Trained on grayscale Game Boy Camera photos. But it can do color photos using the prompt "color photo" and not including "grayscale". The training data was just regular photos with none of the stamps and stickers.
Generate at 512x448 for it to be closest to the look and pixel size. It can work in upscaling, but the pixel sizes stay the same producing more detail than the camera was capable of (see the red & white tulip photo in the examples). To produce more accurate Game Boy Camera photos upscaled, generate at a smaller size like 512, then adjust the image size in Photoshop with the Resample using "Nearest Neighbor (hard edges)".
Z-Image Turbo Version: Works best at around a weight of 0.80.
Qwen-Image Version: Works best with a CFG of 3-3.5. Trained in AI-Toolkit using the 3 bit with ARA method. Leans more heavily into monochrome since the training images were all monochrome. Put either "color photo" or "grayscale" in prompt.
Description
Trained for Qwen-Image in AI-Toolkit using the 3 bit with ARA method