Lukas Sebastian Radl

PhD Student @tugraz. Radiance Fields, 3D, Real-Time Graphics.

[Lukas Sebastian Radl @rafourdl]

🚀 Excited to announce our latest research: “StopThePop: Sorted Gaussian Splatting for View-Consistent Real-Time Rendering”, which will appear in #SIGGRAPH2024! 🎉
Finally, fast and view-consistent rendering of 3D Gaussians, without popping artifacts! https://t.co/bsOHSn0peX

[Felix Windisch @FelixWindi38935]

Check out the code release for "A LoD of Gaussians: Out-of-Core Training and Rendering for Seamless
Ultra-Large Scene Reconstruction" (SIGGRPH 2026) now on Github:
https://t.co/z1JL6ovHRZ

[Lukas Sebastian Radl @rafourdl]

Checkout our new SIGGRAPH paper!

tl;dr: We introduce a method capable of training/rendering ultra-large scale scenes (100+ Million Gaussians) on consumer-grade GPUs without chunk-partitioning.

Congrats to @FelixWindi38935 (project lead), and team!
https://t.co/9xI0GbR6fr

[Felix Windisch @FelixWindi38935]

🧵 1/6 I am excited to announce that A LoD of Gaussians has been accepted to SIGGRAPH 2026! Our method uses out-of-core training directly on the LoD structure to achieve seamless training and rendering of massive 3DGS scenes on consumer GPUs.
Project Page: https://t.co/n5MX00Bg78

[Lukas Sebastian Radl @rafourdl]

Code is now finally out: https://t.co/FEInVD8Luw 🎉
Along with v2 on arXiv, now including some mesh-based NVS results, as well as more details, and clarifications for the normal variance loss: https://t.co/brp0e7mbD3

[Lukas Sebastian Radl @rafourdl]

🧵 1/6 Excited to announce our work on Confidence-Based Mesh Extraction from 3D Gaussians! 🚀
TLDR: We introduce a purely self-supervised confidence framework that extract high-fidelity unbounded meshes.
Check out the project page here: https://t.co/LD4mfXhU3L 🌐

[Lukas Sebastian Radl @rafourdl]

awesome stuff from @janusch_patas (yet again) - the decoupled D-SSIM loss in action, within Lichtfeld Studio!

[MrNeRF @janusch_patas]

I added the decoupled D-SSIM loss. It seems to help a lot with floaters and makes better geometry (video: left w, right w/o). It might be a bit worse on details. Compare the tries. I’m not sure yet, though, since it’s still experimental.

You’ll get it automatically when enabling either PPISP or the Bilateral Grid. Available in the upcoming nightly build!

Let me know how it goes.

[Radiance Fields @RadianceFields]

gaussian splatting meshing is taking another step forward

Confidence-Based Mesh Extraction from 3D Gaussians from @rafourdl et al: https://t.co/EKaVekmTdz

[Lukas Sebastian Radl @rafourdl]

This is likely also a good time to mention that this work builds on SOF, our first unbounded mesh extraction work we presented at SIGGRAPH Asia last year.
Project Page: https://t.co/DH1mzE3Fyy

[Lukas Sebastian Radl @rafourdl]

🧵 1/6 Excited to announce our work on Confidence-Based Mesh Extraction from 3D Gaussians! 🚀
TLDR: We introduce a purely self-supervised confidence framework that extract high-fidelity unbounded meshes.
Check out the project page here: https://t.co/LD4mfXhU3L 🌐

[Lukas Sebastian Radl @rafourdl]

Huge thanks to my amazing collaborators! Particularly noteworthy are Felix and Andreas (both X-less), without whom this project would not have been possible!
Honorable mentions to @Thomy2804, @steimich_tug, and my supervisor Markus Steinberger!

[MrNeRF @janusch_patas]

Confidence-Based Mesh Extraction from 3D Gaussians

Contributions:
– We introduce a self-supervised confidence framework to 3D Gaussian Splatting, along with densification adaptations, improving surface reconstruction

– We demonstrate how reducing variance in color- and normal blending resolves visual and geometric ambiguities, thereby improving the accuracy of surfaces

– Motivated by an analysis of photometric losses in 3DGS, we propose a novel decoupled appearance embedding, improving reconstruction in challenging real-world datasets

[Lukas Sebastian Radl @rafourdl]

Check out our new work on artifact-free Gaussian Rendering! #ICCV2025
Huge kudos to @steimich_tug and @Thomy2804, who led this project.

[Michael Steiner @steimich_tug]

(1/6) We are thrilled to announce that "AAA-Gaussians: Anti-Aliased and Artifact-Free 3D Gaussian Rendering" was accepted as a highlighted poster to #ICCV2025

TLDR: Enabling efficient training and rendering of 3DGS scenes without popping, distortion, and aliasing artifacts.

[Michael Steiner @steimich_tug]

(1/4) We are proud to present our PACMCGIT paper "Frustum Volume Caching for Accelerated NeRF Rendering" 🔥

We speed up real-time and offline rendering of NeRFs (e.g. Instant-NGP) by exploiting their inherent structure and reusing rendering information across multiple frames

[Lukas Sebastian Radl @rafourdl]

This was a cool project to be involved in - Neural Rendering, combined with caching, and highly optimized CUDA code for high quality offline rendering 🚀
@steimich_tug will be presenting this at HPG tomorrow, livestreamed on YouTube!

[Radiance Fields @RadianceFields]

Frustum Volume Caching from @steimich_tug, @Thomy2804, @rafourdl, and Markus Steinberger is pushing NeRF rendering rates into real time!

Also the code is out and MIT Licensed! Look how beautiful the videos are ✨

Article: https://t.co/llYjY0nBMb
Code: https://t.co/PcAk0S2SIq
Project: https://t.co/hfmFzkZt1L

[Lukas Sebastian Radl @rafourdl]

I'll be in Denver for #HPG2024 and #SIGGRAPH2024. At SIGGRAPH, we will present our work on StopThePop! 🚀 Looking forward to connecting with everyone there!

[Lukas Sebastian Radl @rafourdl]

I will be presenting our work on LAENeRF later today at #CVPR2024!
Come by and chat!

https://t.co/OwnTirv1yY

[Lukas Sebastian Radl @rafourdl]

Thanks to all the awesome collaborators! 🙏
@steimich_tug @MParger @AlexAUT_ @Snosixtytwo Markus Steinberger!
See you in Denver! #SIGGRAPH2024

[Lukas Sebastian Radl @rafourdl]

🚀 Excited to announce our latest research: “StopThePop: Sorted Gaussian Splatting for View-Consistent Real-Time Rendering”, which will appear in #SIGGRAPH2024! 🎉
Finally, fast and view-consistent rendering of 3D Gaussians, without popping artifacts! https://t.co/bsOHSn0peX

[Lukas Sebastian Radl @rafourdl]

How are we so fast with all this overhead? 🚀 We incorporate multiple optimizations such as tile-based culling, load balancing, etc. - and they can also be applied to vanilla 3DGS, which increases rendering speed by up to 2x. Our code is available here: https://t.co/tIm1ptA0yb.