65K views · Dec 17, 2017 · 4:17 · 2 HN recommenders · 5 mentions · discussed 2018–2019
A Two Minute Papers episode on NVIDIA's recurrent denoising autoencoder, which cleans up noisy Monte Carlo path-traced image sequences. It shows how a neural network can produce usable renders from very few samples.
Why Hacker News points people here: Cited as an accessible explanation of AI denoising, which underpins real-time ray tracing on RTX hardware.
“Two Minute Papers did a nice explanation of it”HN comment, 2019
“Disney actually does a lot of research combining the world of graphics with deep learning.”
Neural noise filtering for photorealistic rendered video.
Disney actually does a lot of research combining the world of graphics with deep learning. Some examples that you might appreciate (from the excellent channel "Two Minute Papers"): . "Disney's AI Learns To Render Clouds" [0] . "AI Learns Noise Filtering For Photorealistic Videos" [1] [0] link [1] this video
“Two Minute Papers did a nice explanation of it”
Nice explanation of NVIDIA's recurrent denoising autoencoder.
It's only briefly mentioned, but NVIDIA's recurrent denoising autoencoder is pretty amazing. link Two Minute Papers did a nice explanation of it: this video
“I don't think your estimations take into account the massive gains from AI denoising techniques such as”
I don't think your estimations take into account the massive gains from AI denoising techniques such as this video
“This uses a neural network”
Neural network denoising of path-traced results, results from 1:15.
This uses a neural network this video video results start around 1:15 This one does not use a neural network: link
“It's what allows RTX to produce meaningful images at all despite the renderer running at 0.5-2 samples per pixel”
Shows deep learning denoising that makes low-sample ray tracing usable.
Deep learning denoiser. I don't know if NVIDIA came up with some marketing term for it. It's what allows RTX to produce meaningful images at all despite the renderer running at 0.5-2 samples per pixel which looks little better than random noise. I kind of assumed they'd call it DLDN because they called their deep learning supersampling DLSS. I guess not, go figure. this video
