major concern for universities, leading them to invest heavily in strategies to lower attrition rates. Analytical tools are crucial for predicting dropout risks and informing policies on academic and social support. However, many of these tools depend solely on automate tu d predictions, ignoring valuable insights from professors, mentors, and specialists. These experts can help identify the root causes of dropout and develop effective interventions. This paper introduces <em>CSDA-Vis</em>, a visualization system designed to analyze the influence of individual, institutional, and socioeconomic factors on student dropout rates. CSDA-Vis facilitates the identification of actionable strategies to mitigate dropout by integrating counterfactual and survival analysis methods. Unlike traditional approaches, our tool enables decision-makers to incorporate their expertise into the evaluation of different dropout scenarios. Developed in collaboration with domain experts, CSDA-Vis builds upon previous visualization tools and was validated through a case study using real datasets from a Latin American university. Additionally, we conducted an expert evaluation with professionals specializing in dropout analysis, further demonstrating the tool’s practical value and effectiveness.</div></div>","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104489"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145625037","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2025-11-24DOI: 10.1016/j.cag.2025.104500
Zeinab BagheriFard , Miruna Maria Vasiliu , Emma Jane Pretty , Luis Quintero , Benjamin Edvinsson , Mario Romero , Renan Guarese
Immersive technologies offer advantages for the visualization of and interaction with complex setups within manufacturing maintenance processes. The present work catalogs different applications of AR/VR in manufacturing maintenance practices as an extended version of a workshop paper presented at the International Workshop on eXtended Reality for Industrial and Occupational Supports (XRIOS). Through a scoping review in three computing and engineering digital libraries, we outline the key attributes of immersive solutions () for industrial maintenance, categorizing functional prototypes with ten parameters related to interaction, visualization, and research methods. Moreover, we conducted a workshop with three manufacturing experts discussing the future of maintenance interfaces. By bringing forth their recommendations and insights, we targeted a key training challenge in maintenance. We designed and implemented a situated visualization prototype for a safety-critical procedure with real-time, in-depth, spatially relevant instructions. We compared the effects of 2D labels and 3D ghosts in a VR-simulated AR environment. In a preliminary between-subjects evaluation study (), we measured usability, workload, simulator sickness, completion time, and delayed recall. Although we did not find statistically significant differences between conditions, 3D ghosts showed slightly lower perceived workload and discomfort levels, along with shorter completion times. On the other hand, 2D labels produced higher usability. Overall, we contribute by a mapping out the state of the art and discovering knowledge gaps within immersive maintenance, presenting a design and preliminary user study that adheres to our recommendations.
{"title":"Situated visualization towards manufacturing maintenance training: Scoping review, design and user study","authors":"Zeinab BagheriFard , Miruna Maria Vasiliu , Emma Jane Pretty , Luis Quintero , Benjamin Edvinsson , Mario Romero , Renan Guarese","doi":"10.1016/j.cag.2025.104500","DOIUrl":"10.1016/j.cag.2025.104500","url":null,"abstract":"<div><div>Immersive technologies offer advantages for the visualization of and interaction with complex setups within manufacturing maintenance processes. The present work catalogs different applications of AR/VR in manufacturing maintenance practices as an extended version of a workshop paper presented at the International Workshop on eXtended Reality for Industrial and Occupational Supports (XRIOS). Through a scoping review in three computing and engineering digital libraries, we outline the key attributes of immersive solutions (<span><math><mrow><msub><mrow><mi>N</mi></mrow><mrow><mi>p</mi></mrow></msub><mo>=</mo><mn>115</mn></mrow></math></span>) for industrial maintenance, categorizing functional prototypes with ten parameters related to interaction, visualization, and research methods. Moreover, we conducted a workshop with three manufacturing experts discussing the future of maintenance interfaces. By bringing forth their recommendations and insights, we targeted a key training challenge in maintenance. We designed and implemented a situated visualization prototype for a safety-critical procedure with real-time, in-depth, spatially relevant instructions. We compared the effects of 2D labels and 3D ghosts in a VR-simulated AR environment. In a preliminary between-subjects evaluation study (<span><math><mrow><msub><mrow><mi>N</mi></mrow><mrow><mi>u</mi></mrow></msub><mo>=</mo><mn>24</mn></mrow></math></span>), we measured usability, workload, simulator sickness, completion time, and delayed recall. Although we did not find statistically significant differences between conditions, 3D ghosts showed slightly lower perceived workload and discomfort levels, along with shorter completion times. On the other hand, 2D labels produced higher usability. Overall, we contribute by a mapping out the state of the art and discovering knowledge gaps within immersive maintenance, presenting a design and preliminary user study that adheres to our recommendations.</div></div>","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104500"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145625039","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2025-11-19DOI: 10.1016/j.cag.2025.104492
