This is DynaVision, a new merge based off a private model mix I've been using for the past few months. The output is kind of like stylized rendered anime-ish... I guess? I don't know how to classify it, I just know I really like it, and everybody I've let use it really likes it too, and it's unique enough and easy enough to use that I figured I'd share it with the community. It's got noise offset mixed in, so it should be able to create nice dark scenes easily, and really bright ones too. Also, I've found that it's very responsive to Danbooru + clip skip 2. It has the absolutely incredible <a rel="ugc" href="https://civarchive.com/models/58390/detail-tweaker-lora-lora">Detail Tweaker LORA</a> baked in, but it still responds well to adding EVEN MOAR DETAILS with the LORA I've found so go nuts! 🥜🥜For the samples I'm using my own embeddings, <a rel="ugc" href="https://civarchive.com/user/socalguitarist/models">you can find them over on my Civit page</a>, along with a few new ones that I'll be uploading soon to my toolbox. I hope you like this model, I built it with the intention of producing images that are less photorealistic and more artistic stylized output (seriously, it's in-progress name was "ArtsyFartsy"). No face inpainting, no ControlNet, no input masks, all of the samples you see are raw, the only thing I use is High Res Fix for most (though not all). I get best results using DPM++ SDE Karras - Check the gen info on all of my sample images, it's all there. Enjoy! 💖

DynaVision

This model works best at 768 x 768, and I've tested it with landscape imagery all the way up to 768 x 2048 and was still getting coherent output with little repetition. The output at 512 is okay, but it really shines at higher resolutions, and works great for inpainting as well.This version adds a bit more detail and fixes some things that were annoying me about eyes and teeth from v1. It keeps the same general style and look as V1, and should hopefully continue to produce high quality stylized images that are easy to use. This model still leans towards females, but feels less biased overall than v1, and I have no problems "forcing males" just by adding "male" to my prompt. Fingers still aren't perfect, but easy enough to inpaint.

Files with SHA256: 3574ee48997acace773fc09a61a11c8f7e30ce8e61127e457dbf8855b363ec89