This repository contains all the scripts for an AI image detection CSSL NN I am training.
Images are fed to diffusion models as latents - a distilled representation of the image that allows processing with reduced overhead. This is facilitated by a variational autoencoder (VAE), a neural network that encodes/decodes between images and latents. Because decoding by VAE is a lossy process, and a process that is currently used in all major, modern image generation implementations, it seems like an obvious candidate for detecting AI generated images.
I aim to demonstrate this effect in the following VAE's:
These models represent a diverse range of commonly used VAE's:
SDXL: dataset model
DALL·E: dataset model
FLUX: dataset model
These models are either incredibly popular but may be similar to SDXL or are old SD models with less relevance today.
XL_VAE_C: dataset model
djMILXLVAE: dataset model
Pony Pastel: dataset model
SD 1.5: dataset model