Update: I've uploaded a Pony LoRA. It's fidelity is a bit better than the Animagine one, because there was quite a bit more fan artwork since her redebut. I'll update the Animagine version at a later date.Trained on her early sailor collar outfit (sailor) and her later frilly outfit (frilly). Use these prompts in Auto1111 (I've already escaped the brackets for you):Sailor: <code>mint fantome \(sailor\)</code>Frilly: <code>mint fantome \(frilly\)</code>The tokens above alone should be enough. To help with correctness, you can still add additional tags:General: <code>white hair, gradient hair, blue eyes</code>Sailor: <code>corset, paw pose, ghost hair ornament, hair ribbon, black ribbon, blue ribbon, ahoge</code>Frilly: <code>frilly, maid headdress, sleeves past fingers, cross hair ornament, two side up</code>Additional notes:<ul><li>Sometimes produces ghosts unprompted</li><li>The bridal garter'ed legs in the frilly outfit barely work, there's not a lot of artwork with her legs in it that I could find</li><li>I reduced dimensions after training (at dim/alpha=8/4) dynamically with <code>sv_fro</code> and a target of <code>0.9</code>. It has on average the size of a 4/2 LoRA after resizing.</li><li>I've used around double the usual LR. What should be an overfit LoRA seems to perform surprisingly well</li></ul>

Files with SHA256: da9501e8ccf83c8a748b3e8514a10d2ec4c97ed9d301d8781df4bf56ff1f2915