LoRAs and textual inversions are small add-on files that bias a model towards a character, a style or a concept.
Typical 1
A LoRA is loaded alongside the checkpoint and changes what it paints; a textual inversion does the same through a token you put in the prompt.Each carries a strength. For a LoRA, the model and CLIP strengths set how much effect it has on the picture and on the reading of the prompt. Around 1 is the usual starting point, and pushing far past it tends to flatten everything else the model knows.A LoRA is trained for one family of models. One built for Stable Diffusion 1.5 will do little or nothing on an SDXL model. The horde accepts the request either way and gives no error, so the only symptom is a picture the LoRA had no effect on. The picker starts from the family your model belongs to and says so when a selection does not match.When you add one, a specific published version is pinned, and not whichever is newest at the time. That keeps a saved request meaning the same thing after the author publishes an update. A version is always named by number, which is why typing one by hand asks you to confirm it is a version.Many LoRAs are trained against a trigger word and do very little without it. You can ask the horde to insert one for you, or add it to your prompt yourself. Triggers longer than thirty characters cannot be sent for insertion, so those are offered as something to add to the prompt instead.A generation carries at most five LoRAs and twenty embeddings. You can keep more than that selected; the ones past the limit are not sent, and moving one up the list includes it instead.A worker has to be able to fetch the file before it can serve your request, so an early-access or paid version cannot be used at all, and an obscure add-on can leave the job waiting.