santa hat
deerdeer nosedeer glow
Sign In

The Piper LoRA That Never Was


The Backstory

TLDR: I have a new (really old) more accurate and more flexible method for my LoRAs!

This was one (if not the first) of my early LoRAs that I attempted to post to Civitai!

This LoRA ended up staying in early access then draft only and I never actually posted it as I thought it had "problems". Not to mention someone in the community (who actually left a comment while it was in Early Access) posted their own Piper Perri LoRA within the next day. 🤔🙄

Oof, checkmate.

This LoRA never got to see the light of day and it was probably a good thing as it took me more than 24 hours (precisely 1 day, 7 hours and 3 minutes) of GPU time to create it. That is not the kind of precedent you want to set for yourself (my wallet cries thinking about the electricity bill it would produce 😭)!

The Return

Revisiting this LoRA for kicks and grins (I thought my json for it were destroyed with a bad drive, but alas not so!) later with the correct prompts... wow! The subject accuracy is actually stunning and more so the ease and flexibility of achieving the desired output without overly specific prompts. I also noticed it maintained this accuracy without the overbaking side effects I tended to encounter when attempting to capture subject accuracy (ladies and gentlemen of the jury I refer you to my Kira Thorn LoRA). For a while, I was a bit dumbfounded as to how this could be (this was a good state for me to be in for a while).

Truly, training at a lower batch and slower pace with more epochs can create stunning results. No SD 2.x or SD XL needed (though I'm working on SD XL)!

In spite of how long I left this thing to bake, it is still slightly undercooked and it can be a bit tricky to get good results at first (which on top of my initial discouragement, is why I thought it was trash and never released it).

What Now?

First, some of my latest (unreleased) experimental LoRAs:




The rediscovery of this "old, washed up" LoRA has now caused me to experiment and modify my workflow with other LoRAs to recreate (and hopefully optimise!) the method I first accidentally used. Seeing the flexibility of this model, in particular with the truly great EpicPhotoGasm checkpoints has truly reinvigorated me to start creating LoRAs again. Additionally with the inspiration to find that elusive balance between accuracy and flexibility.

Hopefully what I discover over time I can bring across and apply to the SDXL world in the near future.

...and no it doesn't take me anywhere near 24+ hours to generate these new LoRAs! 🤣


Full Gallery

The full high resolution (non-upscaled) gallery of generations above, with a few extras here.

The Poll

So what are your thoughts and should I post it?