When you look at LoRAs for image generation model (like on Civitai), you get a very powerful range of different visual styles - allowing a base model to be constantly refined to suit whatever is trending at the time.

What makes text LLMs different, then? For example, why are there no specific LoRAs for things like python coding, or sci-fi writing, or so on?

I've read discussions about text LoRAs being specific to their base model - which is honestly also largely the case with image gen.

A rich LoRA culture would make a lot of models more versatile and reduce the need to keep downloading new checkpoints