Hacker News Reader

It probably won't learn anything new, but it will learn not to generate bad images if you cherry pick the best ones for training.

Yu would train it in an RL setting rather than actually use generated images in the train set.

Sure

No, it isn't if you don't curate meticulously. Perhaps by accident something new could emerge which is worth it to use as input for an advanced model.

But the rule is that new concepts are very hard to produce, although thanks to countless models, stable diffusion is probably the most flexible approach by quite some margin.

Likely they have humans review the images (or even touch up). Same thing with the dataset scraped from the internet.

Yes, that's the technique. Generate 10 images, then choose the ones that turned out well for the next round. That's the standard way to create a LoRA.