Creating interesting faces to train a LORA (New Lora coming soon)

🎭 Introduction My search for truly interesting faces continues. So far, I have created over 30 character LoRAs, always trying to move away from the typical AI beauty standards. I am fed up with perfect images: smooth faces, flawless skin, always impeccable features. That's why I decided to experiment with a LoRA that aims to produce more realistic, imperfect, lifelike images.

🧬 The Idea: Moving Away from Perfection To create something different, I turned to a very interesting website: Artbreeder, a platform that allows you to generate portraits, combine them, add particular "genes," and distort features to create truly unique characters. 📚 Creating the Dataset I selected about 70 images of women with deliberately accentuated and marked features, trying to vary the dataset as much as possible. These are not perfect images: many have artifacts, asymmetries, and deformations. ➡️ And that's the beauty of it. By training a LoRA for Z Image Base on these images, the result is low-quality faces, but they are ideal as a starting point for creating the initial face to be refined and prepared for the final dataset.

🖼️ Generation Examples Below are four examples: Left: images generated with the imperfect LoRA Right: images "reconstructed" using a workflow based on FLUX.2 Klein 9B, starting from the "ugly" face

✨ Observations on the Result As you can see, Klein always tends to slightly embellish the features, correcting some of the flaws in the original image without completely eliminating them. The result is a series of unique faces, perfect for composing new datasets rich in character and variety. To further accentuate control over style, I recommend using a LoRA trained on amateur photos (there are several on Civitai). I chose: "FLUX.2-klein-base-9B_LoRa_by-AI_Characters_STYLE_SmartphoneSnapshotPhotoReality" – designed for Klein Base, but it also works very well in this workflow.

📢 Coming soon In the next few days, I will publish the complete LoRA, with all the technical specifications, recommended parameters, and additional examples. Below you will find more images generated and their evolution using Klein.