Game development presentations - a useful reference

I honestly don't like the idea of training from a photo set of dubious quality cameras without separating and treating white point. They should have used raw data sets from homogeneous CMOS sources and settings. But yes, such approaches could be used offline by artist content creation pipelines to get less awful effects.

 

They did something like that as well, see here: https://intel-isl.github.io/PhotorealismEnhancement/

But to me this looks less realistic again. CG is always too perfect. The more imperfection we add, the more realistic, somehow.
Ofc. that's subjective, so i wonder if ML will allow more customization to personal preferences as well sometime...

 

I don't see this stuff as producing photorealistic graphics, it's just reshade++, IE with smarts, the parameters for "photorealism" are way too huge for a realtime neural net. But there are neat things that you can do with it, there's a ton of "style transfer" networks that are trained on non realistic art styles. I can see that stuff being really cool for certain art styles. Trying to figure out how to render abstract art or a van goph painting or whatever "properly" seems an absolute pain. Just being to render the basic stuff normally, and then throwing a realtime neural net on top of it to make it look like Spongebob or directly out of a Studio Ghibli film, seems right up the alley for this.

Link: https://medium.com/tensorflow/neura...ing-tf-keras-and-eager-execution-7d541ac31398

 

RTXDI presentation

 

https://schedule.gdconf.com/session/marvels-spider-man-miles-morales-a-technical-postmortem/878267

A Spiderman MM post mortem at GDC 2021 in July

https://schedule.gdconf.com/session/a-deep-dive-into-frostbites-hair-rendering-tech/878214

A frosbite hair rendering deep dive

https://schedule.gdconf.com/session...-fog-geometry-lighting-in-demons-souls/878999

Geometry and lighing in Demon's souls

Tons of good session

https://schedule.gdconf.com/sessions

https://schedule.gdconf.com/session...-procedural-grass-in-ghost-of-tsushima/879611

https://schedule.gdconf.com/session...d-war-a-scalable-approach-to-shadowing/879688

https://schedule.gdconf.com/session/character-technology-of-the-last-of-us-part-ii/878264

MAny others good presentation and I forget many others need to do another pass.

https://schedule.gdconf.com/session...musical-storytelling-in-cyberpunk-2077/878351

https://schedule.gdconf.com/session/emotional-systemic-facial-of-last-of-us-part-ii/878196

https://schedule.gdconf.com/session...e-techniques-in-the-last-of-us-part-ii/878333

https://schedule.gdconf.com/session...-reinforcement-learning-for-navigation/879035

https://schedule.gdconf.com/session...imation-generation-for-expressive-npcs/880503

https://schedule.gdconf.com/session...nd-game-balancing-with-project-chimera/879273

https://schedule.gdconf.com/session...to-3d-rotations-for-non-mathematicians/879197

EDIT:
https://schedule.gdconf.com/session/rope-simulation-in-uncharted-4-and-the-last-of-us-2/878186

https://schedule.gdconf.com/session/the-art-of-ori-and-the-will-of-the-wisps/878256

Loading on Ghost of Tsushima are incredible on PS4
https://schedule.gdconf.com/session/zen-of-streaming-building-and-loading-ghost-of-tsushima/878191

https://schedule.gdconf.com/session...the-ground-the-terrain-of-call-of-duty/879628

 

Pretty good

 

https://wickedengine.net/2021/06/07/future-geometry-pipeline/

Great blog blog post about geometry pipeline

 

June 8, 2021

 

https://momentsingraphics.de/Siggraph2021.html

BRDF importance sampling for polygonal lights

 

Motion matching is definitely neat. In the back of my head I'm wondering if one can author a collection of hand animated sequences and use a motion matching graph/ai generated graph to pick and blend between them; so you get the same production benefits of not having to hand craft a bunch of transitions, but now you get the control and art style of hand authored cycles as well.

 

NVIDIA Research: An Unbiased Ray-Marching Transmittance Estimator

 

June 21, 2021

 

Couple of cool ones, especially the real-time neural radiance cachine. It's a NN that trains itself in real-time for inference.

Nvidia ReSTIR GI: Path Resampling for Real-Time Path Tracing
Real-time Neural Radiance Caching for Path Tracing

 

GTC presentation: https://www.jannovak.info/publications/NRC/NRC_GTC_slides.mp4
Interactive results: https://tom94.net/data/publications/mueller21realtime/interactive-viewer/
Framework: https://github.com/NVlabs/tiny-cuda-nn

Overall, the results aren't all that convincing until 64spp is used where we see a noticeable improvement in half of the scenes ...

 

NVIDIA Research: NeRF-Tex: Neural Reflectance Field Textures

 

I think it makes some really nice in-roads in those areas of the image at 1spp where the path trace + restir merely returns dark + salt and pepper. Or even DDGI for that matter.