←
Bernardo Taveira94 posts
Bernardo Taveira avatar

Bernardo Taveira

@bertaveira

Industrial PhD student @ Chalmers and Zenseact Computer Vision, reconstruction and autonomous driving

◷ Joined October 2011▧ 10 Photos and videos
Bernardo Taveira avatar
Bernardo Taveira@bertaveira·Aug 31

📢📢📢 Introducing VoroTracing Differentiable Voronoi ray tracing that is faster than most rasterized splatting. 623 FPS on Mip-NeRF 360 (RTX 5090) → 3.2× the fastest prior ray tracer → 2.8× 3DGS Paper + Code + Web Viewer: research.zenseact.com/public…

0:08
Bernardo Taveira avatar
Bernardo Taveira@bertaveira·Nov 8

MojoSplat now lighter on memory supporting older hardware like an RTX 3080, it works on Apple Silicon macbooks and had some simple optimizations. Still no CUDA Toolkit involved, no multi-languages, or branched code. A single source of truth for all GPUs. Next: AMD testing?

Post attachment

3DGS now running without CUDA, without Python and with a single language for both training code and GPU kernels. Finally got around to update MojoSplat (using @Modular Mojo), matching @fhahlbohm Faster-GS on CUDA GPUs without even CUDA Toolkit installed github.com/bertaveira/mojosp…

0:20
Minseok avatar
⟳ Bernardo Taveira retweeted
Minseok@Minseok96_kr·Nov 7

Most fun public comment in ICLR2027. Bro reproductivity is important☠️☠️☠️ #ICLR2027

Post attachment
Bernardo Taveira avatar
Bernardo Taveira@bertaveira·Nov 6

After a few tweaks and improvements MojoSplat is now running natively on Apple Silicon! M1 Pro - 1h 03min RTX 5090 - 3min 25sec Not bad for this old laptop… same exact kernel code for all GPUs. No need to maintain multiple branches or paths

Post attachment

3DGS now running without CUDA, without Python and with a single language for both training code and GPU kernels. Finally got around to update MojoSplat (using @Modular Mojo), matching @fhahlbohm Faster-GS on CUDA GPUs without even CUDA Toolkit installed github.com/bertaveira/mojosp…

0:20
Bernardo Taveira avatar
Bernardo Taveira@bertaveira·Nov 6

Hey @Apple @AppleSupport , why do Metal developers need to install 10GB+ of Xcode just to compile GPU kernels? I don't need an IDE, simulators, or iOS SDKs. I need metal, metallib, and a compiler toolchain. Easire to install CUDA Toolkit, and that is a low bar...

Bernardo Taveira avatar
Bernardo Taveira@bertaveira·Nov 6

3DGS now running without CUDA, without Python and with a single language for both training code and GPU kernels. Finally got around to update MojoSplat (using @Modular Mojo), matching @fhahlbohm Faster-GS on CUDA GPUs without even CUDA Toolkit installed github.com/bertaveira/mojosp…

0:20
Bernardo Taveira avatar
Bernardo Taveira@bertaveira·Oct 31

Why are we still only evaluating reconstruction methods in object centric datasets? like (Mip-NeRF 360, DB...) Why did Zip-NeRF dataset never took off?

Bernardo Taveira avatar
Bernardo Taveira@bertaveira·Oct 21

The classic “it is incorrect but the metrics say it is better”. Very interesting to see how the popping artefacts can hide discrete changes between captures

Here's another example, where a book moved during the capture. NASGabor partly fits this via appearance, SH can only do this through popping. Thinking this through, methods that can't rely on popping, e.g., pop-free renderers or opaque surfaces, should benefit, right? (7/10)

0:21
Jun avatar
⟳ Bernardo Taveira retweeted
Jun@NSJunXiu·Oct 10

This video by Apple has aged like wine in today’s superintelligence era.  2013

1:30
Mistral AI avatar
⟳ Bernardo Taveira retweeted
Mistral AI@MistralAI·Oct 6

Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

0:17
Andrea Tagliasacchi avatar
⟳ Bernardo Taveira retweeted
Andrea Tagliasacchi@taiyasaki·Sep 8

What’s the point of the “Cvent Events” for #ECCV26 app if none of the workshop homepages and/or descriptions are up to date? 🤬🤬

Bernardo Taveira avatar
Bernardo Taveira@bertaveira·Sep 5

A few days until #ECCV2026 and Scholar-Inbox still hasn’t added the conference planner… So I threw all my tokens at the problem, scraped the conference website + arXiv, and built a graph of similar papers. Worked so well I decided to share: scholar-draftbox.com Enjoy 🎓

0:43
etn. avatar
⟳ Bernardo Taveira retweeted
etn.@etnshow·Sep 3

JUST IN: NVIDIA's $12,930,300,000 acquisition of Hugging Face contains an easter egg. The number 129,303 is the decimal conversion of Unicode point U+1F917. The 🤗 emoji.