Ben Beilharz

Ben Beilharz

@engramm.ing

Cognitive Science PhD student working on (inverse) light transport for sensing, planning & acting @tsawallis.bsky.social's Perception Lab. AI for science. I like my models uncertain. he/him

297 followers529 following252 posts

Ben Beilharz
@engramm.ing

If you are interested in MRD, I am editing a blog post (will probably be updated a couple of times): engramm.ing/journal/rese... This will more casually outline the challenges and ideas of the metamer work. I am happy for people to chime in and discuss these things. 2/2

engramm.ingMRD: Using Physically Based Differentiable Rendering to Probe Vision Models for 3D Scene Understanding・エングラムWhile deep learning methods have achieved impressive success in many vision benchmarks, it remains difficult to understand and explain the representations and decisions of these models. Though vision ...
Ben Beilharz
@engramm.ing

Good morning @ecvp.bsky.social. Sadly, the conference is already nearing its end, but @lschmittwilken.bsky.social still has something nice for you! So now you have a date: 10:30 in the Active Vision and Visual Salience session (Bayview Suite). Have a great day everyone! :) #ecvp2026

Tom Wallis@tsawallis.bsky.social

TU Darmstadt's Perception Lab is at #ECVP2026 in force. We hope you can make it to see our work! @engramm.ing @lreining.bsky.social @lschmittwilken.bsky.social

Ben Beilharz
@engramm.ing

G'day @ecvp.bsky.social, it's time for another treat: Lina will present her poster on "Causality perception under visual uncertainty" at W1AM2. In the afternoon, I will be giving my first in-person talk in the Material Perception session about differentiable rendering in vision science. #ECVP2026

Tom Wallis@tsawallis.bsky.social

TU Darmstadt's Perception Lab is at #ECVP2026 in force. We hope you can make it to see our work! @engramm.ing @lreining.bsky.social @lschmittwilken.bsky.social

Ben Beilharz
@engramm.ing

Hey @ecvp.bsky.social we got twice the fun for you today (order of visit recommended): T5AM6 Swantje will talk about Dead Leaves stimuli with which you can model natural image statistics. T5AM1 @lreining.bsky.social will talk about how to segment these dead leaves, testing NNs for optimality.

Tom Wallis@tsawallis.bsky.social

TU Darmstadt's Perception Lab is at #ECVP2026 in force. We hope you can make it to see our work! @engramm.ing @lreining.bsky.social @lschmittwilken.bsky.social

Ben Beilharz
@engramm.ing

goooooood mornin‘ @ecvp.bsky.social! If you are an early bird and you like to see something and talk about multi focal display, or just want to know how to easily implement your experiments on one, go ahead and pay Tarek a visit at M3AM1 :) #ecvp26

Tom Wallis@tsawallis.bsky.social

TU Darmstadt's Perception Lab is at #ECVP2026 in force. We hope you can make it to see our work! @engramm.ing @lreining.bsky.social @lschmittwilken.bsky.social

Ben Beilharz reposted

ECVP 2026
@ecvp.bsky.social

🎉 TODAY’S THE DAY! 🎉 #ECVP2026 begins today in Bournemouth! 👁️🧠 We’re ready for a fantastic week of vision science, workshops, talks, symposia, posters and plenty of opportunities to connect with the ECVP community. Welcome to ECVP 2026 — let’s get started! 🚀 #VisionScience #Perception

Ben Beilharz reposted

Lars C. Reining
@lreining.bsky.social

Excited this is finally out! I am happy to have built this together with Rabea Turon, @tsawallis.bsky.social and all the others!🎉

Tom Wallis@tsawallis.bsky.social

New preprint! Have you ever wanted to measure psychometric functions in high-dimensional stimulus spaces, but realised that this is infeasible? We present and validate a technique to adaptively estimate psychometric functions for stimulus spaces up to 50 dimensions. www.biorxiv.org/content/10.6...

biorxiv.orgAdaptive experiments in high-dimensional feature spaces: A particle filtering approachBehavioral experiments are often infeasible when stimulus spaces have many dimensions or when testing time is limited. One way to address this challenge is adaptive stimulus selection, where informati...
Ben Beilharz
@engramm.ing

Happy to announce that MRD got accepted by the Journal of Vision (marks the first submission of my phd). Thanks to all editors and reviewers making the paper better over time. If you don't want to wait for the official JoV version, I updated the preprint & repo with the latest code and results.

