arXiv:2511.13684v1 Announce Type: cross 
Abstract: We introduce GS-Light, an efficient, textual position-aware pipeline for text-guided relighting of 3D scenes represented via Gaussian Splatting (3DGS). GS-Light implements a training-free extension of a single-input diffusion model to handle multi-view inputs. Given a user prompt that may specify lighting direction, color, intensity, or reference objects, we employ a large vision-language model (LVLM) to parse the prompt into lighting priors. Using off-the-shelf estimators for geometry and semantics (depth, surface normals, and semantic segmentation), we fuse these lighting priors with view-geometry constraints to compute illumination maps and generate initial latent codes for each view. These meticulously derived init latents guide the diffusion model to generate relighting outputs that more accurately reflect user expectations, especially in terms of lighting direction. By feeding multi-view rendered images, along with the init latents, into our multi-view relighting model, we produce high-fidelity, artistically relit images. Finally, we fine-tune the 3DGS scene with the relit appearance to obtain a fully relit 3D scene. We evaluate GS-Light on both indoor and outdoor scenes, comparing it to state-of-the-art baselines including per-view relighting, video relighting, and scene editing methods. Using quantitative metrics (multi-view consistency, imaging quality, aesthetic score, semantic similarity, etc.) and qualitative assessment (user studies), GS-Light demonstrates consistent improvements over baselines. Code and assets will be made available upon publication.

يقدم المقال GS-Light، وهو خط أنابيب جديد مصمم لإعادة إضاءة المشاهد ثلاثية الأبعاد الموجهة بالنص، باستخدام Gaussian Splatting (3DGS). يعمل هذا النظام دون الحاجة إلى تدريب، حيث يوسع نموذج انتشار الإدخال الفردي للتعامل مع المدخلات متعددة المناظر. من خلال تفسير مطالب المستخدمين المتعلقة بخصائص الإضاءة عبر نموذج لغة رؤية كبير (LVLM)، يقوم GS-Light بإنشاء خرائط إضاءة ورموز كامنة أولية لكل منظر. تعمل هذه العملية على تحسين دقة مخرجات إعادة الإضاءة، مما يجعلها تتماشى بشكل أفضل مع توقعات المستخدمين، خاصة فيما يتعلق باتجاه الإضاءة.

El artículo presenta GS-Light, un nuevo pipeline diseñado para el relighting de escenas 3D guiado por texto, utilizando Gaussian Splatting (3DGS). Este sistema opera sin necesidad de entrenamiento, extendiendo un modelo de difusión de entrada única para manejar entradas de múltiples vistas. Al interpretar las solicitudes de los usuarios relacionadas con atributos de iluminación a través de un gran modelo de visión-lenguaje (LVLM), GS-Light genera mapas de iluminación y códigos latentes iniciales para cada vista. Este proceso mejora la precisión de las salidas de relighting, alineándolas más es…

L'article présente GS-Light, un nouveau pipeline conçu pour le relighting de scènes 3D guidé par du texte, utilisant le Gaussian Splatting (3DGS). Ce système fonctionne sans nécessiter d'entraînement, étendant un modèle de diffusion à entrée unique pour accueillir des entrées multi-vues. En interprétant les invites des utilisateurs liées aux attributs d'éclairage via un grand modèle de vision-langage (LVLM), GS-Light génère des cartes d'illumination et des codes latents initiaux pour chaque vue. Ce processus améliore la précision des sorties de relighting, les alignant davantage sur les attent…

The article presents GS-Light, a novel pipeline designed for text-guided relighting of 3D scenes using Gaussian Splatting (3DGS). This system operates without the need for training, extending a single-input diffusion model to accommodate multi-view inputs. By interpreting user prompts related to lighting attributes through a large vision-language model (LVLM), GS-Light generates illumination maps and initial latent codes for each view. This process enhances the accuracy of relighting outputs, aligning them more closely with user expectations, particularly regarding lighting direction.

Training-Free Multi-View Extension of IC-Light for Textual Position-Aware Scene Relighting

arXiv:2512.03621v1 Announce Type: new 
Abstract: We propose ReCamDriving, a purely vision-based, camera-controlled novel-trajectory video generation framework. While repair-based methods fail to restore complex artifacts and LiDAR-based approaches rely on sparse and incomplete cues, ReCamDriving leverages dense and scene-complete 3DGS renderings for explicit geometric guidance, achieving precise camera-controllable generation. To mitigate overfitting to restoration behaviors when conditioned on 3DGS renderings, ReCamDriving adopts a two-stage training paradigm: the first stage uses camera poses for coarse control, while the second stage incorporates 3DGS renderings for fine-grained viewpoint and geometric guidance. Furthermore, we present a 3DGS-based cross-trajectory data curation strategy to eliminate the train-test gap in camera transformation patterns, enabling scalable multi-trajectory supervision from monocular videos. Based on this strategy, we construct the ParaDrive dataset, containing over 110K parallel-trajectory video pairs. Extensive experiments demonstrate that ReCamDriving achieves state-of-the-art camera controllability and structural consistency.

