This model was trained at 512px x 640px.
Consequently it rarely 'doubles up' when using the popular 512px x 768px format.

It is best used at a LoRA strength of around 0.7.

A LoRA of 1960s/70s British glamor model and actress Margaret Nolan using 150 ref images and 1980 (real) reg images, for 15,000 steps over two hours on a 3090 GPU. Trained from scratch in Kohya, not extracted from a DreamBooth.

This is the hardest LoRA I ever made, because good data was impossible to find: Margaret Nolan appeared usually only in walk-on parts in minor movies that never made it to Blu-ray, or on British TV, where the source video was very poor, even if you could get your hands on the original.

Please manage your expectations - I had to colorize 95% of the material for this LoRA, and it is going to show in the output.

I also had to create intermediate LoRAs in order to attempt to get better synthetic data for subsequent versions. There's a limit on how well that works, and this too will be obvious - don't expect that you can paint the face in and get a ton of detail, because that detail was not there in the data, despite all the augmentation. However, face inpainting does work well with this model, so long as you don't zoom in for a huge close-up. That's how the accompanying images for this model were done, in at least two steps.

Additionally, further passes on eyes and lips, if well-prompted to add detail, can be very beneficial.

USAGE

For some reason, this LoRA performs very well in an extraordinarily wide range of CFG levels and steps, so long as you keep the LoRA strength at around 0.7. You can slide the CFG value practically anywhere without blowing the render out. I have no idea why - the data was pretty terrible.

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 NSFW and nude renderings unless counter-prompted.