Background [skip to next section for actual info]

This model was originally an experiment to see if I could get similar results to Umbrox's Alarune models for SDXL and Pony in SD1.5 by training on a few of the generation images automatically captioned. The result did a decent job of transferring but also, due to other LoRA's results in the dataset, including Goblin_DND's Dragonborn model and crystalkalem's Novakid model, the resulting SD1.5 LoRA learned an interesting head shape and facial structure somewhere in between dragon, bird, and robot. I named it Xlalarune (since it seemed like some sort of fantasy name and is a concatenation of XL (the original base model) and "Alarune" (the name used for the original LoRA). I ended up using it occasionally on my YouTube channel as a symbol of AI creativity and standing against uninformed AI hate.

The specific checkpoint I tested was really bland, so I ran the training again months later and posted it here as "2nd Run" (which has a much darker and more alien feel). However I found out a few months after that that I still had some alternate checkpoints from the first one. One of those shows more promising results and more accurately depicts the canonical Xlalarune face shape shown in some video thumbnails (with a pointed dragonlike head with long empty eye sockets; the 2nd run model only gets this sometimes)

How to use

For Run 1, I suggest you download the 72 MB model. The 1GB+ LoRA file download listed is the entire contents of Run1's saved training progress, with 20 separate checkpoints, in case you want to try other ones (step00000500 has a more wild, dishevelled look (because it hasn't fully learned and is bad, but its fails actually look cool in this case))

I recommend actually just using it with the Stable Diffusion base model if you're okay with 512x512 images (and the hi-res fix in A1111 doesn't work well so yeah). Other models will work but have widely varying effects, including some sort of new DND concept creature (Species https://civarchive.com/models/17564?modelVersionId=93446), armored people in the forest (CleanLinearMix https://civarchive.com/models/42433?modelVersionId=186152), or more standard but much leafier and more flexible as far as prompts go (this model which despite its nature and dataset isn't exclusively nsfw https://civarchive.com/models/21728?modelVersionId=249427 ). These alternative interpretations are generally pretty flexible as far as prompts go, while the base model generates the "true" interpretation of the Xlalarune creature but isn't super flexible. You can however put armor on it (it already kind of has built in armor but you can change the metal, add gems, etc.) in the base model, and some other LoRAs have interesting combinations with it, such as the Minecraft Warden fan art model I made earlier (https://civarchive.com/models/644740?modelVersionId=721231), though that one tends to mess up the face in an AI-fail kind of way.