Sign In

A look back on a year of LoRAs earning buzz

0

Aug 1, 2025

(Updated: 4 months ago)

ML Research
A look back on a year of LoRAs earning buzz

On August 1st 2024, Civitai creators began earning buzz when their models are used for on-site generation. Being a fairly active creator and pretty open about the creation side, I figured I'd share some of my observations from the now full year of data.

  • Buzz totals increased quite a bit in November 2024 (over 5x) and staying that way. I think it’s me doing more relatively famous stuff then.

  • Model use is feast or famine. Most models will never make 100 buzz (my propensity towards obscure subjects exacerbates this, but its not unique), let alone the 500 training cost, and often ones that make that ~100 but not 500 are because of what appears to be a single user using it a lot for a single spike in usage (I simply have enough enough variety for these spikes to be observable). The ones that do make back their training cost however will make it back several times (but not nearly enough to fund other stuff).

  • Most models rapidly lose interest past the first few days. I think this is mainly a factor of Civitai's lack of discovery meaning it's impossible to find stuff after the first few days unless you're looking for it or it's turbo popular.

  • Gundam X style seems to get popularity spikes after I use it for example pics of other models (my own and random things I post it to). Presumably from people seeing pic and figuring out where the style came from. That's at least one way to get around lack of discovery (characters I use to showcase styles, by contrast, essentially never experience these spikes)

  • If you want buzz, make sexual content, memes, content from currently airing anime, and women from 80s/90s/early 2000s content that were once very popular but don’t have a LoRA (or don’t for current models). Don’t let that discourage you from making models because you think they’re neat though.

  • Male characters do poorly, even from (relatively) well known series or ones that are heavily memed.

  • I finished converting my models to Illustrious on Feb 18 2025. Illustrious based model gens began to at least match Pony based ones in April, while Pony based ones became a tiny minority in May onward.

  • If you think of some concept or character and wonder “How has no one done this yet?” it’s either A: Got really rubbish training data options B: Going to be at least somewhat popular as a LoRA, or C: Both.

  • The intended to depict a minor flag is a kiss of death for LoRA preformance. Few see these and most have no idea they don't see everything.

Feel free to ask me questions on what my data says, or (for fellow creator) if my observations align with your own. I'm struggling to think of stuff to pick apart with it.

If you're an American, consider contacting your legislators in favor of the Fair Access to Banking Act to avoid platforms being debanked like Civitai has been. If you want to make models to see if people use them, you can check out my guide on how to make them.

0