Hacker News Reader

Comment by panki27 | Hacker News Reader

I'm pretty sure the generation could easily run locally on a low-to-mid tier graphics card.

While it might take a bit longer to generate, you're still saving network and authentication latency.

…but then you just have a graphics card, built to render graphics, that you could tap instead through traditional tooling that’s already widely known and which produces consistent output via local assets.

While the results of the experiment here are interesting from an academic standpoint, it’s the same issue as remote game streaming: the amount of time you have to process input from the player, render visuals and sound, and transmit it back to the player precludes remote rendering for all but the most latency-insensitive games and framerates. It’s publishers and IP owners trying to solve the problem of ownership (in that they don’t want anyone to own anything, ever) rather than tackling any actually important issues (such as inefficient rendering pipelines, improving asset compression and delivery methods, improving the sandboxing of game code, etc).

Trying to make AI render real-time visuals is the wrongest use of the technology.

Nah, vibe coding is the wrongest use of technology, this is way to go. Why? Because good rendering isn't the necessarily most physically accurate one. You might actually want non-realistic rendering, and (depending on the specific style) it might be hard to impossible to get the right one on traditional pipeline. E. g. take "cartoonish" look - toon shading is, frankly said, total crap, because the artists rely on explicitly non-physical geometry to provide visual clues. This is definitely the future. Render on a normal pipeline (maybe with no lighting model at all), then put it through a style transfer network.

>Render on a normal pipeline (maybe with no lighting model at all), then put it through a style transfer network.

in 16ms? We keep needing higher resolutions to render to, so I doubt we'll ever have the budget to fit this kind of thing in.

Also, if Toon Shading is crap, this is completely unusable, even as a prototype. The present is being stripped away so it's a bit tiring talking about "the future". Let's actually secure it first.

I do not see how it replaces or substitutes network and authentication latency, especially in terms of a single player game in which neither is necessary.

This was built in September 2022, and it's still pretty mind-blowing today:

https://madebyoll.in/posts/game_emulation_via_dnn/demo/

https://madebyoll.in/posts/game_emulation_via_dnn/

Hook world state up to a server and you have multiplayer.

2025 update:

https://madebyoll.in/posts/world_emulation_via_dnn/

https://madebyoll.in/posts/world_emulation_via_dnn/demo/

That's really cool, thanks for sharing!

Reply on news.ycombinator.com