What I've been working up to! Better quality! Less mistakes! More consistent head frillies (they're external gills)! Happy with the results. This version plays much better with other loras compared to previous versions. I'm using a custom node in comfyui for the showcase that is able to capture all of the generation info except for the lora stack. I'm mostly using eldritchadam's loras, I'll have them linked in the suggested resources. I'm a fan and I like using his loras with my models.

If your prompts are really lengthy and you're using a bunch of other loras, i find it useful to negative prompt things like "rodent", "whiskers", and "fins". This is because stable diffusion's idea of an axolotl is a pink salamander but sometimes their head frillies get turned into something that looks like mouse ears and then it just runs with making something that looks like a mouse. This won't come up as much because I regularized this model against generations of what stable diffusion thinks an axolotl is, but it can come out if there are a bunch of other things fighting for control. I'll try to make a pdxl version soon, I don't use it, but I've seen some fun loras I've been missing out on and it seems like my models don't have much of an affect on those checkpoints.