Oh look, an image model with actual prompt adhesion! Well, I've got the tagged images already, so I may as well...There's <a rel="ugc" href="https://civarchive.com/models/2008854/jim-aparo-style-80s-dc">fanboying about the artist</a> in my Illustrious LORA intro, so I'll spare you the waffle in this one and stick to the practical stuff:<ul><li>Trained on Krea2Raw... Very slowly on my machine.</li><li>Sample images were generated in <a rel="ugc" href="https://civarchive.com/models/2734469/krea2centersemiraw">Krea 2 Center Semiraw</a> with no turbo LORA for 16 steps at CFG 1.4 'cause I like neg-prompting. A mix of Euler Ancestral/Simple and Euler Ancestral/Beta.</li><li>Samples are a mix of oneshot gens and oneshot gens with an upscale pass. The upscale pass is very iffy - it tends to over-render things and make flat colours blotchy. That might just be a skill issue on my part.</li><li>The LORA was trained with a jimaparo trigger word (style) and an aparobats trigger word (specific batman outfit), but as far as I can tell they're entirely unnecessary.</li><li>Strength at 0.8 to 1.0. 1.0 gives you proportions that are closer to Aparo's, but has the drawback that it over-learned the characteristics of the '80s comics that comprised the data set. Specifically, crappy '80s colour printing with plenty of bleed. It's accurate, at least!<ul><li>Cut the strength in half if you're going to do a second pass.</li></ul></li></ul>

Files with SHA256: 9b18980db02605c4e1d909f8c7a1a3114f2deef4225e9d805721709fc388236f