V2.0 Released: A pretty big upgrade from 1.1, primarily achieved by mixing in the excellent <a rel="ugc" href="https://civarchive.com/models/79872?modelVersionId=84846">Brixel Core</a> at 30%. It adds a massive amount of flexibility to the model, letting it do a variety of lineart styles, and even fun chibis, while retaining the soul and core of the style. This also tamps what peculiarities in prompting were present, and makes it a touch more responsive to many things.RoseRendered is focused mainly on achieving expressive characters in a western, somewhat rendered style. Previously it was a counterpart to my Drawn Mix, but recent developments mean it can do most things that model can do, with a bunch of other improvements and added capability.The original form of this model was trained off of <a target="_blank" rel="ugc" href="https://civarchive.com/models/2583/hassaku-hentai-model">Hassaku(NSFW)</a>, though it's diverged very significantly over the course of multiple sessions of training. As well as my additions, some of <a rel="ugc" href="https://civarchive.com/models/79872?modelVersionId=84846">Brixel Core</a> has been added to greatly improve the range of styles it is able to achieve.Tips for use:<ul><li>I recommend using 40+ steps, at least with Euler A. You can get away with fewer, but it leads to more inpainting</li><li>I use <a rel="ugc" href="https://civarchive.com/models/71961/fast-negative-embedding">FastNegativeEmbedding</a>, which includes a hand improvement. If you don't use this, or even if you do, a negative embedding such as one of the badhand embeddings can improve hands</li><li>"detailed face" is recommended for non anime face styles.</li><li>Trained with primarily booru tags. If you're having trouble getting a result, figuring out what it would be on danbooru may help.</li></ul><s>Image workflow and settings for examples: </s>(Most of the initial gallery of examples are no longer inpainted, as I've improved my GPU and no longer have to generate at quite as low resolution. <ul><li><s>Euler A generating at (usually) 512x512 with 36 steps at CLIP SKIP 2</s></li><li><s>Upscaled 1.5x using the 4x-UltraSharp and Hires. fix</s></li><li><s>Quickly inpainting faces that required it using the method above, and in a couple cases a second pass to fix a wonky eye or small detail</s></li><li><s>No inpainting of hands or held objects was performed.</s></li></ul>Nerdy stuff you may or may not need to know:<ul><li>WD-VAE-v1 is baked in, inherited from Hassaku. I may experiment and upload a version without a VAE.</li><li>Trained at CLIP SKIP 2</li></ul>

RoseRendered

Update. Some refining of concepts, beefing up of the dataset of various ethnicities, and some extra refinement and variety.

Files with SHA256: 18c9e2d9497f12e7318e97118b6833701da4cdb203f2b35b1bac7287aff5842d