If you find any improvements, please let me know and I will refer to them.Version 2.1<ol><li>Increased dataset.</li><li>Poses have become more dynamic</li><li>The amount of farts increased.</li></ol>Since this is an increased dataset in addition to the 2.0 dataset, you can consider using both at the same time to find an appropriate value.<hr />Version 2.0I've completely changed the way the model is trained.Learning was done using animagine v3.0 base,but since it was learned individually, it can be used to some extent in a pony environment.You can apply the sentence, such as "fart squirts from the ass" and "huge fart suqirts from the ass", "green fart", "yellow fart" etc.And when changing the color, you must add the desired color in front of all fart-related tags.Farting and farts tags with similar token values ​​are also possible to some extent.Of course, it's not perfect.One thing is for sure, this is the most perfect fart model I have made so far.<hr />Version 1.311 changed the data set while learning using the MLP method.Not intended, but it also works with ponydiffusion.<hr />LOHA has a very strong norm=tensor by nature, requiring a low weight of 0.12 to 0.15.(This is not the correct answer because it is a personal recommended value.)First of all, there are three learning tags: fart, farting, and sprays, but due to the nature of SDXL, these three keywords can be applied to other sentences.An example is “There is a fart below.”I built this model to act as a secondary, so I don't recommend using it exclusively.For stability, it is recommended to adjust the tag weight rather than the lora or loha weight.I bet it's much better than just using v1.2 alone.<hr />Version 1.2 is the same dataset as before, but with a different training method and tags.It is relatively more purposeful than 1.1.<hr />The 1.1 version still has a lower hit rate than the 5.1 (old) version.I plan to supplement this part additionally.However, the overfitting is so severe that I recommend the 1.1 version.<hr />5.1 (old) Due to promiscuous active tags, the Lora model had to be weighted down to overfit, and the model itself was unstable due to frequent mergers.Version 1.1 recommends a weight of about 0.7.The reason for version 1.1 from the beginning is that version 1.0 has only increased the same setting and the number of learning steps on the dataset.<hr />

Fart_SDXL

When I used to make it, I used to use promiscuous active tags. Now I've reduced it to two simple tags, which reduces tag collisions and architecture.

Files with SHA256: ce77412025806b3a74653866f95b97ef2558284bfa055df74bda43dd3978f2e8