Jin-Feng Li , Sen-Zhe Xu , Qiang Tong , Peng-Hui Yuan , Ling-Long Zou , Er-Xia Luo , Qi Wen Gan , Song-Hai Zhang
Redirected walking (RDW) is a virtual reality locomotion technique that enables users to explore large virtual environments within a limited physical space. While state-of-the-art methods based on physical trajectory planning make effective use of physical space, some of them often compromise user comfort due to frequent directional reversals in curvature gain. To address this, this paper proposes a novel RDW method that integrates strafing gain with pose score guidance. Our approach discretizes the physical space into a series of standard poses, each with a long-term safety score, and redirects the user toward the optimal pose. The main contribution is a path generation algorithm that decomposes redirection into two sequential stages to ensure stable gains for each planned path: it first uses the curvature gain to steer the user along an arc for orientation alignment, and then inserts a straight path segment with constant strafing gain to achieve positional alignment with the target pose. Simulation experiments demonstrate a reduction in resets, while the user study shows lower Simulator Sickness Questionnaire scores compared to previous methods. Our work explores the potential of combining novel gains with state-of-the-art methods to create a more effective and comfortable RDW controller algorithm.
{"title":"Incorporating strafing gain into redirected walking with pose score guidance","authors":"Jin-Feng Li , Sen-Zhe Xu , Qiang Tong , Peng-Hui Yuan , Ling-Long Zou , Er-Xia Luo , Qi Wen Gan , Song-Hai Zhang","doi":"10.1016/j.cag.2025.104492","DOIUrl":"10.1016/j.cag.2025.104492","url":null,"abstract":"<div><div>Redirected walking (RDW) is a virtual reality locomotion technique that enables users to explore large virtual environments within a limited physical space. While state-of-the-art methods based on physical trajectory planning make effective use of physical space, some of them often compromise user comfort due to frequent directional reversals in curvature gain. To address this, this paper proposes a novel RDW method that integrates strafing gain with pose score guidance. Our approach discretizes the physical space into a series of standard poses, each with a long-term safety score, and redirects the user toward the optimal pose. The main contribution is a path generation algorithm that decomposes redirection into two sequential stages to ensure stable gains for each planned path: it first uses the curvature gain to steer the user along an arc for orientation alignment, and then inserts a straight path segment with constant strafing gain to achieve positional alignment with the target pose. Simulation experiments demonstrate a reduction in resets, while the user study shows lower Simulator Sickness Questionnaire scores compared to previous methods. Our work explores the potential of combining novel gains with state-of-the-art methods to create a more effective and comfortable RDW controller algorithm.</div></div>","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104492"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145580258","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2025-12-17DOI: 10.1016/j.cag.2025.104526
Leonardo Sacht, Marcos Lage, Ricardo Marroquim
{"title":"Foreword to special section on SIBGRAPI 2025","authors":"Leonardo Sacht, Marcos Lage, Ricardo Marroquim","doi":"10.1016/j.cag.2025.104526","DOIUrl":"10.1016/j.cag.2025.104526","url":null,"abstract":"","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104526"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145796670","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2025-12-17DOI: 10.1016/j.cag.2025.104523
A. Augusto de Sousa , Miguel Angel Guevara López , Traian Lavric
{"title":"Foreword to the special section on the 29th international ACM conference on 3D web technology","authors":"A. Augusto de Sousa , Miguel Angel Guevara López , Traian Lavric","doi":"10.1016/j.cag.2025.104523","DOIUrl":"10.1016/j.cag.2025.104523","url":null,"abstract":"","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104523"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145883787","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2026-01-02DOI: 10.1016/j.cag.2025.104528
Yifei Tian, Xiangyu Li, Jieming Yin
