DreamS Archive

<h1 id="heading-350">SketchyDreamS Rev2 (SDS Rev2)</h1><p>Updated recipe of SDS, less damaging to understanding, so it follows prompt better.</p><p>It is less suitable for expressive sketches and lineart, if you want them - use first version.</p><p>S12 used is v4.6.2, trained on 768 reso, so i would recommend to step up your native gen to 640px, like 640x960, or even to 768px.</p><p></p><p>While it is less abstract and expressive, it is capable of more sophisticated composition, because it doesn't get garbled up in sketchy stuff that would make trypophobes delete this model immediately.</p><h1 id="heading-350">SketchyDreamS (SDS)</h1><p>Overhaul, rather than a direct continuation of ElusiveDreamS. Respects anime style much better, increased compatibility with loras, etc, etc.</p><p>I love how it works with sketchy styles, well, that was one of the points when i made it.</p><p>It utilizes Dreamwave v2 and Sketchy(unreleased) loras, + polishing by S12.</p><p></p><h1 id="heading-2356">ElusiveDreamS (EDS)</h1><p>This model is a sort of continuation of ACMAR G Mix, as this is it, but spiced up with 3 trainings on top of it, combined using deep merge in 3 steps.</p><p>S3Full - Full merge of S12 on OUT0-11. Might have less leak, more abstraction, and a bit better control, but will surely have higher presence of various particles. Requires stron negative on (3d, realistic) to produce flat, or positive flat color.</p><p>S3Half - Half merge of S12 on OUT0-11. Might have less control, but will be less leaky in abstraction, i think. Doesn't demand negative (3d, realistic) to say 2.5d/anime.</p><h2 id="heading-2357">Highlights</h2><p>Usually quite aesthetically pleasing.</p><p>Partially capable of high level understanding of combination of various concepts, thanks to S12 being a cherry on top of this. + Possibility of high level of compositional abstraction.</p><p>Absolutely fire sketchy look, imho. <br /><br />Refer to S12 for possible new capabilities.</p><h2 id="heading-2358"></h2><h2 id="heading-2359">Downsides</h2><p>```Note: downsides are not stacked, but unique for each model, so read only section for one you're choosing.```</p><h3 id="heading-1919">SDS rev2:</h3><p>Honestly, not much. Solid model that is now my favorite for generating wide variety of semi-abstract things, or unique compositions.</p><p>It might be too abstract and a bit unstable compared to common models, but other than that it's very solid.</p><h3 id="heading-1920">SDS:</h3><p>Meant for extremely sketchy things, so low-level(small) details are lost in most cases, and attempts at detailed background are met with abstract walls of color.</p><h3 id="heading-1921">EDS:</h3><p>Quite rigid by default, so you will need to use more detailed prompts, which most of you already do anyway.</p><p>Is biased towards female characters, so you'll have to increase weight, and/or put appropriate tags in negative to have consistent male output.</p><p>Bias towards flowers. While they are aesthetically pleasing, sometimes you don't want them. Use nagative <code>flowers</code> for that.</p><p>Possibly could be hard to adapt to loras, but that needs testing.</p><p>Best used for 2.5, or 2d, as it's a bit f'ed up in terms of faces, if you promt <code>realistic, 3d</code></p><h2 id="heading-2360">TLDR Information</h2><p>This model is based on ACMAR G Mix, and both EDS and SDS contain 3 additional trainings, which were added through deep merge technique:<br />EDS: Dreamwave(Full), ElusiveWorld, S12(4.2 NAI)</p><p>SDS: Dreamwave(v2), Sketchy(unreleased), S12(4.2 NAI)</p><p>SDS rev2: Dreamwave(v2), Sketchy(unreleased), S12(4.6.2, unreleased)</p><p>EDS: Dreamwave and Elusive are used for dreamy looks and aestethics, while S12 is a polisher on top, that adds conceptual understanding. It is merged at very high percentages at some, and with gradual stepping on rest to stabilize style at anime, rather than what Dreamwave and Elusive look like, while adding control over output style.</p><p>SDS: Sketchy and v2 Dreamwave provide exceptional sketchy look, if you want to get it. That model excels at such. It is, however, not good at deep details, as they are going to get lost in sketchy style. Fixed in rev 2, but sketchy no more.</p>