IcedCoffeeXL Continues the general style of my 1.5 Model IcedCoffeeMix, it is more limited in terms of species that it can create but the ones it can do it does better with less About this versionChanges:<ul><li>Merged in <a rel="ugc" href="https://civarchive.com/models/131784?modelVersionId=159512">FurtasticXL --Beta-EPOCH 3 </a>as such, this is no longer a purely trained model. I'll continue training it but will also be considering merges as well.</li></ul>Issues:<ul><li>It has problems with men, not gunna lie. In fact, in many cases I was unable to get ANY male characters to generate, not sure why but it happened. I'm working on bringing them back (with nudity!) as soon as I can.</li></ul>Example:<ul><li>Prompt: Stylized cinematic anime artwork, furry, Dingo, female, fit, (loose shirt, downblouse), side-tie bikini bottom, detailed face, detailed digital acrylic painting, color graded, studio ghibli, highly detailed</li><li>Negative: Blurry, ugly, deformed, deformed hands, bad anatomy, extra limbs, western comics, photograph</li><li>Sampler: I am using KDPM2, DPM 1s, and Euler A and they work best for me in SDNext, but you should play around with what works best in your environment of choice.</li><li>Steps: 30-40</li><li>CFG: 3-9,</li><li>Resolution: 832x1280</li><li>Seed: 1597256910</li></ul>I do not recommend clip skip or using the refiner (I haven't tried it, but it really doesn't need it)Note: the term anthro isn't really necessary as it has a really good grasp on furries, but you can add it to reinforce it.

IcedCoffeeXL

While this is the first version I'm uploading its actually the 3rd version overall. I started by finetuning <a rel="ugc" href="https://civitai.com/models/119012?modelVersionId=138828">Blue pencil XL v0.0.9</a> for 2500 steps using 1000 high quality images but no tags, that got me closer so I trained a lora that I'll upload as well using around 150 high resolution and high quality images with individual captions and tags for 10 epochs at a batch size of 16 with 5 repeats then merged that into the model with a strength of 1. I've trained it on a variety of subjects and styles, so it can do quite a bit including NSFW Content and it pretty nails female anatomy at this point (though its still pretty bad when it comes to males, sorry guys I'm working on it). I recommend trimming back your tags from what you're used to with 1.5 models. Example: Prompt: Stylized anime artwork, (A female wolf, remarkably iris, detailed face), digital painting, pallet knife, photoshop acrylic painting, anime studio, subsurface scattering, HDR, masterpiece, best quality, detailed, absurd resNegative: blurry, poorly drawn, out of focus, worst quality, low quality, simple background, deformed, bad anatomy, bad hands, extra limbsSampler: I prefer Euler A or UniPC they work best for me in SDNext, but you should play around with what works best in your environment of choice.Steps: 20-30, it doesn't need much.CFG: 3-5 will get you more creative/abstract stuff, 6-9 will get more precise. I've gone all the way up to 14 and it works but its really intense, you'll want to add in some negative quality tags like fried/overcooked/etc. For the example image I use 6.Resolution: 832x1216Seed: 244841152I do not recommend clip skip or using the refiner (I haven't tried it, but it really doesn't need it)Note that the terms furry, anthro, etc aren't really necessary as it has a really good grasp on furries, but you can add them to reinforce that, also if you find yourself getting smooth skin furries try adding human, skin to the negative prompt.

Files with SHA256: a2cf9d3debfdc7245babafdb771f39f21791d4734317e62695b25f4a2ec862db