Most existing point cloud registration methods heavily rely on accurate correspondences between the source and target point clouds, such as point-level or superpoint-level matches. In dense and balanced point clouds where local geometric structures are relatively complete, correspondences are easier to establish, leading to satisfactory registration performance. However, real-world point clouds can be sparse or imbalanced. The absence or inconsistency of local geometric structures makes it challenging to construct reliable correspondences, significantly degrading the performance of mainstream registration methods. To address this challenge, we propose P2NCorr, a pseudo-to-non-correspondence registration method designed for robust alignment in point clouds with missing or low-quality correspondences. Our method leverages an attention-guided soft matching module that uses self- and cross-attention mechanisms to extract contextual features and constructs pseudo correspondences under slack constraints. On this basis, we introduce a geometric consistency metric based on the thickness-guided self-correction module, which enables fine-grained alignment and optimization of micro-surfaces in the fused point cloud. This thickness evaluation serves as a supplementary supervisory signal, forming a comprehensive feedback from the post-registration fusion to the feature extraction module, thereby improving both the accuracy and stability of the registration process. Experiments conducted on public datasets such as ModelNet40 and 7Scenes demonstrate that P2NCorr achieves high-precision registration even under challenging conditions. Especially when point clouds are sparse, sampling is imbalanced, and measurements are noisy, experiments demonstrate strong robustness and promising potential.
{"title":"From pseudo- to non-correspondences: Robust point cloud registration via thickness-guided self-correction","authors":"Yifei Tian, Xiangyu Li, Jieming Yin","doi":"10.1016/j.cag.2025.104528","DOIUrl":"10.1016/j.cag.2025.104528","url":null,"abstract":"<div><div>Most existing point cloud registration methods heavily rely on accurate correspondences between the source and target point clouds, such as point-level or superpoint-level matches. In dense and balanced point clouds where local geometric structures are relatively complete, correspondences are easier to establish, leading to satisfactory registration performance. However, real-world point clouds can be sparse or imbalanced. The absence or inconsistency of local geometric structures makes it challenging to construct reliable correspondences, significantly degrading the performance of mainstream registration methods. To address this challenge, we propose P2NCorr, a pseudo-to-non-correspondence registration method designed for robust alignment in point clouds with missing or low-quality correspondences. Our method leverages an attention-guided soft matching module that uses self- and cross-attention mechanisms to extract contextual features and constructs pseudo correspondences under slack constraints. On this basis, we introduce a geometric consistency metric based on the thickness-guided self-correction module, which enables fine-grained alignment and optimization of micro-surfaces in the fused point cloud. This thickness evaluation serves as a supplementary supervisory signal, forming a comprehensive feedback from the post-registration fusion to the feature extraction module, thereby improving both the accuracy and stability of the registration process. Experiments conducted on public datasets such as ModelNet40 and 7Scenes demonstrate that P2NCorr achieves high-precision registration even under challenging conditions. Especially when point clouds are sparse, sampling is imbalanced, and measurements are noisy, experiments demonstrate strong robustness and promising potential.</div></div>","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104528"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145938376","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2025-12-02DOI: 10.1016/j.cag.2025.104507
George Westergaard , Mark Ellis , Jacob Barker , Sofia Garces Palacios , Alexis Desir , Ganesh Sankaranarayanan , Suvranu De , Doga Demirel
In this work, we present a real-time virtual reality-based open surgery simulator that enables realistic soft-tissue suturing with bimanual haptic feedback. Our system uses eXtended Position-Based Dynamics (XPBD) for soft body and suture thread simulation, allowing stable real-time physics for complex interactions like continuous sutures and knot tying. In tests with all four common suturing techniques, purse-string, Connell, stay, and Lembert, the simulator maintained high frame rates (50–80 FPS) with up to 4155 simulated particles, demonstrating consistent real-time performance. As part of our work, we conducted a user study using our suturing simulator, where 24 surgical trainees and experts used the Virtual Colorectal Surgery Trainer – Rectal Prolapse simulator. The user study showed that 71% of participants (n=17) rated the anatomical realism as moderate to very high. Half (n=12) found the force feedback realistic, and 54% (n=13) participants found the force feedback useful, indicating effective immersion while also highlighting the need for improved haptic fidelity. Overall, the simulation provides a low-cost, high-fidelity training platform for open surgical suturing, addressing a critical gap in current virtual reality educational tools.
