Hassaku aims to be a anime model with a bright and distinct anime style.My <a target="_blank" rel="ugc" href="https://discord.gg/zSR5FcYWWE">Discord</a> for everything related to anime models and art. If you'd like to support my work, you can do so at: <a target="_new" rel="ugc" href="https://subscribestar.adult/citrus-models">SubscribeStar</a>, every bit helps and is truly appreciated! ____________________________________________________________ Supporters:Thanks to my supporters tamashiicolle, pttcot, SETI and Kodokuna ____________________________________________________________ Using the model:<ul><li>The model was trained primarily on images with minimal disruptive elements, such as floating text, logos, speech bubbles, or signatures. If any of these elements appear in an image, please include "signature" in your negative prompt.</li><li>Danbooru <a target="_blank" rel="ugc" href="https://danbooru.donmai.us/wiki_pages/tag_group%3Ametatags">Metadata</a> and franchise tags are not used in training, so please avoid using them in prompts. Metadata such as "highres" and franchise tags like "re:zero kara hajimeru isekai seikatsu" are not used in training.</li><li>A CFG scale between 3 and 7 is recommended. For best results, use CFG 6 with the Euler a sampler.</li><li>Refer to the example images, to understand, how the model should be used.</li><li>The model is designed to be simple and straightforward to prompt. Use the following tag order:<ol><li>Number of characters</li><li>Character Tags</li><li>All remaining tags Example: 1girl, rem \(re:zero\), standing, masterpiece, upper body</li></ol></li><li>Here are some resolution options for SDXL:<ul><li>1536 x 640</li><li>1344 x 768</li><li>1216 x 832</li><li>1152 x 896</li><li>1024 x 1024</li><li>896 x 1152</li><li>832 x 1216 (most recommended)</li><li>768 x 1344</li><li>640 x 1536</li></ul></li></ul>Lora use:<ul><li>NoobAIs Loras working for the most part better in compare to <a target="_blank" rel="ugc" href="https://civarchive.com/models/795765/illustrious-xl">Illustrious loras</a></li><li>Versions 2 and 3 is less compatible to most loras trained on Illustrious or Noob, because it got further trained away of both. Style A is still the most compatible for it.</li></ul> ______________________________________________________Version and License info:<ul><li>Below V1 merges <a target="_blank" rel="ugc" href="https://civarchive.com/models/260267">ANIMAGINE XL 3.0</a></li><li>Version V1 use <a target="_blank" rel="ugc" href="https://civarchive.com/models/795765/illustrious-xl">Illustrious-XL</a> &amp;<a target="_blank" rel="ugc" href="https://civarchive.com/models/827184?modelVersionId=1068947"> WAI-NSFW-illustrious-SDXL</a> with additional training</li><li>V2 is trained on its own and don't include any extra merge. Base was V1, it is trained to include newer or missing characters. It was also used for further training tests</li><li>V3 is a merge of <a target="_blank" rel="ugc" href="https://civarchive.com/models/795765/illustrious-xl">Illustrious-XL</a> and V2.2, to fix some issues and was also trained to include newer or missing charactersAll Models using the <a target="_blank" rel="ugc" href="https://freedevproject.org/faipl-1.0-sd/">Fair AI Public License 1.0-SD</a> license.</li></ul>

Hassaku XL (Illustrious)

The pure trained model output. Completly unaltered, as Nature intended. This model does not have any style merge stabilisation and is intended as base for further training, merges, using Loras, etc. for anyone, that are interested in that. Be aware that it can have every kind of problem, a base model can have. But it have a little specialty to it, the style got natively trained into the model, instead of merging it into the model per loras. It enhance the default output of the pure model. If you compare it with 3.1, you will notice some similar looking images, because of the low style merge of 3.1 itself. In some cases, this pure model may produce better looking images, but it can also create significantly worse images. Results will vary greatly. It have one trained epoch more (400k images -&gt; not a big differece) in, in compare to 3.1. With slitly changed training: Trained now on LION, with Stochastic Rounding, <a target="_blank" rel="ugc" href="https://arxiv.org/abs/2412.08894">SMMF, </a><a target="_blank" rel="ugc" href="https://github.com/kyleliang919/C-Optim">C-Optim</a> and higher Alpha1. Before with adafactor Stochastic Rounding and Fused Back Pass.

Files with SHA256: 5e709efe7a51ba4409bf425c7435d09ce76b6a44784f02a6441a8cc6884829a2