<h3 id="note:-i'm-still-figuring-out-how-well-it-actually-works.">Note: I'm still figuring out how well it actually works.</h3>This is an attempt to make a style Lora for <a target="_blank" rel="ugc" href="https://en.wikipedia.org/wiki/Pili_(TV_series)">Pili</a> puppet plays (of which the most well known outside Taiwan is <a target="_blank" rel="ugc" href="https://en.wikipedia.org/wiki/Thunderbolt_Fantasy">Thunderbolt Fantasy</a>). I don't have much experience with SDXL but it seemed like the best choice. Pony and Illustrious just seem like a bad idea, even realism variants enforce their face look way too much - however, for funsies you can just use the SDXL LoRA with them to get some <a target="_blank" rel="ugc" href="https://civarchive.com/posts/10226066">elaborate outfits</a>.Note that this is meant to be a style lora and was trained only on screenshots from the show. It does know the characters (by their Chinese names in ASCII, ex: Shang Bu Huan), but it's probably not going to help you much with generating Dan Fei porn in whatever style you want.<h3 id="notes:">Notes:</h3><ul><li>Better not to use "puppet" in the prompt - it will give you a puppet of a puppet.</li><li>Works best with photorealistic model variants. Juggernaut or DreamShaper give it a nice soft look, but it's likely not what you're looking for. NoobAI etc will bulldoze all over the style with their own. Flux is of course very good at realism.</li><li>It's not good at actually generating specific Thunderbolt characters, but their names can be used as a rough guidance for the general look. This is even worse for Flux, where you'll generally have to describe what you want very specifically.</li><li>Flux version works best with slightly lowered guidance.</li><li>There's a lot of Shang Bu Huan in the dataset, so characters (especially male ones) look a lot like him by default. You might want to put him in the negatives to mitigate this. You might also want to request females specifically, unless you like crossdressing maids etc.</li><li>Only really properly good at portraits it seems. Detail shots become overly artificial, while wide shots just turn into a mess. This might just need better prompting or settings.</li></ul>

Files with SHA256: a05a491c21d5a45aceec0d2f29611550f5b1b872952ff150dee7563ed4d9276a