{"title":"Real-time haptic-based soft body suturing in virtual open surgery simulations","authors":"George Westergaard , Mark Ellis , Jacob Barker , Sofia Garces Palacios , Alexis Desir , Ganesh Sankaranarayanan , Suvranu De , Doga Demirel","doi":"10.1016/j.cag.2025.104507","DOIUrl":"10.1016/j.cag.2025.104507","url":null,"abstract":"<div><div>In this work, we present a real-time virtual reality-based open surgery simulator that enables realistic soft-tissue suturing with bimanual haptic feedback. Our system uses eXtended Position-Based Dynamics (XPBD) for soft body and suture thread simulation, allowing stable real-time physics for complex interactions like continuous sutures and knot tying. In tests with all four common suturing techniques, purse-string, Connell, stay, and Lembert, the simulator maintained high frame rates (50–80 FPS) with up to 4155 simulated particles, demonstrating consistent real-time performance. As part of our work, we conducted a user study using our suturing simulator, where 24 surgical trainees and experts used the Virtual Colorectal Surgery Trainer – Rectal Prolapse simulator. The user study showed that 71% of participants (n=17) rated the anatomical realism as moderate to very high. Half (n=12) found the force feedback realistic, and 54% (n=13) participants found the force feedback useful, indicating effective immersion while also highlighting the need for improved haptic fidelity. Overall, the simulation provides a low-cost, high-fidelity training platform for open surgical suturing, addressing a critical gap in current virtual reality educational tools.</div></div>","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104507"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145693885","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-01Epub Date: 2025-12-05DOI: 10.1016/j.cag.2025.104512
Chun-Hau Yu, Yu-Hsiang Chen, Cheng-Yen Yu, Li-Chen Fu
3D human avatar reconstruction has become a popular research field in recent years. Although many studies have shown remarkable results, most existing methods either impose overly strict data requirements, such as depth information or multi-view images, or suffer from significant performance drops in specific areas. To address these challenges, we propose HARDER. We combine the Score Distillation Sampling (SDS) technique with the designed modules, Feature-Specific Image Captioning (FSIC) and RADR (Region-Aware Differentiable Rendering), allowing the Latent Diffusion Model (LDM) to guide the reconstruction process, especially in unseen regions. Furthermore, we have developed various training strategies, including personalized LDM, delayed SDS, focused SDS, and multi-pose SDS, to make the training process more efficient.
Our avatars use an explicit representation that is compatible with modern computer graphics pipelines. Also, the entire reconstruction and real-time animation process can be completed on a single consumer-grade GPU, making this application more accessible.
{"title":"HARDER: 3D human avatar reconstruction with distillation and explicit representation","authors":"Chun-Hau Yu, Yu-Hsiang Chen, Cheng-Yen Yu, Li-Chen Fu","doi":"10.1016/j.cag.2025.104512","DOIUrl":"10.1016/j.cag.2025.104512","url":null,"abstract":"<div><div>3D human avatar reconstruction has become a popular research field in recent years. Although many studies have shown remarkable results, most existing methods either impose overly strict data requirements, such as depth information or multi-view images, or suffer from significant performance drops in specific areas. To address these challenges, we propose HARDER. We combine the Score Distillation Sampling (SDS) technique with the designed modules, Feature-Specific Image Captioning (FSIC) and RADR (Region-Aware Differentiable Rendering), allowing the Latent Diffusion Model (LDM) to guide the reconstruction process, especially in unseen regions. Furthermore, we have developed various training strategies, including personalized LDM, delayed SDS, focused SDS, and multi-pose SDS, to make the training process more efficient.</div><div>Our avatars use an explicit representation that is compatible with modern computer graphics pipelines. Also, the entire reconstruction and real-time animation process can be completed on a single consumer-grade GPU, making this application more accessible.</div></div>","PeriodicalId":50628,"journal":{"name":"Computers & Graphics-Uk","volume":"134 ","pages":"Article 104512"},"PeriodicalIF":2.8,"publicationDate":"2026-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145796668","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号