<h1 id="heading-709">Ideal</h1>Hi everyone, in this page you will find my "new" long term project (ofc i will keep training loras in the meanwhile), the ideal model serie.Those are models obtained by merging loras and checkpoints to obtain flexible models, that excells in something.Each "Version" of paradise will have a different "use" as objective, so there won't be a classic v1, v2 etc.For info on each version check the apposite section in the page.<h3 id="heading-710"></h3><h3 id="heading-711">This serie will use nsfw models in the merging process, so the models will be able to produce nsfw content!</h3>

Ideal

This version is a merged model obtained from:<ul><li>one of my private models (Forbidden), based on anylora</li><li>SCH excelsior</li><li>BBEDA</li><li>Origin</li><li>薄星夏V6</li></ul>That said the objective of this model was to have a flexible model that can excell at portrait images, but can still produce other kind of images without any issues.The model has been tested in the last week by me and a friend of mine, and it is somewhat decent; but it has a few requirementes:The use of the following embeddings, is strongly recommended:<ul><li>EasyNegativeV2</li><li>Bad-Hand-5</li></ul>The use of highres.fix is also strongly recommended.Adetailer is mandatory to achieve the intended expressions result.That said this model works good with the classic: vae-ft-mse-84000It works with most samplers, but i suggest to use:Euler a - cfg 7 - between 20 and 40 steps (it can probably go higher)DPM++ SDE - cfg 7 - between 30 - 40 stepsClip skip 2 is the best (but some times also the 1 produce good outputs)

Files with SHA256: 8ab69e0c5e94e7db7c53a2ba9ab03755b6038e1be08a87fe8cb6fd6f5b461d21