Version 5.0 started as a 50/50 merge between minor finetunes of NoobAI and UnsafeXL 4.4 to bring their text encoders closer together. This was followed by 145 training iterations using over 40k images, with most of the new images being photographs. The final version of the model should default to a photorealistic style while still supporting the concepts of booru tags, including more fantastical and science fiction concepts. This model was not trained with quality tags, so prompts with "masterpiece", "best quality", and the like will pull it back towards its anime roots. If you just want anime style, which it still understands, you can prompt for "anime". Version 4.1 was trained with PonyXL v6 as the base, using a dataset of over 24k images, selected to teach concepts that were missing from the model, and 4.0 was based on AnimagineXL. 4.2 was a merge between 4.1 and 4.0, with further training afterwards to stabilize the merge and add styles besides anime. As a result, usage of this model needs to adhere to the licenses of both Pony and Animagine. The main focuses of this model are nsfw, holding weapons, and horror/monster themes, (which is where the "unsafe" name comes from) but it is meant to be a generalist model capable of handling images with or without those concepts. There is also a focus on being able to prompt facial expressions and general art styles.Due to using an anime model trained on booru tags as a base, this version fully trained the text encoder, so it should have much better prompt adherence when using booru tags, but will have worse natural language performance. Because it has been trained on a variety of styles, you'll get more consistent results if you specifically prompt for the style you want, such as "photorealistic", "realistic", "3d", "anime", "cel shading", "painterly", "1980s (style)", "oil painting (medium)".

UnsafeXL

This version was trained from version 3.2, which uses Juggernaut v6 as a base, but also had the DPO Lora merged in near the end. The major change this time is training the text encoder on a number of tags, but only one tag at a time to avoid affecting the natural language part more than necessary. The dataset is now 12k images to train towards and 4k images as negative training.

Files with SHA256: 975f2816f5168eb4d53716dbce55a21c12d4a7c10ded28c58698bda10b0dfae7