Yeah, but is this really that great? Are these models going to remember the town you wandered through on your session yesterday and want to return to?

Imagine playing Read Dead Redemption 2 and you attempt to ride your horse from Saint Denis to Valentine and Valentine no longer exists, or is a completely different town located half a mile off from where it was originally.

I just don't see how this would work...

If I had to use the models as they exist right now I'd use them in a procedural Myst-like where I incorporate the temporal inconsistency into the setting. The player's actions and state would affect the prompts used for conditioning the video generation. It would probably be weird and buggy but could be fun.

You could also use these models to generate assets for a game during development whether that's simple cutscenes or assets produced through gaussian splatting or some other process.

If these models and others can be run cost effectively on a cloud service or even locally at some point then you could do some interesting things by combining them with 3D mesh generation, img2img, vid2vid, etc. just think about even simple games like Papers Please and the whole genre it spawned that uses short episodes where you have to make a guess based on what you see, there's a lot of potential for creating new mechanics around generative imagery.

Same prompt, same seed, and yes you can ensure you get the same output, but also imagine using it as a game designer and recording the output. Imagine level editors where you prompt to fill in details, walk through it, decide which parts you don't like, and prompt for a replacement of those parts.

It's not hard to imagine a system that combines deterministic state tracking with diffusion generated scenes.

Yes, a lot of models don’t state this explicitly, but they can be made deterministic. Not the generation itself, but the same prompt, with a generation seed will always result in the same output.