ModelsLab/Snakebite 2 - V2.2 Turbo

2.2 was finetuned on around 700 high-quality photographs to bring the model's baseline style much closer to that of real life. It also has better anatomy and more stable backgrounds.Training details:Used Flow Matching objective for maximum quality.Photos were selected for unique backdrops, little to no grain/artifacts, strong composition, and subject diversity.Plenty of NSFW material.All watermarks removed.Other styles--such as "lofi" or "cinematic"--were excluded from the training set, in order to establish the baseline style.Trained for 20k steps, then pruned a few overcooked blocks.Captions generated with JoyCaption as it's what bigASP knows best.Takeaways:Terms like onlyfans or photorealistic are optional in 2.2. Using photo of a... is usually enough.I'm using a quantized version of JoyCaption Beta One that runs on my local hardware, and it's not perfect. It often goes into detail about watermarks that don't exist. I was worried this would harm the model's ability to spell words, but that didn't seem to happen. If anything, Snakebite 2.2 might be a little better at text than ever before. No idea why.Despite the carefully cultivated dataset, some image artifacts are present in 2.2. Based on my experience, increasing the dataset to 1-2k images could alleviate this. Improving the captions would also help.2.2 likes adding some soft focus/depth of field by default. It's a bit difficult to steer away from. Adding more "sharp focus" photography to the training set would help.For me, Snakebite 2.2 represents peak photorealism in SDXL, especially when you factor in the prompt adherence. But if you like making other types of images, you may find that 2.1 is more appropriate for the task.Have fun!