Have you considered Alembic? I'd be curious to know if/why you discarded it, as sharing dynamic meshes is its core use-case, and it is supported in most 3D programs AFAIK. It serves a different purpose than OpenUSD, tailored towards "baked" representations rather than "editable".

Thanks, I'm glad you liked it!

As a result, our method consistently reduces the build time although this comes at the cost of an increase in tracing time. We show that this tradeoff is worth it for real-time raytracing of complex dynamic scenes since rebuilds are particularly expensive. Check out the paper for more details!

In this paper we focus on maximizing build speed by removing this additional traversal. To do so we introduce a bottom-up collapsing algorithm, which we integrate into an existing high-performance BVH2 builder to obtain a fast wide BVH construction algorithm.

However most of the GPU BVH construction litterature focuses on binary hierarchies.
A notable exception is H-PLOC by Benthin et al which uses a binary builder and a separate top-down collapsing step to convert the binary BVH into a wide BVH.

The standard acceleration structure for ray-tracing is the wide BVH, and fast GPU construction of these hierarchies is a key challenge for realtime raytracing. This is especially important for complex dynamic scenes where BVH rebuilds are a major bottleneck. Example: Flooded Sponza by Intel.

How do you get AABBs for the skinned clusters without transforming all the geometry? As the AABB also needs to bound the finer LODs this doesn't seem trivial. And transforming the bounding box is not conservative as the cluster weights vary. Would love to have more details on your solution for this!

Thanks! Glad you like it, your paper was a big inspiration for this project.

Oops small mistake, I mistyped @axelparis.bsky.social

These pruned trees enable efficient evaluation of the SDF for rendering or discretization, resulting in >100x speedups for complex scenes compared to brute-force evaluation. Come to my talk next week for more details, or read the paper here: wbrbr.org/publications...

Our algorithm takes as input an analytical SDF represented as a tree of smooth booleans and a 3D region of space, and computes a much smaller pruned tree equivalent within the region. This procedure is simple and efficient, using the Lipschitz prop. of the SDFs instead of costly interval arithmetic

Genuine question, why do you want to use Nanite if not for the performance ? With fully dynamic runtime procgen you can't escape the O(n) rendering cost which is the whole point of the Nanite LoD system IMO. Or is it the software rasterizer that you'd like to use with your geometry ?

Is this for CPU ray-tracing ? In my experience for GPU ray-tracing compressed wide BVHs do perform better (see CWBVH paper).

Are there any talks about their engine that you'd recommend (free or vaulted) ? I haven't heard much about it so I'm interested.