Hacker News Reader

Comment by ikishade | Hacker News Reader

I have a tool that can up-scale heuristic images based on a training sample, and a smaller input image. But I've found the reception to that to be very lukewarm, maybe because technical proc-gen is quite niche?

View on HN

> I have a tool that can up-scale heuristic images based on a training sample, and a smaller input image.

What, in this context, is a "heuristic image" as distinct from any other image and how is this different (in utility, not in mechanism) from the vast array of AI image upscaling tools that already exist? What does this do that you cannot do (or cannot do as well, cheaply, or easily) without it?

Explaining—and demonstrating—this would probably help gain traction.

In this context, an image where there's some data represented by each unique RGBA value, and less of the "smooth" complexities of normal images like from a camera. So ones with small color palettes. Depending on the output size, it can run very fast and I've used it to up-scale a simple 2-color map for an island in water off one 512x512 training image.

I tried to demo this to HN back in July, but my post back then was buried.

For that particular application, I think the key part is going to be explaining/demonstrating the advantage over non-AI approaches, especially given that the non-AI approaches are not patent-encumbered.

That's a very good point. Thank you!

I'd recommend putting together some side-by-side comparisons of your tool with a set of diverse input images against other popular upscalers like ESRGAN [1], ControlNet Tile approaches, SUPIR, DAT [2], etc.

- [1] https://arxiv.org/abs/1809.00219

- [2] http://arxiv.org/abs/2308.03364

Thank you for the advice! Mine's mainly for limited palette upscaling, so that might be Apples to Oranges, but I'll give it some thought.

Reply on news.ycombinator.com