OIDN Tensorflow update: Polished CPU side data handling for various types Spoiler alert, it's slow Getting data from three off the GPU takes ~100ms Running the denoiser takes ~4ms Pulling the TF tensor ~1600ms Processing & pushing back to GPU ~100ms Time to resolve GPU sharing

Excited to launch react-three/xr🤳🥽 Leverage the latest WebXR features, reuse your existing interactions, and advance what is possible in AR and VR on the Web. ↳ Turn your existing R3F app into an AR or VR experience with only a few lines of code. 🎉

AO / indirect (no denoiser; IGN pictured)

Realized I was blending it wrong. Here's with AgX.

Fixes + 512spp since I don't denoise yet. Need to get back down to 16spp. GTAO was already memory bound where Hi-Z greatly helps (unfortunately awful to implement in WebGL). Not sure how to include diffuse, starting with including it in the (depth) reduction anyway.

GTAO + visibility masks. Single pass at full resolution, no Hi-Z or denoiser yet.

Made it to SIGGRAPH.

InstancedBatchedSkinnedMesh with GPU mixer. jsfiddle.net/cbenn/Las0poyu

Smoke machines weren’t such a good idea.

Reminder this was reverted in error, and I put a $1k bounty to get through Meta's red tape and reinstate it. This is strictly an enhancement and has no performance ramifications for future/existing apps; instead of rebinding to an unrenderable canvas, it binds to the XR layer.

Implemented reverse depth in three.js with EXT_clip_control for [0, 1] NDC range. This is a strict improvement to logarithmic depth buffer where supported, namely in both performance and accuracy. Also works well with MSAA, unlike logarithmic depth. github.com/mrdoob/three.j…