Type | |
Stats | 184 0 |
Reviews | (22) |
Published | Feb 26, 2024 |
Base Model | |
Hash | AutoV2 A9B70917B5 |
KPC-Rv2
This model is well trained to make 2.5d anime-like images right out of the box with minimal prompting. I've been working on this model for a few months now, and it's the result of crafting my own "style". Heavily inspired by artists like incase, and the styles of spiderverse/arcane. KPC-Rv2 makes what appears to be true digital paintings, with a serious sense of artistic style to faces and people. The model is both SFW and NSFW capable. This is an update to my previous KPC-R model and focuses on honing in on a painterly style with a ghibli or cutout color style while also improving the legibility of details and staying on-prompt.
NEGATIVE PROMPTING:
Negative prompting will produce far better results. For the image above, I used the following negative prompt:
(ng_deepnegative_v1_75t, easynegative, bad quality, low quality, bad artist, (deformed:1.5), cropped, out of frame, (bad hands, badhandv4:1.5),gray skin, young, tattoos, child, ceiling, indoors, Asian:1.4), blurry, amputee, monochrome, fat, cold, thick thighs, [high contrast], (!text:1.2!), watermark, (pubic hair), bad-picture-chill-75v
Definitely use easynegative all the time, and bad-picture-chill-75v for more photorealistic results.
Furthermore, the prompt tends to generate younger looking people, which I'm not a huge fan of, so prompting against it is advised.
RECOMMENDED GENERATION SETTINGS:
Personally, I prefer DPM++ 2M Exponential, the more steps the better, at a CFG scale of 8. You can use other samplers of course, but this will provide the best quality & consistency.
For upscaling, latent models will give you a more soft/stylish look, but R-ESRGAN 4x+ is best overall, as it provides more consistency. For models other than those provided by A1111, 4x-Ultrasharp is my favorite, and my preferred model for upscaling.