arXiv:2511.11231v1 Announce Type: new 
Abstract: High-fidelity gaze redirection is critical for generating augmented data to improve the generalization of gaze estimators. 3D Gaussian Splatting (3DGS) models like GazeGaussian represent the state-of-the-art but can struggle with rendering subtle, continuous gaze shifts. In this paper, we propose DiT-Gaze, a framework that enhances 3D gaze redirection models using a novel combination of Diffusion Transformer (DiT), weak supervision across gaze angles, and an orthogonality constraint loss. DiT allows higher-fidelity image synthesis, while our weak supervision strategy using synthetically generated intermediate gaze angles provides a smooth manifold of gaze directions during training. The orthogonality constraint loss mathematically enforces the disentanglement of internal representations for gaze, head pose, and expression. Comprehensive experiments show that DiT-Gaze sets a new state-of-the-art in both perceptual quality and redirection accuracy, reducing the state-of-the-art gaze error by 4.1% to 6.353 degrees, providing a superior method for creating synthetic training data. Our code and models will be made available for the research community to benchmark against.

تقدم الورقة DiT-Gaze، وهو إطار جديد لإعادة توجيه النظر بدقة عالية، والذي يعزز النماذج الحالية لتقنية 3D Gaussian Splatting مثل GazeGaussian. يستخدم DiT-Gaze مزيجًا من Diffusion Transformer، والإشراف الضعيف عبر زوايا النظر، وفقدان قيود التعامد لتحسين عرض التحولات الدقيقة في النظر. تشير التجارب إلى أن DiT-Gaze يحقق مستوى جديدًا من الجودة الإدراكية ودقة إعادة التوجيه، حيث يقلل من خطأ النظر بمقدار 6.353 درجة ويحسن الدقة بنسبة 4.1%.

El artículo presenta DiT-Gaze, un nuevo marco para la redirección de la mirada de alta fidelidad, que mejora los modelos existentes de 3D Gaussian Splatting como GazeGaussian. DiT-Gaze utiliza una combinación de Diffusion Transformer, supervisión débil a través de ángulos de mirada y una pérdida de restricción de ortogonalidad para mejorar el renderizado de cambios sutiles en la mirada. Los experimentos indican que DiT-Gaze establece un nuevo estado del arte en calidad perceptual y precisión de redirección, reduciendo el error de mirada en 6.353 grados y mejorando la precisión en un 4.1%.

Cet article présente DiT-Gaze, un nouveau cadre pour la redirection de regard haute fidélité, qui améliore les modèles existants de 3D Gaussian Splatting comme GazeGaussian. DiT-Gaze utilise une combinaison de Diffusion Transformer, de supervision faible à travers les angles de regard et d'une perte de contrainte d'orthogonalité pour améliorer le rendu des changements de regard subtils. Les expériences montrent que DiT-Gaze atteint un nouvel état de l'art en qualité perceptuelle et en précision de redirection du regard, réduisant l'erreur de regard de 6,353 degrés et améliorant la précision de…

The paper presents DiT-Gaze, a new framework for high-fidelity gaze redirection, which enhances existing 3D Gaussian Splatting models like GazeGaussian. DiT-Gaze utilizes a combination of Diffusion Transformer, weak supervision across gaze angles, and an orthogonality constraint loss to improve the rendering of subtle gaze shifts. Experiments indicate that DiT-Gaze achieves a new state-of-the-art in perceptual quality and gaze redirection accuracy, reducing gaze error by 6.353 degrees and improving accuracy by 4.1%.

3D Gaussian and Diffusion-Based Gaze Redirection

One More Thing in AI – Your Shortcut to AI Mastery

3D Gaussian and Diffusion-Based Gaze Redirection

Was this article worth reading? Share it

One More Thing in AI

LucidQuery AI

Eyeye

Fakeface

Deptho.ai

GPTHumanizer

Ready to build your own newsroom?