iiinova — where your digital self keeps evolving

And here is a request from me: if you have a nanobind extension, could I ask you to test it in "split mode"? I would like this to cook for bit before making a regular release. Any feedback helps!

100

Changelog - nanobind documentation

Basically nanobind has adopted its own kind of "stable ABI". The new release also brings a ton of additional performance improvements and goodies. Please check out the full release notes at nanobind.readthedocs.io/en/latest/ch...

nanobind.readthedocs.io

Split mode cuts nanobind into a backend & frontend part. The backend tightly couples to each Python version and is available on PyPI (i.e., not your problem). Your extensions only contain the tiny frontend and become deployable on Python 3.10+ without downsides.

100

Python’s stable ABI was supposed to fix this but creates new issues: performance-critical parts of nanobind need access to Python internals for efficiency, which is forbidden by the stable ABI. It locks nanobind to the ancient Python 3.12 feature set and can't be used on < 3.12.

100

It tackles a pain point felt by many developers who distribute Python extensions with natively compiled parts: it's necessary to not only compile binaries ("wheels") per OS/processor (which kind of makes sense) but also for all Python minor versions (which is kind of crazy).

110

Wenzel Jakob @wjakob.bsky.social · 21/07/2026

I'm excited to announce a first development release of nanobind 3.0. It introduces *split mode*, a new distribution model where a single Python wheel runs across all Python versions, while achieving outstanding performance and gaining rather than losing features! 🧵

1248

This is so useful, thanks a lot Keenan!

000

Reposted by Wenzel Jakob

Delio Vicini @deliovicini.bsky.social · 26/06/2026

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. 🧵

1143

If your projects still on the old version, you will want to upgrade.

000

Changelog - nanobind documentation

I've released a nanobind 2.13 a few days ago. It improves performance on the critical path of Python bindings (call dispatch, array casting). In the few days that Fable 5 was available, I ran it nonstop to search for bugs, and it found some *nasty* stuff: nanobind.readthedocs.io/en/latest/ch...

nanobind.readthedocs.io

1120

Upgrading should be as simple as "pip install --upgrade drjit mitsuba". Give it a try and let us know how much faster your Mitsuba/Dr.Jit project now runs :-).

020

I want to thank Sébastien Speierer, Nicolas Roussel, Delio Vicini, and Philippe Weier, Tobias Jüterbock, and Thomas Auzinger, who all contributed to this release.

Release notes - Mitsuba 3

Mitsuba features faster envmap map sampling, 8-bit textures (finally!), and a new patch-based JIT for data structure conversion on Intel/ARM processors. The ray tracing core was rewritten to scratch to support so many different backends in a unified way. mitsuba.readthedocs.io/en/stable/re....

mitsuba.readthedocs.io

140

.. or evaluate them traditionally using tensors. We also added support for the Muon optimizer and made loops differentiable in reverse-mode. See drjit.readthedocs.io/en/stable/ch... for details.

drjit.readthedocs.io

Changelog -

100

Dr.Jit gains many tensor operations with NumPy-compatible semantics, including matrix multiplication, advanced indexing/slicing, stacking, reshaping, reductions, convolutions. A generalized neural network interface can fuse MLPs into computation using cooperative vectors, ...

110

Jacco Bikker 🎦 @jbikker.bsky.social · 21/06/2026

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.🧵

1318

Reposted by Wenzel Jakob

TinyBVH 1.7.2 (dev branch) now adds "Bring Your Own Threadpool" (thanks @wjakob.bsky.social!): Replace the built-in job system (adapted from Wicked Engine) by an alternative, using three hooks in BVHContext. TinyBVH remains single-header and dependency-free of course. github.com/jbikker/tiny...

github.com

GitHub - jbikker/tinybvh at dev

Single-header dependency-free BVH construction and traversal library. - jbikker/tinybvh

0295

Wenzel Jakob @wjakob.bsky.social · 27/11/2025

Quick reminder about the EPFL PhD program deadline (EDIC) on Dec 15.

042

Wenzel Jakob @wjakob.bsky.social · 22/08/2025

You can dump the PTX intermediate representation (see the documentation), but figuring out the calling convention of the kernel for your own use will be tricky. The system is not designed to be used in this way.

010

Wenzel Jakob @wjakob.bsky.social · 18/08/2025

Just write your solver in plain CUDA. How hard can it be? 😛

Wenzel Jakob @wjakob.bsky.social · 18/08/2025

This approach is restricted to software that only needs the CUDA driver. If your project uses cuSolver, you will likely need to have a dependency on the CUDA python package that ships this library on PyPI (similar to PyTorch et al.)

Markus Worchel @mworchel.eurosky.social · 12/08/2025

Reposted by Wenzel Jakob

Differentiable rendering has transformed graphics and 3D vision, but what about other fields? Our SIGGRAPH 2025 introduces misuka, the first fully-differentiable path tracer for acoustics.

