Conversation

Wild paper synthesizes photorealistic imagery for corresponding video game footage, roughly via a generator informed by extracted semantic scene info and a discriminator minimizing differences vs. photographic dataset. intel-isl.github.io/PhotorealismEn

0:07

8.9K views

189

Andy Matuschak

@andy_matuschak

This summary by

@twominutepapers

gives a nice overview with lots of comparisons. Also suggests an interesting possible future for this work—that photorealistic game engines could become very simple, emitting coarse scene info rather than complete renders.

youtube.com

Intel's Video Game Looks Like Reality!

Check out Perceptilabs and sign up for a free demo here: https://www.perceptilabs.com/papers

The paper "Enhancing Photorealism Enhancement" is available...

6:36 PM · Jun 5, 2021Twitter Web App

Replying to

Thank you very much for the kind shoutout Andy! 🙏

This Tweet was deleted by the Tweet author. Learn more

jalil.eth

@jalil_eth

Jun 5, 2021

Take a close look at the grass. Also the reflections in cars, etc. in the video. Crazy.

Replying to

and

Would each run through on the racing course produce slightly different details? If you made slightly different paths early on, would the trees on the hills be different? Each choice you made might affect the render?

Andy Matuschak

@andy_matuschak

Jun 5, 2021

Conceptually, yes. Frames are fully independent, though, so there's a similar issue temporally, which (unlike the prior work reviewed) this method seems to handle surprisingly well.

Show replies

Replying to

and

To the latter point, it could also make designing worlds much more simple, since finer details could be filled in by the renderer. This paper a great example- minecraft world model produces amazing results nvlabs.github.io/GANcraft/

Andy Matuschak

@andy_matuschak

Jun 5, 2021

Ah, I hadn't seen GANcraft—thank you!