Florian Hahlbohm

PhD student @tuBraunschweig, Computer Graphics Lab
Bsky:

[Florian Hahlbohm @fhahlbohm]

2.5 years, >7.7k citations. 3D Gaussian Splatting follow-ups are plenty.
Our new paper shows how far these works can advance the original by combining them in a refactored, performance-optimized implementation to speed up future research.

Paper + code: https://t.co/NVV5AZCALx

[Florian Hahlbohm @fhahlbohm]

@3DVconf You might want to add the abstract deadline here.

[Florian Hahlbohm @fhahlbohm]

@benhouston3d Really cool stuff, have you tried to compute the per-frame color in a prepass that doesn't run the SH evaluation 4x per splat? That was a lot faster for me on low end GPUs like my (shitty) phone.

[Florian Hahlbohm @fhahlbohm]

Radiance field meetup/dinner anyone? 👀

[European Conference on Computer Vision #ECCV2026 @eccvconf]

#ECCV2026 encourages community-organized social events! 🎉

Whether it’s a dinner, tutorial, or casual meetup, we’ll help promote your event to conference attendees.

Interested in organizing an event or want to see what’s already planned? Check out:
https://t.co/vJiwNfScOH

[Florian Hahlbohm @fhahlbohm]

@slimbuck7 Nice! :)

[Jorge Condor @Arcanous98]

Our work with @nvidia, Neural Harmonic Textures, has been accepted as an ORAL in ECCV'26! To celebrate, we bring fully fused kernels, up to 50% faster than before! Plus other new features, check them here https://t.co/vPqakTnecK

If you work on high quality #3DGS give it a try!

[Vincent Woo @fulligin]

I GOT GAUSSIAN SPLATTING ON THE LOCAL TV NEWS!

[Vanessa Wirth @vwirth_]

(1/4) What lies beneath an object’s visible surface—and what does an imaging radar perceive differently from an optical sensor? 👀

🎉 Thrilled that our paper MAROON was accepted to ACM ToG and excited to present it at @siggraph 2026!
#SIGGRAPH2026

[Keenan Crane @keenanisalive]

For #SIGGRAPH2026 attendees:

I made the webpage that we should have for the Technical Papers schedule, that lets you search, filter, and scroll through linearly (rather than opening a zillion tabs):

https://t.co/u20ecSvhfI

Also provides a flat JSON & CSV of all talks & times.

[Florian Hahlbohm @fhahlbohm]

@keenanisalive @eccvconf, this would be nice to have!

[Jorge Condor @Arcanous98]

Also, I wanted to mention that as I approach my graduation next year, I am slowly starting to consider my next steps. I've been talking with people over the last few months, but if you have some interesting project in academia or industry where you think I could fit, let's talk.

[Jorge Condor @Arcanous98]

Today we are releasing the code for Gabor Fields https://t.co/I9OgajCSrv. A new way to understand volume rendering with primitives (think Gaussians, but different!), that allows us to do principled LOD, extreme acceleration (\w control variate theory) and much more #SIGGRAPH26

[High-Perf Graphics @HPG_Conf]

🎬 #HPG2026 Day 1 is LIVE!

🎤 Keynote: "Complex Models, Simple Solutions" — Patrick Cozzi & Sean Lilley
⚡ Hot3D: neural.slang · Papers: Ray Tracing & Denoising

📺 https://t.co/SV40RAUgFf

#HighPerformanceGraphics #GPU #RayTracing

[OnTheFly @onthefly3d]

3D for everyone, everywhere.
Today, we’re proud to introduce OnTheFly, a GraphDeco spin-off from @Inria founded by @GDrettakis, @AndreasMeuleman, @cammonteCG and Anthony Schoofs.
Our goal: put high-quality 3D capture in the palm of your hand.
https://t.co/spyKgbF1Ij
1/5

[Ishaan Shah @ishaan_n_shah]

We are excited to announce our work on efficient precomputation for real-time global illumination, presented at EGSR'26 last week!
Learn more: https://t.co/P6zMuq2H7M.

This work was conducted with @SimonLUCAS_CG and @GDrettakis.
Funded by the ERC Advanced Grant NERPHYS.

[Florian Hahlbohm @fhahlbohm]

@iquilezles Haven't fully watched the video but I find it quite interesting that the formula you derived is effectively Squareplus(x - y, lambda), where Squareplus is @jon_barron's alternative to Softplus from https://t.co/ssdAaIKD8e

[Florian Hahlbohm @fhahlbohm]

It also borrows from our own Faster-GS, and while I built it inside NeRFICG, the rasterizer can of course be built standalone as a PyTorch extension too.

Go give it a try and let me know what you think. I was personally surprised how fast rendering converges on a 5090.

[Florian Hahlbohm @fhahlbohm]

PSA: I built a small CUDA project with two sort-free, stochastic 3DGS renderers with zero VRAM overhead. 250M Gaussians? Easy!

No paper attached from my side. I built it to better understand the advantages of stochastic transparency.

Code + write-up: https://t.co/PHkzBKjSXI 🧵

[Florian Hahlbohm @fhahlbohm]

A resolve pass then does the SH evaluation for the single winning Gaussian per pixel. Limitations on GitHub.

Of course, the stochastic rendering requires averaging multiple samples per pixel. When doing that, it converges exactly to the result produced by the 3DGS rasterizer.