A lora of the famous instagram model elisha herbert. Detailed prompts will yield better results. Use the example images for guidance. They are all first generations. I will update the model as time goes by and new techniques for flux become available. 1.2 vastly enhances the ease of prompting and quality outputs. It seems that flux responds extremely well to "Person XYZ is depicted in/as" I managed to get insanely good results after only 3 epochs of training using this captioning style and also using the same prompting style. For best results, please look at the example images provided. The best bet to get a high quality output is by using an language model and tell it to start the prompt with "Elisha Herbert is depicted". There are no real settings involved, the lora responds well in combination with other loras and settings. Just stick to the basics - guidance scale of 3.5, clip skip 1. Please note that flux training is still very new, currently the training bleeds into the main token a lot, so it transforms all women into the character - this is happening for basically all loras currently. I will try to negate this issue by adding women to the dataset, but i guess this is currently just a lora issue.

Elisha herbert

elisha_herbert-000037.safetensors

Files with SHA256: 1a64cceb5ee32e56c9549e9eea9f2b4efa1eb91c8eebbba9600b6f1832b96ae2