> The diffusion models understand the physics of optics better than our hand-written math.
How well will they do with something like creating two adjacent mirrors with 30 degree angle between them with one of them covered with varying-polarized red tinted checkerboard pattern?
Haha, despite the 1000s of instances where it DOESNT simulate correctly. Very specifically chosen 10000-shot generated videos show somewhat impressive fluid physics... And even then, it MUST be something it's seen before. Diffusion is in NO way modeling physics in a realistic matter, there is not an infinite amount of training data to show all fluid dynamics...
Now I know you're too far down the hype rabbit hole. Either that, or you lack a cursory understanding of diffusion models.
How well will they do with something like creating two adjacent mirrors with 30 degree angle between them with one of them covered with varying-polarized red tinted checkerboard pattern?