18113

Wenzel Jakob @wjakob.bsky.social · 15/08/2025

Wasn’t that something.. Flocke (German for “flake”) says hi!

150

Wenzel Jakob @wjakob.bsky.social · 12/08/2025

If you are fitting a NeRF and you want a surface out at the end, you should probably be using the idea in this paper.

010

Wenzel Jakob @wjakob.bsky.social · 11/08/2025

Given the focus on performance, I would suggest to switch from pybind11 to nanobind. Should just be a tiny change 😇

150

For the paper and data, please check out the project page: mokumeproject.github.io

mokumeproject.github.io

Mokume

000

The Mokume project is a massive collaborative effort led by Maria Larsson at the University of Tokyo (w/Hodaka Yamaguchi, Ehsan Pajouheshgar, I-Chao Shen, Kenji Tojo, Chia-Ming Chang, Lars Hansson, Olof Broman, Takashi Ijiri, Ariel Shamir, and Takeo Igarashi).

100

To reconstruct their interior, we: 1️⃣Localize annual rings on cube faces 2️⃣ Optimize a procedural growth field that assigns an age to every 3D point (when that wood formed during the tree's life) 3️⃣ Synthesize detailed textures via procedural model or a neural cellular automaton

110

The Mokume dataset consists of 190 physical wood cubes from 17 species, each documented with: - High-res photos of all 6 faces - Annual ring annotations - Photos of slanted cuts for validation - CT scans revealing the true interior structure (for future use)

100

Wood textures are everywhere in graphics, but realistic texturing requires knowing what wood looks like throughout its volume, not just on the surfaces. The patterns depend on tree species, growth conditions, and where and how the wood was cut from the tree.

100

How can one reconstruct the complete 3D interior of a wood block using only photos of its surfaces? 🪵 At SIGGRAPH'25 (Thursday!), Maria Larsson will present *Mokume*: a dataset of 190 diverse wood samples and a pipeline that solves this inverse texturing challenge. 🧵👇

27615

My lab will be recruiting at all levels. PhD students, postdocs, and a research engineering position (worldwide for PhD/postdoc, EU candidates only for the engineering position). If you're at SIGGRAPH, I'd love to talk to you if you are interested in any of these.

0127

The reason is that the "volume" of this paper is always rendered as a surface (without alpha blending) during the optimization. Think of it as an end-to-end optimization that accounts for meshing, without actually meshing the object at each step.

Yining Karl Li @yiningkarlli.bsky.social · 08/08/2025

To get a triangle mesh out at the end, you will still need a meshing step (e.g. marching cubes). The key difference is that NeRF requires addtl. optimization & heuristics to create a volume that will ultimately produce a high quality surface. With this new method, it just works.

163

Reposted by Wenzel Jakob

Wow, this is such a cool paper! Basically with a surprisingly small modification to existing NeRF optimization, this paper gets a really good direct surface reconstruction technique that doesn't require all of the usual mess that meshing a NeRF requires (raymarching, marching cubes, etc).

1235

Check out our paper for more details at rgl.epfl.ch/publications...

091

Our method minimizes the expected loss, whereas NeRF optimizes the loss of the expectation. It generalizes deterministic surface evolution methods (e.g., NvDiffrec) and elegantly handles discontinuities. Future applications include physically based rendering and tomography.

241

Instead of blending colors along rays and supervising the resulting images, we project the training images into the scene to supervise the radiance field. Each point along a ray is treated as a surface candidate, independently optimized to match that ray's reference color.

100

By changing just a few lines of code, we can adapt existing NeRF frameworks for surface reconstruction. This patch shows the necessary changes to Instant NGP, which was originally designed for volume reconstruction.

Methods like NeRF and Gaussian Splats model the world as radioactive fog, rendered using alpha blending. This produces great results.. but are volumes the only way to get there?🤔 Our new SIGGRAPH'25 paper directly reconstructs surfaces without heuristics or regularizers.

310523

Release notes - Mitsuba 3

It also adds support for function freezing so that the process of rendering a scene can be captured and cheaply replayed. See mitsuba.readthedocs.io/en/latest/re... for details.

mitsuba.readthedocs.io

020

Release notes - Mitsuba 3

Mitsuba significantly improves performance on the OptiX backend and adopts SER (Shader Execution Reordering) throughout various integrators. It adds special shapes and an integrator for gaussian splatting, a sun-sky emitter, and fixes missing partial derivative terms in differentiable integrators.

mitsuba.readthedocs.io

Dr.Jit adds matrix operations that compile to tensor core (CUDA) or vector instruction sets like AVX512, neural network abstractions, grid/permutohedral encodings, function freezing, and shader execution reordering (SER). Many improvements simplify development. drjit.readthedocs.io/en/stable/ch...

drjit.readthedocs.io

Changelog -