Only one thought deeply troubled me: why was the quality of the first model so unsatisfactory. It only dawned on me long after publishing the first version that I had been taking game screenshots in far from the best format (JPEG), when I could have saved them as full-quality PNGs without any compression. My experience with pixel-art models also pushed me to ensure all images in the dataset were multiples of 64. I believe this prevents any size distortion during training, meaning no compression occurs and the pixels remain unchanged.
In the first training run, I didn’t use as many images as I did in the second version.
The purpose? Oh, guys, I myself don’t even know what I wanted from this model. I just took some screenshots and, with the laziest approach possible, slapped on auto-tagging without giving a second thought to how nonsensical the image descriptions were—after all, it was just an experiment. Many of my friends think that even this level of tagging is a plus.
Judge for yourselves—I’ll post all the works that I found amusing in a way that makes them relatively convenient to browse in a gallery. Though, of course, everyone has their own taste when it comes to viewing galleries.
Just to remind you, the screenshots from Hades 2 were used as the basis for training.