Ben Beilharz@engramm.ing

Sorry for the wait (been to China for 4 weeks between the years), but I finally managed to update the pre-print with more results and the code is now also available on: github.com/ag-perceptio... If you run into issues, let me know.

github.comGitHub - ag-perception-wallis-lab/MRD: Accompanying repository for MRD (Metamers Rendered Differentiably)Accompanying repository for MRD (Metamers Rendered Differentiably) - ag-perception-wallis-lab/MRD
Ben Beilharz
@engramm.ing

It is fucking horrendous what happened today at the Berlin Pride Parade. Kinda sucks that hate progresses quicker than tolerance and acceptance. Currently only German news: www.tagesschau.de/inland/gesel... TL;DR some sicko drove into a group of people after the demonstration. 1 dead person.

tagesschau.deBerliner CSD abgebrochen - mutmaßlich Fahrzeug in Menschenmenge gefahrenDer Berliner Christopher Street Day ist abgebrochen worden. Laut Polizei ist mutmaßlich ein Fahrzeug in eine Menschenmenge gefahren, mehrere Menschen seien verletzt.
Ben Beilharz
@engramm.ing

After the neural material/shading talks at EGSR, I was wondering if there's a need for a package to facilitate this type of research. It would be nice to have a package with reusable building blocks and a set of reimplementations to benchmark against. This is currently what I am working on. 1/5

View full thread
Ben Beilharz
@engramm.ing

I've implemented a couple of models I found interesting. The idea would be to have people filing PRs with their experiments to gather a library of neural material approaches in which we can easily benchmark and evaluate different approaches. 4/5

Ben Beilharz
@engramm.ing

Disclaimer: I wrote the code, Claude the documentation and tests. I want to do another pass on it, which is why I want to postpone the release a bit. 5/5

Ben Beilharz
@engramm.ing

After the neural material/shading talks at EGSR, I was wondering if there's a need for a package to facilitate this type of research. It would be nice to have a package with reusable building blocks and a set of reimplementations to benchmark against. This is currently what I am working on. 1/5

Ben Beilharz
@engramm.ing

It is not ready yet, but what I have so far: - Common helpers to tabulate and reparameterise BSDF data - Loaders for MERL and RGL's data - Pipelines, losses, and evaluations - A few models - Adapters to use the fitted BSDFs (Mitsuba; coopvec). - RTC-like neural texture compression. 2/5

Ben Beilharz
@engramm.ing

Just wrapped up implementing ボンド's path space API. Previously, I had a loop that constructed the path incrementally without storing the vertices. Now, the whole path is created at once and then evaluated. While this is still a unidirectional case, it would enable me to implement BDPT next.

Ben Beilharz
@engramm.ing

Just seeing that @guerrilla.bsky.social has a new tab in their nav bar pointing to the Decima engine: www.guerrilla-games.com/decima I don't want to hypothesise, but if that means that the engine is being released at one point, we could see serious competition on the market for UE, Unity and Godot.

guerrilla-games.comWe create awe-inspiring worlds - GuerrillaGuerrilla is one of Europe's leading game companies and a wholly-owned subsidiary of Sony Interactive Entertainment Europe. We started in 2003, and have pushed the boundaries of technical and artistic...
Ben Beilharz
@engramm.ing

#EGSR26 has come to an end, and I am sad it is already over. Tons of interesting chats, nice, approachable people, and overall just a great experience. Thanks to the organisers for making this possible; looking forward to next year in Lugano! My highlights: 1/5

