I started digitizing Anais Dior's face a few weeks ago and i..

I started digitizing Anais Dior's face a few weeks ago and it has been a journey.

These are all txt2img with not controlnets in place. I wanted to have a pure model without influence of anything other than the prompt.

The first four images were my first attempts using the Lora method, I did about 5 attempts and sent her the best from those groups and she said "they look nice, but they don't look like me" and I was a little offended (because of the amount of work that went in lol), but she was absolutely correct. While it has some elements that are similar to her's, it's off by enough that it doesn't even come close to passing the side-by-side test.

I started doing more research into other SD users who post trained models to Civit and all the bigger ones seem to swear on Textual Inversion (TI) for people's faces vs Lora, which apparently is better for styles. Now this isn't reflected in the NUMBERS of models that have been trained under Lora vs TI, but I'm starting to realize it's probable because of 1. hardware requirements to train a TI and 2. Lora's are just flat out easier to train. So this past friday I started from scratch to train a TI model of Anais and it was a HUGE pain in the dick. I made 6 attempts, with every iterations being almost a complete do-over because the problems seemed rooted in the initial tags that you do for each sample image (1girl, anaisdior, smiling, etc). There are TONS of writeups online from these 'trainers' but they are anecdotal at best as we are all fucking winging it with this tech. Regardless, I ended up with a prettty good model for txt2img that can be improved by using inpaint (where you replace parts that didnt render correctly, like a face or hands or tiddies lol). It also insists on making her eyes blue by default, even tho they are clearly hazel in the training dataset? AI gonna just lie I guess.

Everything from image 5 onward is the latest model, all txt2img with no inpainting or touchups as I want demonstrate how close we can get to "real" people. Is it perfect? Nah. Would it trick someone who perhaps hasn't seen Anais in a minute? Absolutely...maybe. One thing is for sure, for the last couple of images, there is enough "Anais" there characters in other styles that are inspired by her *look*. Disney Anais, Anime Anais, anything is possible with this tech! The last pic, btw, is what happens when the AI doesn't know what a "tron lightcycle" is, so instead it puts the character over a regular bicycle with neon lights in the back lmao.

Anais fans, what do you guys think?

Thanks for letting me nerd out about this stuff here, I don't have anywhere else to do it!