This was originally posted here a decade ago. I’m happy to see it’s still alive.
I’ve been using some generated assets for a game with voxelized art. I intend to take a deeper look at this and see if it can simplify parts of my workflow.
This is fascinating. I see its powered by weights and probabilities - would this be a very simple ancestor of things like Stable Diffusion that we have now, or would this be on a completely different branch (different approach)
It’s procedural generation but that’s pretty much where the similarities end. People today might use a big generative NN model to do this, using maybe a thousand times as much energy to get essentially the same result. Gen AI is definitely a big step forward in our relentless drive to make software more inefficient in order to compensate for any efficiency gains that the hardware guys come up with.
I always wondered how this compares to the 1999 algorithm Texture Synthesis by Non-parametric Sampling [1]. The results look very similar to my eyes. Implementation here [2] — has anyone tried both?
7 comments:
This was originally posted here a decade ago. I’m happy to see it’s still alive.
I’ve been using some generated assets for a game with voxelized art. I intend to take a deeper look at this and see if it can simplify parts of my workflow.
https://news.ycombinator.com/item?id=12612246
"Wave function collapse" - such a fancy name for a relatively simple algorithm without any connection to actual wave functions.
This is fascinating. I see its powered by weights and probabilities - would this be a very simple ancestor of things like Stable Diffusion that we have now, or would this be on a completely different branch (different approach)
It’s procedural generation but that’s pretty much where the similarities end. People today might use a big generative NN model to do this, using maybe a thousand times as much energy to get essentially the same result. Gen AI is definitely a big step forward in our relentless drive to make software more inefficient in order to compensate for any efficiency gains that the hardware guys come up with.
An explanation of how this works here: https://robertheaton.com/2018/12/17/wavefunction-collapse-al...
It’s interesting this article uses the phrase, “you feed it the vibe your going for,” about 5 years before “vibe coding” became a common term.
I always wondered how this compares to the 1999 algorithm Texture Synthesis by Non-parametric Sampling [1]. The results look very similar to my eyes. Implementation here [2] — has anyone tried both?
[1] https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/p...
[2] https://github.com/goldbema/TextureSynthesis