This is a test/beta model based on <a target="_blank" rel="ugc" href="https://civarchive.com/models/105047?modelVersionId=164370">Break the Darkness v08ma</a> (BTDv08ma) by <a target="_blank" rel="ugc" href="https://civarchive.com/user/br_d">br_d</a>, which aims at making a new SD1.5 anime style model that has no NAI-based models merged into it. It's a merge of multiple models that br_d has trained directly from stock SD1.5.After some testing and continuing on the same idea, I reused some custom test models that also have no NAI-based models in them and merged them with the model to try to stabilize the results without modifying the model's style too much. While it's still a bit hard to control and far from perfect, I think the results are at least interesting and might be useful for others to test with.Similar to BTDv08ma, results are 2D anime style, with thin outlines and mostly with flat colors and soft gradients. It also has limited flexibility in the styles range (unless additional models like LoRAs are used) and because of it not having NAI (or similar) influence, it is also weak regarding anime characters and normally used danbooru tags (where again additional models could help). Different from the original, this merge can produce clearer and more colorful results more consistently.While each version of BTD gets much better than the previous one, it still takes some work to tame the results. It's the same case with this model, which tends to:<ul><li>put subjects off-center of the image (can be mitigated by adding "centered" in the prompt</li><li>have some issues with hands and arms</li><li>not work too well with some LoRAs that depend too heavily on either the style or the danbooru tags of NAI-based models (LoRAs trained on SD1.5 might fare better)</li><li>not follow the prompt consistently</li><li>add some exaggerated "effects" in the composition</li><li>make subjects in very similar poses when altering the prompt, but keeping the same seed</li></ul>The examples in the showcase were all made using the same seed, a simple negative prompt, very few LoRAs (if any) and no hi-res fix. The prompts were copied from other models to try to have a point of reference to compare to.If you test and have any interesting ideas that can be explored, let me know. Also, take a look at br_d's awesome models! :D

Files with SHA256: fb2ee969aedb6a1b38bef18aa58df5dd6b13c7cc541fb48ab985036d2a83aa49