I lead Systems Research & Engineering (Acceleration, Efficiency, Scale for ML/Research) at Luma AI, working on making large-scale training and inference as efficient as possible — from low-level kernel optimizations to architectural improvements that scale workloads to thousands of GPUs. Previously, I was a University Assistant and PhD student at Graz University of Technology, researching the intersection of real-time computer graphics and machine learning including neural rendering and texture-space shading for VR. In my spare time I like to develop games, play music, and contribute tools to the speedrunning and game modding communities.
Thomas Neff is Head of Systems Research & Engineering at Luma AI, where he leads efforts in ML acceleration, training efficiency, and inference optimization at scale — ranging from low-level CUDA optimizations to distributed training and inference systems spanning thousands of GPUs. He has contributed significantly to building and shipping Luma AI's flagship products including 3D Capture (NeRF/Gaussian Splatting) with an Unreal Engine 5 integration, Genie (text & image-to-3D), Dream Machine, Ray2, Ray3, and Uni-1 (multimodal understanding and generation). He gave a talk at the PyTorch Conference 2025 on how modern PyTorch supercharges multimodal training and inference at Luma AI.
Previously, Thomas was a University Assistant and PhD student at Graz University of Technology, Austria, as part of the GPU Computing group at the Institute of Computer Graphics and Vision. He (co-)authored several papers in the domains of real-time texture-space shading (TOG'19, TOG'21, CGF'22), neural rendering (CGF'21, ECCV'22) and medical image analysis (MICCAI'18), including a best paper award (OAGM'17).
PyTorch Conference 2025 — October 23, 2025 • San Francisco, CA
Covers how Luma AI designed training and inference code that scales across thousands of GPUs using modern PyTorch features including torch.compile, torch.distributed, and custom operations that accelerate multimodal workloads. Event page • Watch on YouTube
Large-scale multimodal generative video model training, architecture improvements, and efficient distributed inference at scale.
Large-scale multimodal training, inference, and optimization for Luma AI's unified understanding and generation model.
Contributions to Photon and Photon Flash, Luma AI's high-quality, cost-efficient image generation models.
Full-stack contributions to Dream Machine v1, Luma AI's original text-to-video product and the platform that hosts Ray and Photon today.
Research and engineering for Genie, Luma AI's generative 3D AI product turning text and images into interactive 3D assets.
Built and shipped Luma AI's 3D Capture product (NeRF/Gaussian Splatting), including design and implementation of the Unreal Engine 5 plugin for real-time NeRF/Gaussian Splat rendering.
A split rendering approach for ambient occlusion that offloads point-based AO queries to a server, enabling high-quality real-time AO on low-power clients. Published at High-Performance Graphics 2023 (Meta Reality Labs Research). Paper • Code
By predicting discrete depth values along each ray, we can reduce the required number of samples for neural raymarching by 24-64x without sacrificing quality or compactness.
We use a student-teacher distillation scheme to train a sampling network for efficient rendering of raymarching-based neural radiance fields.
Texture-space temporal reuse of full shading information, saving more than 57% of all shader invocations in VR without noticable loss in quality.
Efficient texture-space shading for virtual reality, allowing for high latencies compared to conventional streaming approaches.
Experiments with Wasserstein GANs to automatically generate medical image data for deep learning.
Procedurally generated Roguelike set in an engineering environment using Unity.
Colorful endless runner written in C for the Pebble Smartwatch.
Educational platformer with the intent to teach simple mathematics written in Objective C.
Third iteration of TileGame, also written in C#. Has metroidvania-style rooms and powerups to progress.
Second iteration of TileGame, also written in C#. Has scrolling levels, level editor, and better physics.
My first 2D platformer project, written in C#.
I love listening to music (mostly progressive metal and video game music) as well as composing my own songs. I've been playing guitar for a long time now, and I'm familiar with basic audio engineering and production.
Guitars:
Schecter BlackJack ATX C-1 ABS
Ibanez RGA7
Monitors:
Yamaha HS7
Audio-Technica ATH-M50 Headphones
Interface:
Focusrite Scarlett 2i2
Graz, Austria
thomasneff93@gmail.com