View full thread
Ben Beilharz
@engramm.ing

* Mirco's Optimized and Aligned Anisotropic Monte Carlo Sampling Patterns: diglib.eg.org/items/3ba69a... * My personal favourite (and best paper awarded): Elie's Volumetric Inverse Rendering via Neural Radiative Transfer diglib.eg.org/items/dcda7e... 5/5

diglib.eg.orgVolumetric Inverse Rendering via Neural Radiative TransferVolumetric inverse rendering seeks to recover the optical properties of participating media from images. Existing approaches either rely on differentiable stochastic light transport simulation, which ...
Ben Beilharz
@engramm.ing

#EGSR26 has come to an end, and I am sad it is already over. Tons of interesting chats, nice, approachable people, and overall just a great experience. Thanks to the organisers for making this possible; looking forward to next year in Lugano! My highlights: 1/5

Ben Beilharz
@engramm.ing

* Song's Conditional Product NEE for Gaussian Process Implicit Surfaces: diglib.eg.org/items/732105... * Liliana's Keynote about Avian colouration: egsr2026.inria.fr/keynotes/#alba * Louis' A Hybrid Neural-Microfacet BRDF Model for Real-Time Rendering: diglib.eg.org/items/add38f... 3/5

diglib.eg.orgConditional Product Next Event Estimation for Gaussian Process Implicit SurfacesWe propose a new next-event estimation (NEE) technique for Gaussian Process Implicit Surfaces (GPISes). We show that the distribution of surface normals available for sampling at a GPIS hit point coll...

Ben Beilharz reposted

Godot Engine
@godotengine.org

For years we have been overwhelmed by the volume of code contributions we have to review, especially from new contributors. AI has made the problem much worse. That's why we are taking steps to reduce the burden on maintainers while still welcoming new contributors: godotengine.org/article/cont...

godotengine.orgChanges to our Contribution Policies – Godot EngineGodot is growing fast, so here is how we are dealing with the huge increase in contributions.

Ben Beilharz reposted

Delio Vicini
@deliovicini.bsky.social

I'm excited to share the open-source release of hdMitsuba, a USD Hydra delegate for the Mitsuba 3 renderer. By leveraging the Hydra API, this project brings Mitsuba directly into OpenUSD workflows. The GIF below shows interactive rendering within usdview. 🧵

Ben Beilharz reposted

Wenzel Jakob
@wjakob.bsky.social

I am excited to announce Dr.Jit 1.4 and Mitsuba 3.9 which now feature GPU-accelerated compute/ray tracing on Apple Silicon. Significant optimizations in both projects accelerate tracing/code generation ~2x. They generate better code and have faster Python bindings.🧵

Ben Beilharz
@engramm.ing

embree goes brrrr... (so yes, just refactored my acceleration structure into its own plugin, so I can easily swap out different backends).

Ben Beilharz
@engramm.ing

Apparently, for smaller scenes, my SAH BVH holds up, but the higher the poly count, the more Embree smashes my prior implementation. Embree: 344.61s user 2.25s system 988% cpu 35.091 total SAH BVH: 436.40s user 2.21s system 928% cpu 47.215 total

Ben Beilharz
@engramm.ing

One nice thing about having the interactive part of the renderer done is having fun with materials. Implemented a couple of diffuse models recently: Lambert, Oren Nayar, EON, d'Eon 21 fast, d'Eon 21 full model

Ben Beilharz
@engramm.ing

Aaaand references: Oren, M. and Nayar, S.K. 1994. Generalization of Lambert’s reflectance model. Portsmouth, Kutz, and Hill 2024. EON: A practical energy-preserving rough diffuse BRDF. d'Eon 2021. An analytic BRDF for materials with spherical Lambertian scatterers

Ben Beilharz
@engramm.ing

One nice thing about having the interactive part of the renderer done is having fun with materials. Implemented a couple of diffuse models recently: Lambert, Oren Nayar, EON, d'Eon 21 fast, d'Eon 21 full model