This is a LoRA in ComfyUI format for Flux.2 Klein 9B trained to generate images of the person who appears in the reference image using a facial recognition loss
along with a teacher distillation loss that keeps it in line with the distilled model's outputs.
Trained on UTKFace and a subset of VGGFace2 images.
The original distilled model does this pretty well, but often generates a similar-but-different person.
Here are a few examples (Facial Recognition Score from CVLFace-ViT-webface4m. Higher is better)
