<h2 id="ultimate-the-definitive-version-v3">ULTIMATE - The definitive version v3</h2>Check on "About this model" on the tab at the right side. Thanks.______________________________________________________________________¬ The best sizes are combinations of 832 up to 1344 (vertical, horizontal, square). Greater values often generates artifacts or gray borders. ¬ Steps from 10 to 15. Less or more decreases quality¬ CFG scale should be from 3 (bit more realistic) to 4.4 (hard limit, since it starts adding artifacts) Try values from +-0.2 ¬ Sampler Euler Ancestral for quick pictures, DPM++ 2M for better quality (slower)¬ Scheduler (from better and slower to regular and faster): sgm_uniform, Karras, normal¬ Use facefix always (with medium shots or closer, it behaves better)Standard positive Prompt:<pre><code>( score_9, score_8_up, score_7_up, rating_safe / explicit),</code></pre>Negative Prompt:<pre><code>worst quality, bad quality, jpeg artifacts, source_cartoon, blurry, lowres, censor, source_cartoon, (watermark, bad anatomy, long chest, bad torso, missing fingers), (bad hand, bad feet, extra fingers, extra toes), (moe themed, chibi, cute themed), score_4, score_5</code></pre>Very last version. No more upgrades so enjoy it._________________________________________________________<h3 id="last-version-2.5d">Last version 2.5D</h3>~ After 3K of 2.5D images, after merging "Better clothes", "Better breasts", "Better lighting" and after some trial and error, here I come with the last (and final) version of this model. Since no one is using it, I don't see a reason to fine tune further. ¬ The best sizes are combinations of 768, 896, 1024, 1152 (vertical, horizontal, square) ¬ Steps from 10 to 15. More than 16 tend to ignore prompts, add artifacts and supress LoRas¬ CFG scale should be from 3 to 5.5 (hard limit, since it starts adding artifacts or burning the skin)¬ Samplers Euler Ancestral for quick manga-like pictures, DPM++ 2M Karras for better quality (slower)¬ Scheduler: simple of sgm_uniform¬ Use facefixStandard positive Prompt:<pre><code>( score_9, score_8_up, score_7_up, source_anime ), </code></pre>Manga-like positive Prompt: (always Euler Ancestral with this)<pre><code>( score_9, score_8_up, score_7_up, source_anime, B&amp;W, grayscale), /and after all prompts/ (Fix anatomy)</code></pre>Negative Prompt:<pre><code>worst quality, bad quality, jpeg artifacts, source_cartoon, blurry, lowres, censor, source_cartoon, (watermark, bad anatomy, long chest, bad torso, missing fingers), (bad hand, bad feet, extra fingers, extra toes), (moe themed, chibi, cute themed), score_3, score_4, score_5</code></pre>Negative Prompt for manga:<pre><code>(color:1.2), (watermark, bad anatomy, long chest, bad torso, missing fingers), (bad hand, bad feet, extra fingers, extra toes), (moe themed, chibi, cute themed), censor, source_cartoon, score_3, score_4, score_5</code></pre>- - - Outputs are now a bit 2.5D too - - -CORE: Pony + Illustrious Prompts suggested: [before prompts] ( score_9, score_8_up, score_7_up, source_anime, B&amp;W, grayscale) [after all prompts, only if you're making humans / furry] _Flat (fix anatomy) STEP: 10 ~ 16 CFG: 2.0 ~ 4.0 EULER A - Sobre el modelo / About the model - * Tomé una vieja versión hyper de uso libre desde Github, le agregué algunos LoRa de foco en manga que hallé en Hugging Faace así como algunas imágenes en blanco y negro o escala de grises en muy buena calidad de fan arts de Chainsaw Man, Dragon Ball, Gantz, Berserk y Kagurabachi. No te sorprendas si actualizo en la semana, porque me lleva muchas horas (mi GPU es muy chica), pero estoy añadiendo 600 imágenes manga más así como 200 de comic. * Took an old hyper version (of free use) from a github repo, added some manga LoRa from Hugging Face as well as some high quality B&amp;W or grayscale fan arts of Chainsaw Man, Dragon Ball, Gantz, Berserk and Kagurabachi. Expect a new upload in the week, since making this takes several hours and I'm low on GPU. I'm adding 600 more manga pictures as well as 200 comic images.

PonyMan [Hyper]

First of all, thanks to FeuerWalkure, AddemiHosze and TomoKozaki for your support to this model. Now, I think I fixed what needed to be fix and improved what I was asked for.¬ The best sizes are still combinations of 832 up to 1344 (vertical, horizontal, square). Greater values often generates small artifacts or gray borders. ¬ Steps from 10 to 17. Less decreases quality, more start adding noise.¬ CFG scale should be from 3 (bit more realistic) to 5 (hard limit is 6 with 17 steps) Experiment jumping 0.3 in any direction ¬ Sampler Euler, Euler_ancestral or DPM++ SDE¬ Schedulers simple, Karras, normal. Combine those 3 with different samplers to get interesting outcomes.¬ Use facefix in full body or long shots (not needed with medium shots or portraits)Standard positive Prompt: (only add "Rating_explicit" if needed)<pre><code>( score_9, score_8_up, score_7_up),</code></pre>Negative Prompt: (add "3d, real" if you want flat anime)<pre><code>bad quality, worst quality, worst detail, artifacts, signed, censor, (bad hand, bad feet), long chest, chibi, cute themed, jpeg artifacts, lowres, watermark</code></pre>

Files with SHA256: 736f1de8d477e8e8edaa92c9935882953555b931fd142f854302b89744cef9cd