!!! Use LCM Lora, or an LCM model. You will be very disappointed if you don't.

Large images are best at 512x896 at 10 steps and you can leave hires 1.5x on most of the time.

Her dataset is mildly spicy and most of the images will be. I take no responsibility for any Taylor Swift incidents.

This is using my LCM optimized training which has better likeness, but the cost of likeness is reduced pliability of concepts outside the Lora, so you will mostly just get her kind of images with cutouts and midriff. If you want to do abstract prompts or concepts you'll need to reduce the strength to 0.6-0.7 and inpaint or Adetailer her likeness back in or use controlnet tile/blur at 0.1-0.2 strength and Adetailer her likeness on top with the Lora and work it iteratively.

Workflow update: To enable Lora and Negatives and better prompt adherence use the CFG fix extension. Copy the first image for settings and a good lora flavoring.