Fully fused Zip-NeRF hits 6-10x speedup: 36-min training, 300ms per megapixel on RTX 5090
jon_barron · x · 2026-11-01
- Developer @fhahlbohm applied the JIT-fusion approach Thomas used for Instant-NGP to Zip-NeRF, with help from Opus, producing a fully fused implementation: 36-minute training and 300ms per megapixel rendering on an RTX 5090, matching official code quality (-0.1 PSNR) at roughly 6-10x the speed.
- Further optimization reportedly pushes toward 10fps rendering and 20-minute training via more optimization plus a smaller network, at only -0.1 PSNR.
- Code is open-sourced; Jon Barron, co-author of the original Zip-NeRF, reposted it.
Related event: Fully Fused CUDA Zip-NeRF Trains 6-7x Faster, Renders 10x Faster(3 posts)→
More from Multimodal
- 267-word prompt generates epic action sequence, creator shares the full prompt — umesh_ai · 2026-11-01
- Kimi autonomously used a 4x3090 rig and YuE2 to score its own 45-second launch film — minchoi · 2026-11-01
- Ghibli-style hand-drawn 3D game built in one day: Kimi models assets via Blender Python — minchoi · 2026-11-01
- Kimi 3D Make turns one reference image into a 3D model in ~2.5 minutes — minchoi · 2026-11-01
- A Rolex Day-Date built in pure Three.js code via Kimi — 32 minutes, $6.25, no imported models — minchoi · 2026-11-01
- One prompt to real 3D: Kimi drives Blender, CAD and Three.js in 10 wild community demos — minchoi · 2026-11-01