An uncensored model mainly capable of photorealism.v2.1Small weight adjustment of v2. Very similar but better imo.v2Clamped add difference v1 with <a target="_blank" rel="ugc" href="https://civarchive.com/models/312530/cyberrealistic-xl">cyberrealistic 3.1</a>. More contrast.v1Two models are used (+ <a target="_blank" rel="ugc" href="https://civarchive.com/models/840857">itercomp</a>).<ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/502468?modelVersionId=991916">bigASP 2.0</a> <a target="_blank" rel="ugc" href="https://civarchive.com/articles/8423/the-gory-details-of-finetuning-sdxl-for-40m-samples">(training details)</a></li></ul><ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/573152?modelVersionId=938628">Lustify 4.0 DMD2</a></li></ul>I chose these models because they are trained on large comprehensive datasets of SFW and NSFW images, with tags and natural language captions.The following steps were performed with <a target="_blank" rel="ugc" href="https://github.com/ljleb/sd-mecha">sd_mecha</a>/<a target="_blank" rel="ugc" href="https://github.com/ljleb/comfy-mecha">comfy-mecha</a>:<ol><li>Clamped add difference with the two models against sdxl base 1.0.</li><li>Merge "rotate" (orthogonal procrustes) the result model with <a target="_blank" rel="ugc" href="https://civarchive.com/models/502468?modelVersionId=991916">bigASP 2.0</a>.</li><li>Clamped add difference with the result of step 2 against <a target="_blank" rel="ugc" href="https://civarchive.com/models/840857">IterComp</a>.</li><li>Add the original CLIP and VAE from step 2 into the model of step 3.</li></ol>

Files with SHA256: b28f54da540b535b16cad051f548f6a31b7742729ad4216e5f171804e718fad3