تم تقديم ReCamDriving كإطار جديد لتوليد مسارات الفيديو التي يتم التحكم فيها بالكاميرا دون استخدام LiDAR، حيث يعتمد بدلاً من ذلك على رسومات 3D Gaussian Splatting (3DGS) الكثيفة لتوفير توجيه هندسي أفضل. تهدف هذه الطريقة إلى التغلب على قيود الأساليب الحالية التي تواجه صعوبات مع العيوب المعقدة أو البيانات النادرة. يستخدم الإطار عملية تدريب من مرحلتين لتحسين التحكم في الكاميرا وزيادة دقة توليد الفيديو.

Se ha presentado ReCamDriving como un marco novedoso para la generación de trayectorias de video controladas por cámara sin el uso de LiDAR, basándose en renderizaciones densas de 3D Gaussian Splatting (3DGS) para una mejor guía geométrica. Este enfoque busca superar las limitaciones de los métodos existentes que luchan con artefactos complejos o datos escasos. El marco emplea un proceso de entrenamiento en dos etapas para refinar el control de la cámara y mejorar la precisión de la generación de video.

ReCamDriving a été introduit comme un cadre novateur pour générer des trajectoires vidéo contrôlées par caméra sans utiliser de LiDAR, s'appuyant plutôt sur des rendus 3D Gaussian Splatting (3DGS) denses pour une meilleure guidance géométrique. Cette approche vise à surmonter les limitations des méthodes existantes qui peinent avec des artefacts complexes ou des données éparses. Le cadre utilise un processus de formation en deux étapes pour affiner le contrôle de la caméra et améliorer la précision de la génération vidéo.

ReCamDriving has been introduced as a novel framework for generating camera-controlled video trajectories without the use of LiDAR, relying instead on dense 3D Gaussian Splatting (3DGS) renderings for enhanced geometric guidance. This approach aims to overcome limitations of existing methods that struggle with complex artifacts or sparse data. The framework employs a two-stage training process to refine camera control and improve video generation accuracy.

ReCamDriving: LiDAR-Free Camera-Controlled Novel Trajectory Video Generation

arXiv:2508.08136v2 Announce Type: replace 
Abstract: The success of 3DGS in generative and editing applications has sparked growing interest in 3DGS-based style transfer. However, current methods still face two major challenges: (1) multi-view inconsistency often leads to style conflicts, resulting in appearance smoothing and distortion; and (2) heavy reliance on VGG features, which struggle to disentangle style and content from style images, often causing content leakage and excessive stylization. To tackle these issues, we introduce \textbf{FantasyStyle}, a 3DGS-based style transfer framework, and the first to rely entirely on diffusion model distillation. It comprises two key components: (1) \textbf{Multi-View Frequency Consistency}. We enhance cross-view consistency by applying a 3D filter to multi-view noisy latent, selectively reducing low-frequency components to mitigate stylized prior conflicts. (2) \textbf{Controllable Stylized Distillation}. To suppress content leakage from style images, we introduce negative guidance to exclude undesired content. In addition, we identify the limitations of Score Distillation Sampling and Delta Denoising Score in 3D style transfer and remove the reconstruction term accordingly. Building on these insights, we propose a controllable stylized distillation that leverages negative guidance to more effectively optimize the 3D Gaussians. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art approaches, achieving higher stylization quality and visual realism across various scenes and styles. The code is available at https://github.com/yangyt46/FantasyStyle.

يمثل تقديم FantasyStyle تقدمًا كبيرًا في تقنية 3D Gaussian Splatting (3DGS) من خلال معالجة التحديات المتعلقة بنقل الأنماط، وخاصة عدم التناسق بين وجهات النظر وتسرب المحتوى. يعتمد هذا الإطار على تقنيات تقطير نماذج الانتشار لتعزيز التناسق بين وجهات النظر والتحكم في التزيين، بهدف تحسين جودة المحتوى ثلاثي الأبعاد الناتج.

La introducción de FantasyStyle marca un avance significativo en el 3D Gaussian Splatting (3DGS) al abordar desafíos en la transferencia de estilo, particularmente la inconsistencia de múltiples vistas y la fuga de contenido. Este marco utiliza la destilación de modelos de difusión para mejorar la consistencia entre vistas y controlar la estilización, con el objetivo de mejorar la calidad del contenido 3D generado.

L'introduction de FantasyStyle représente une avancée significative dans le domaine du 3D Gaussian Splatting (3DGS) en s'attaquant aux défis du transfert de style, notamment l'incohérence multi-vues et les fuites de contenu. Ce cadre utilise la distillation de modèles de diffusion pour améliorer la cohérence entre les vues et contrôler la stylisation, visant à améliorer la qualité du contenu 3D généré.

The introduction of FantasyStyle marks a significant advancement in 3D Gaussian Splatting (3DGS) by addressing challenges in style transfer, particularly multi-view inconsistency and content leakage. This framework utilizes diffusion model distillation to enhance cross-view consistency and control stylization, aiming to improve the quality of generated 3D content.

Training-Free Multi-View Extension of IC-Light for Textual Position-Aware Scene Relighting

Was this article worth reading? Share it

LucidQuery AI

CreativeDevJobs

4o Image Gen