Happy 2025! Earlier last year I stumbled upon NVIDIA Warp, a differentiable, kernel-based framework that lets you write CUDA-like code in python. I've been making a new NeRF app with it and have found it to be extremely performant and fun to use! Enjoy this short training video. This is Warp + PyTorch + fVDB + tcnn. Enjoy!
Happy 2025! Earlier last year I stumbled upon NVIDIA Warp, a differentiable, kernel-based framework that lets you write CUDA-like code in python. I've been making a new NeRF app with it and have found it to be extremely performant and fun to use! Enjoy this short training video. This is Warp + PyTorch + fVDB + tcnn. Enjoy!
@BenR82225 Splatting gaussians offers better rendering framerates for the most common hardware currently, that is the benefit. Besides that, NeRFs tend to take up less memory and can have higher realism
This is a Foliage-emulating Radiance Network, or FeRN for short. Ok just kidding it's just a NeRF of a fern I spotted on a hike but it's my post for the day. Enjoy. #WarpNeRF
@natanielruizg I had this issue too and after manually using the "I don't want to see this" option on a bunch of posts I'm back to academic research feed!
Siren vibes. NeRF quality improvements are on hold for now while I work on the utterly ridiculous idea of implementing a remote websocket renderer for use in a custom Blender addon. Will it work? Probably! Will it be fast? Hell no! #WarpNeRF
@mazy1998 Yes! I haven’t announced an official release yet, but the code is visible at https://t.co/LcvZcCk1D5
Be forewarned - it is entirely undocumented.
Balloon Dog by Jeff Koons at the Broad Museum, Los Angeles. Continuing with static images until I can get camera movements working in the Blender addon/renderer.
NeRF for now, because the raymarching algorithm has a ton of untapped artistic potential to create some really unique stuff. I’ve worked with gaussians and they have a lot of benefits over NeRF, also I’d like to experiment with raytracing them, but splatting them isn’t too useful to my goals.
#WarpNeRF - Posting a NeRF a day until I get a job😄 10k steps usually trains in under 60s @ 4096 rays/sec (largely thanks to tiny-cuda-nn) and this can be improved much further. Currently working on a Blender addon to control the camera, then will fix the image quality :]
Portrait of two fashionable desert nomads, trained in WarpNeRF. Without a contraction function, background collapse causes blurriness in scenes with a large amount of depth.