Baguette-Anime XL: v1.0: baguetteAnimeXL_v10.safetensors

SDXL 1.0 Standard Model

Baguette-Anime XL is a model that I created based on images which suited my tastes.

Trained on more that 85K images, it is a model that I want to maintain and work on to make it better each generation.

As it is my first ever model, I still have a lot to learn and improve, nonetheless, I still find the results satisfactory.

The first version of Baguette-Anime XL was quite lacking on a lot of point, and some tags were not well recognized by the model. Thus, I created a second version, based on the first, with more images.

I also ran the training with images that were not resized, to let the model learn how to create details on zoomed/cropped items, this leads to way better results with Hires. Fix and Ultimate SD Upscale.

For this reason, I strongly recommend you to use Ultimate SD Upscale, when you upscale images with this model, and to use 1024x1024 tiles while upscaling.

Important keywords :

This model was trained to be able to recognize danbooru tags.
Since the dataset was still relatively "small", there are some tags that the AI didn't understand well, or didn't see at all.

In addition to common danbooru tags, there are a total of four "special" keywords that the model was trained on :

  • bad quality & worst quality : Works that the AI was trained to not reproduce, because of distorted limbs, or because the image didn't suit my tastes

  • best quality & masterpiece : Works from which I wanted the AI to learn the most

As you can see in the exemple images, I use these 4 keywords everytime.

Please, note that the masterpiece keyword may change the composition of the image in version 1 of the model. This problem seems to be fixed in version 2.

Known problems :

In version 2, the model is way better than in its first iteration. However, during my tests, it seems like this model still has some problem that may be fixed in future versions.

First of all, the model still has a quite weak understanding of tags involving multicolored hair. streaked hair and two-tone hair do essentially the same in many cases, while colored inner hair doesn't work at all.

Secondly, the model still has a lot of problem recognizing characters. While being better than the first version, on which no character was known by its name, the version 2 still makes a lot of errors, and outputing image of a certain character may be difficult.

Thirdly, the outputed images when adding NSFW tags are nowhere near satisfactory. From my tests, at most, version 2 is able to generate characters in underwear.

Roadmap :

  • Make the model able to recognize more characters by their name

  • Improve the NSFW capabilities of the model, that are quite lacking at the moment

I will gladly accept any feedback about this model, to improve it even more in the future !