Speaking of Lahanna, I've got an AI update on her! This is..

Speaking of Lahanna, I've got an AI update on her!

This is another lengthy post, so don't judge the pictures accuracy just yet as they represent the process from not-quite to pretty-much-nailed-it. I trained a total of about 8 AI models of Lahanna over the past few weeks, and it was my most difficult face training job to date.

First, there were issues with teaching the AI how to represent nipple and nose piercings (nipples in img 2-3), so I went through the dataset, which is the collection of images and tags that are used to train the AI, and photoshopped out all the piercings and removed all mentions of piercings in the tags. I *thought* I had done a good job, and visually you can see that the nipples are clean and free of barbell piercings, but the ai STILL found a way to see it. Upon further research, it turns out the method that it uses to train a model can also reveal even the most minor of blemishes. I did another pass few passes, doing more and more photoshop work with each one and I ultimately got them to be *almost* perfect, but nothing that can't be fixed in post or using inpaint (img 4). Piercings were the bulk of the initial issues.

Next hurdle was her face structure and hair, which were actually very much connected in the 'eyes' of AI. Lahanna's look is inherently alluring, and what I mean by that is a woman with long wavy full bodied hair that's falling over her face while she's eyeballing the viewer is, like I said, a sexy look. Obv that's not what these AI checkpoints were trained with so it suffers in multiple ways when you try and force it to put hair over even ONE of the eyes, and this is a trademark Lahanna look! So it seems to compensate by substituting incorrect features and proportions on her face (img 5-6). You can tell these are at least inspired by her look, but they are still incorrect. This led me to revisiting the dataset one more time, this time ADDING "hair in face" to the training data.

The actual training parameters hardly changed, as I certainly feel I have that close to nailed down from my Heather and JMO models, so the problems resided almost exclusively in the damn dataset lol. After adding the appropriate tags, recropping some of the images, and adding a few more shots where hair wasn't covering Lahanna's face to the dataset, I was able to generate probably about as close to the real thing as we're gonna get, and it's pretty damn close imo (Img 7-11 and 1, ignore the gimp arm and focus on the face! lmao). Included in these final pictures are a few artistic variations like anime, pop art, etc, to demonstrate how flexible the model is along with how well it captured some of her notable features and proportions.

Lahanna loves they way they turned out and comments that it's still not quite a replacement for the real thing, which I totally agree with, and is down for doing some fun stuff with it in the near future. I'll probably start putting together some halloween collections with some of our favorite models, starting with Lahanna, naturally 😈

Don't forget to check out @inkedbratlala's OF!