University of Crete (UOC)
Radiance transfer prediction for real-time rendering using neural networks and geometric algebra
Abstract
dc:descriptionPrecomputed Radiance Transfer (PRT) has long been a cornerstone of real-time rendering, enabling high-fidelity global illumination under complex lighting. However, traditional PRT suffers from prohibitive precomputation costs, limited scalability to high-resolution meshes, and impracticality for dynamic or animated scenes--severely constraining its use in interactive applications. Three key challenges arise in particular: (i) the expensive computation of per-vertex spherical harmonic (SH) coefficients, (ii) the lack of adaptability to changing geometry and lighting, and (iii) the memory and latency constraints that hinder deployment in real-time pipelines. Recent advances in neural rendering and neural radiance field (NeRF) research have demonstrated the potential of deep learning networks to model complex light transport efficiently. Yet, these methods are typically restricted to image-space rendering and remain poorly suited for SH-based lighting representations or integration within mesh-based pipelines. In parallel, Conformal Geometric Algebra (CGA) has gained attention in computer graphics for its ability to represent geometric entities and transformations in a unified algebraic framework, offering a mathematically consistent way to encode spatial and lighting relationships directly on meshes. To address these limitations, we introduce Neural-GASh, a neural rendering framework that reformulates PRT within a geometric deep learning paradigm. By encoding vertex-normal pairs as CGA multivectors and utilizing neural networks to predict radiance transfer coefficients, we bypass the costly precomputation while maintaining physical interpretability. The contributions of this work are threefold. First, we show that CGA-based encodings improve training stability and expressiveness by utilizing multivectors that capture spatial and angular relationships more richly than Euclidean inputs. Second, we show that graph neural networks (GNNs), leveraging mesh connectivity, capture self-occlusion and shadowing effects that MLP-based surrogates fail to reproduce. Third, we show how to integrate the trained models into Unity for real-time rendering, for both static and animated assets. Experimental evaluation across diverse 3D meshes confirms that Neural-GASh reduces precomputation by orders of magnitude while preserving high-quality illumination, including complex shadowing. The framework scales to tens of thousands of vertices in real-time, and the Unity prototype demonstrates its feasibility for interactive use cases. In summary, this thesis establishes GNN-based PRT as a scalable and effective alternative to classical precomputation, laying groundwork for future neural rendering pipelines in graphics and virtual reality.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Geronikolakis, Efstratios
- Γερονικολάκης, Ευστράτιος
Subjects
dc:subject × 23- Precomputed Radiance Transfer
- Νευρωνική Απεικόνιση
- Νευρωνικά δίκτυα γράφων
- Σύμμορφη Γεωμετρική Άλγεβρα
- Σφαιρικές αρμονικές
- Γεωμετρική Βαθιά Μάθηση
- Ενσωμάτωση στην Unity
- Εικονική πραγματικότητα
- Neural Rendering
- Graph neural networks
- Conformal Geometric Algebra
- Spherical harmonics
- Geometric Deep Learning
- Unity Integration
- Virtual reality
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Γραφικά υπολογιστή και Σχεδιασμός με χρήση υπολογιστή
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
- Computer Graphics and Computer-Aided Design
- Artificial Intelligence
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60351