Hacker News Reader

How much of it is it `painting` though? It looks like your typical startup, combine two ideas into one. (Image) Search engine + Photoshop. You enter some prompts, it searches it's database for matches, and meshes all found results into a single visually pleasing image. Can it draw something that is not a mesh or a variation of it's database? Can you?

View on HN

I don't that "searches it's database for matches" description is a great metaphor for how this really works.

As Andy points out in this article, the model itself is a 4.2GB file. That's way, way too small to work as a "database" of examples it can stitch together.

I think of it instead as an enormous mass of loosley assembled impressions of concepts - everything from a low-level primitive like a triangle to a Star Destroyer. You can use text prompts to combine those primitives - so you could get it to generate something just from dots and lines and shapes, or you could mix in extremely detailed concepts like the Seattle skyline - or anything in between.

What is painting if not a mesh of ideas?

At bare minimum I can think of the 3D space the character is in, and rotate and shift it to any location and position I can dream of, even if I have never seen it before. When we will have 3D aware AI, that would be interesting.

Reply on news.ycombinator.com