Hacker News Reader

Comment by jakub_g | Hacker News Reader

Since talking images, are there any AI models that can output real transparent gifs/pngs?

And not a (botched) fake white/gray grid background that is commonly used to visualize transparency?

This is exactly the use case I just tried, and not only does one get a non-transparent checkerboard, the squares are inconsistent & jumbled.

Example: https://gemini.google.com/share/36d66cad1764

ChatGPT's image generator has been able to do this since last year. That NBP still can't is baffling. They should at least train it to respond to requests for transparency with a solid colour pink background.

This. Gpt-image-1/1.5 are the only ones that have this built in - though I'd love to have an insider view if its natively considering the alpha channel or just feeding it through a rembg-style post processor.

Certainly the initial versions were post processing rather than native. I'd be interested to know if that has changed on subsequent releases.

You can output to a plain background and use any number of tools to mask it.

I know. It sounds like a perfect task for AI to do it though (wasn't the whole premise of AI do to mundane things for us), yet they fail to do it, and I need to use an external tool.

Alpha is a 4th image channel that 99%+ of images in the training data do not use, so it makes more pragmatic sense to just not allow it.

Weirdly, MSPaint (or whatever it is now) is really good at this (with, I assume, an AI model)

The output from Nano Banana, even when it is ostensibly drawing "single color" shaded areas, is so jittery that it can be challenging to threshold it.

Nano Banana Pro is better at that type of thing, so I suspect Nano Banana 2 will be sufficient.

Reply on news.ycombinator.com