<h3 id="heading-115"><a target="_blank" rel="ugc" href="https://civarchive.com/models/260267/animagine-xl-v3?modelVersionId=293564">Since the release of Anymagine XL 3.0, there is no need to update the model.</a></h3><h3 id="heading-115"><a target="_blank" rel="ugc" href="https://civarchive.com/models/260267/animagine-xl-v3?modelVersionId=293564">1.32 is the last update.</a></h3><h3 id="heading-115"><a target="_blank" rel="ugc" href="https://civarchive.com/models/260267/animagine-xl-v3?modelVersionId=293564">The Anymagine XL 3.0 is superior to this model.</a></h3><h3 id="heading-2096">version 1.3</h3>I extracted LORA from previously merged models and added LORA and lycoris that I learned before to increase tag accuracy.(kohakuXL extract + DPO extract + LORA learned personally + hassaku extract)<h3 id="heading-2096">version 1.21</h3>Deephentai-mix-v1.2 (This is a different model than the test version.)+<a target="_blank" rel="ugc" href="https://civarchive.com/models/119012/bluepencil-xl">BluePencil 2.0.0</a><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ebcf0cde-8545-4b89-9b49-5ab5d6e30b31/width=525/ebcf0cde-8545-4b89-9b49-5ab5d6e30b31.jpeg" />Version 1.21 merges <a target="_blank" rel="ugc" href="https://civarchive.com/models/119012/bluepencil-xl">BluePencil 2.0.0</a> and block tuning to improve stability.Version 1.0 was released for about a month with no consideration for stability.just focused on merging a lot of information.Because of this, it was very unstable.The 1.21 model is a compilation of models created by people much better than me,compensating for weaknesses in my model with block tuning,and merged in other additional LORAs.It is still being worked on for stability and tag compatibility.I recommend <a target="_blank" rel="ugc" href="https://civarchive.com/models/231989?modelVersionId=261856">my embedding</a>I recommend the default VAE.positive: 2 or 3 big tags like style, fonts, etc. + 1 sentences describing the situationToo many sentences can rather damage the image.Intuitive and consistent tags and 1-2 clear sentences are optimal.negative: text, watermark, <a target="_blank" rel="ugc" href="https://civarchive.com/models/231989?modelVersionId=261856">negative</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/231989?modelVersionId=264111">negative_v2</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/118418/negativexl">negativeXL_D</a>It is a model that merged several checkpoint models and learned Lora after learning it.(hentaiMixXLRoadTo_v21+_hassakuXLSfwNsfw_alphaV09-sdXL_v10_x0.5.fp16+ssd-1b-dalle-3-step00077700.fp16 + (kohakuXLBeta_beta7Pro+_reproductionSDXL_2v12-sd_xl_base_0.9-fp16_x0.25.fp16+ssd-1b-dalle-3-step00077700.fp16 - sd_xl_base_0.9-fp16) x 0.2) x deephentaiMix_v40_Stableversionx0.775+envyagendaxl02_v10x0.225.fp16 x 0.75 + bluePencilXL_v200 x 0.25 x deephentaiMix_v50 + (hentaiMixXLRoadTo_v3025DHCNSFW - sdXL_v10) x alpha(cosine) + Merge with Lora personally made

SDXL-Anime-Deephentai-mix

We used ViT-L-14/openai to learn captions for 2D images. This version is unstable.

Files with SHA256: 454c63f9eac1bdd9d1b3475edbc0eacee0f0bac2da937fc213b897a757e47cd6