PLEASE NOTE - ALL MY CELEBRITY LORAS WILL BE DELETED (BY MYSELF) BY THE END OF JUNE 2024.
LoRA of UK actress Madeline Smith, who was one of the most ubiquitous British faces of the 1970s, featuring in a number of films and TV programs of the era.
Uses 114 ref images and 1989 (real) reg images, trained for 12,000 steps at a batch size of 1, using BF16 on a 3090, with a training time of four hours.
Despite its shortcomings (see below), this is almost certainly the best and most flexible/disentangled model in the catalogue - if only I could figure out why! Sadly, just repeating what I did with this model does not automatically produce equally effective models of other subjects; with LoRA, the data changes everything, every time.
Though this model uses the same data as my previous Madeline Smith checkpoint, it is not a Kohya extract from a DreamBooth model, but was trained from scratch in Kohya, after a lot of manual annotation of images.
Please keep the CFG very low (about 3.5 - and usually it's best to keep the sampling steps low too, unless inpainting), and the LoRA strength at about 0.8.
I generally use these models in complex workflows, inpainting faces after initial T2Is, and using ControlNet extensively. So if you're hoping for one-click prompt magic, my models aren't data-curated with this in mind, but rather as tools for traditional workflows that use Photoshop and other older methods.
This model is very likely to produce renderings unless counter-prompted.
Trained from scratch on Kohya_ss, on NVIDIA 3090, 24GB VRAM, for 4.5 hours on BF16, with 114 training images.
1. Dieses Modell dient nur Lernzwecken. Urheber- und Auslegungsrechte liegen beim Originalautor.
2. Bist du der Originalautor eines Modells, kontaktiere uns bitte zur Authentifizierung über unsere offiziellen Kanäle. Wir schützen die Rechte aller Schöpfer. Hier klicken, um es zu verifizieren.
