Trained on Kroma 0.2 + Krea 2 Raw. Works "ok" (not great) on raw, I recommend using with Kroma 0.2.General NSFW and Furry. Lots of broad and general concepts.Probably works well on Kroma 0.2-based alternative stuff as well, or Krea 2 in general. Kroma 0.3 maybe not.MULTIPLE RANK VERSIONS AVAILABLE IN THE MODEL DROP DOWN.The rank 256 one is not thoroughly tested.My preferred generation setup:<ul><li>Use Turbo LoRA at strength 0.4</li><li>CFG 2.5 20-24 steps on gradient estimation + beta, switching to euler around 12-14 steps in. Don't use only gradient estimation, it can look pretty bad. Most other Krea 2-specific or Kroma-specific settings should work fine too.</li><li>dpmpp_2m_sde (beta) is good too, and er_sde maybe.</li><li>You can scale down the MLP.down in blocks 0-9 for less "noise" sometimes.</li><li>Res 1024+ (higher res should work decently well)</li></ul> Strengths<ul><li>General NSFW, usually</li><li>Realistic scenes and generally unique photography with depth</li><li>Darkness</li></ul>Prompting (not required, but helps coherency and unlocks better features)<ol><li>Prose, 2 to 5 paragraphs, in plain, specific sentences. It names the species, the fur and skin colours, the pose, where each body part is, and exactly what is inside what. Anatomy is named directly: penis, shaft, knot, sheath, testicles, vulva, anus, nipples.</li><li>Tint line (optional): a short comma list for lighting, colour and rendering.</li><li>Tag line (optional): "by &lt;artist&gt;.", then pose, act, species and setting tags. Anatomy and penetration tags were removed from the tag line on purpose; that information lives only in the prose.</li></ol>Prompting Specifics/Keywords- Every entity you describe gets drawn. A back-reference such as "the one on the left" can produce a second copy of that character. Describe each character once, fully, and give positions in absolute terms. - Be explicit about anatomy and placement. - Open with a style. For anime: "Digital anime art style with clean line art and cel shading." For photos: "hyperrealistic, realistic." End with a tint line, using the words the model was trained on, any combo of these: - dim lighting, brightly lit, very dark scene, deep shadows, crisp highlights, high contrast, low contrast, pitch black - vivid colors, muted colors, near monochrome, color wash, oversaturated colors, warm orange tint, golden tint, cool blue tint, purple tint, red tint, teal tint, olive tint - flat cell shading, painterly shading, smooth gradient shading, bold lines, smooth clean lines, sketchy lines, lineless, textured, heavy texture, very smooth, very detailed - clean composition, minimal composition, clutteredArtists<ul><li>Not too many artists trained in depth. Here are the ones that MIGHT show an effect, from most examples trained on to least:</li><li>by hakya11, by johnwulffe, by furflection, by braeburned, by mawmain, by dagasi, by cold-blooded-twilight, by r-mk, by drako1997, by feretta, by zackary911, by claweddrip, by zourik, by codyblue-731, by oro97, by truegrave9, by loimu, by hooves-art, by sincastermon, by keadanger, by the-minuscule-task, by tom fischbach, by pikajota, by dimwitdog, by miso souperstar, by magnetus, by jishinu, by dragonplayer, by securipun, by twang, by sijimmy456, by riska (artist), by nepentz, by felino, by backsash, by taran fiddler, by photonoko, by inu-sama, by cooliehigh, by angellsview3, by zaush, by jay naylor, by desertkaiju</li></ul>Issues<ul><li>Does not really try to expand much on broad or diverse art styles, some capability there from base model may be lost.</li><li>"Anal" stuff is still kind of a WIP. More versions to come with improvements likely.</li><li>May not work super well with other loras at high strength</li><li>Sometimes blends two character features together a little bit, especially if they are stacked</li><li>Readable text and sound-effect lettering, sometimes, and watermarks</li><li>Double anuses</li></ul>How it was trainedAbout 24,500 unique images, mostly anthro/furry, plus human and humanoid characters across cel/flat, painterly, 3D and photo styles. Each epoch samples about 34,500 rows, because some sets are repeated. Training was split across GPUs between two bases, Kroma v0.2 and raw Krea 2. Every image has a long, Gemini 3.8-built caption, most were double checked by other models and a lot by me specifically in areas where AI vision isn't great.Credit in merges plz so I know if anyone has a use for it at all

Files with SHA256: 4f4b4b94065045dea9d1c38dec442528f48925576056bed15896c36fb2d0f80f