Gijs Fiten

About

I'm a PhD researcher at KU Leuven working on generative AI for extended reality, advised by Prof. Maria Torres Vega. My focus is Gaussian splatting, neural rendering, and 3D scene understanding for multi-user XR — I care most about building systems that actually run in real time on real hardware, not just about the numbers in a paper.

Before and alongside the PhD, I've worked across the stack from low-level rendering to the ML models that drive it, and I bring that engineer-first mindset to research: prototype fast, measure honestly, and ship things other people can build on.

Publications

First author

1. GSI-GEN: Gaussian Splats for Resolution Invariant Image Generation

   Gijs Fiten, Jit Chatterjee, Maria Torres Vega

   Under review, WACV 2027

2. Edge-XReal: Distributed Holistic Scene Understanding for Edge XR Devices

   Gijs Fiten, S. Winant, Jit Chatterjee, Maria Torres Vega

   Under review, Springer Multimedia Systems

3. On the Impact of the Embodiment of AI-agents on Active Engagement in Virtual Reality Experiences

   Gijs Fiten, Reaven Jansen, Shubham Selleri, Jit Chatterjee, Javad Sameri, Maria Torres Vega

   ACM ISIM '26 (proceedings forthcoming)

4. Redirected Walking for Multi-User eXtended Reality Experiences with Confined Physical Spaces

   Best Student Paper Award

   Gijs Fiten, Jit Chatterjee, Kobe Vanhaeren, Mattis Martens, Maria Torres Vega

   QoMEX 2025

   Image placeholder

   Paper Code

Co-author

1. Pose-Aware Multi-View Diffusion for Consistent 3D Texture Generation from a Single RGB Image

   Jierong Wen, Jit Chatterjee, Gijs Fiten, Maria Torres Vega

   GRIVAPP 2026

   Paper

2. GGD-Tex: Geometry-Guided Diffusion for 3D Texture Synthesis from Single RGB Image

   Jit Chatterjee, Gijs Fiten, Maria Torres Vega

   VISAPP 2026

   Paper

Projects

• GSI-GEN

  A feed-forward VAE that generates 2D Gaussian Splats directly, instead of pixels — a resolution-invariant reconstruction method, not a generation-quality play. 9.21M parameters, 6.1–14.5× smaller than pixel-based baselines, with an open-sourced 70k-image Gaussian Splat conversion of FFHQ.

  Image placeholder

  • Gaussian Splatting
  • Feed-Forward VAE
  • Resolution-Invariant

  Trained on the VSC (Flemish Supercomputer): the paper's main run used 4×H100 GPUs for just over a day; scaling tests ran on 8×B200 for a few hours — full VRAM utilization throughout.

• Edge-XReal

  A distributed architecture that offloads 3D scene-understanding models from edge XR headsets to a remote server, cutting processing time by 40–48%. Deployed as a working proof-of-concept on the Meta Quest 3.

  Image placeholder

  • Edge Computing
  • Scene Understanding
  • Meta Quest 3

  Originally trained via transfer learning on a single GTX 1070 over a few days, as part of the master's thesis that seeded this project — later work moved onto the VSC (Flemish Supercomputer).

  Code

• Live demo pipeline

  A small image classifier running on the on-demand GPU serving pipeline that will host GSI-GEN and Edge-XReal demos once those models are ready to share publicly.

  Image placeholder

  • Deployment
  • PyTorch
  • FastAPI

  Demo

CV

Download CV (PDF)