<h2 id="heading-824">READ THIS!</h2>If you've previously downloaded IcedCoffeeMix V60 before 03/24/2024 I had originally uploaded the wrong version of the model, resulting in good upscale performance but terrible initial generation results. I've fixed the issue and uploaded the correct model, please redownload the model if you were having issues with my last version.<h3 id="heading-825">About this version</h3>Pretty big update, but for strange reasons. I was messing around merging some of my older models into ICM 5.0 and was able to find a really good combination with ICM 5.0 and ICM 3.0, I also subsequently merged in the Add Sharpness lora I have linked and it also boosted quality by quite a bit.A few notes:-Faces in general seem to be improved, eyes especially.<s>-It's MUCH easier to generate humans without anthro features while still being ridiculously easy to prompt furries, add "human, skin" to the negative prompt and "furry, </s><s>species"</s><s> to the positive prompt. If you want a human just prompt without the furry tags in the positive prompt and add "furry, anthro, animal ears" to the negative.</s><h3 id="heading-1221">UPDATE: 3-25-24</h3>-current version is capable of both human and furries without adding anything extra to the negative prompt, if you want a human just describe the character, if you start seeing furry features just add furry to the negative prompt. For furries you only need a species tag. If you want more defined furry features you can add those tags as well (paws, snout, whiskers, wings, feathers, etc), and of course adding furry to the prompt will enhance the effect and more or less get rid of any skin patches.One major change is the usage of the add detail Lora, it no longer produces muddy, overly messy results and instead simply enhances the anime effect by adding more features. This does mean it's harder to get a 2.5D effect, however it is still possible with the right tags, and of course if you want to use a Lora to add the effect that works too.<h2 id="heading-1222">Example prompt for the Fennec Fox in a leather jacket:</h2><h2 id="heading-1223"></h2>Prompt: A beautiful female fennec, snout, aged brown leather jacket, canvas pants, white undershirt, black collar, sly, awesome pose, detailed face, detailed eyes, gorgeous, short (wavy) hairStylized anime screencap, masterpiece, best quality, absurdres, solo, by sligarthetiger, by meesh, by avante92, by ashblueflames, by djpuppeh, by personalami, by kenket, by desubox, by gideon, by taran fiddler, by chunieNegative prompt: blurry, out of focus, worst quality, (bad anatomy, extra limbs, extra ears), hands at side, skin, human, baldSize: 800x1200Hires upscale: 2Steps: 35Second pass: TrueHires Steps: 20Sampler: DPM++ 2MHires Sampler: Euler aSampler options: rescale betaCFG scale: 6Clip skip: 2CFG rescale: 0.7Second Pass CFG scale: 6Hires force: TrueHires upscaler: RealESRGAN 4x+ Anime6BFace restoration: Face HiRes, Denoising strength: 0.3HDR clamp: 0.95/2.6Seed: 2111518298FreeU: b1=1.2 b2=1.4 s1=0.9 s2=0.2Hypertile UNet: 384Please note that I am using SDNext for my images, so your results will probably not be like mine unless you are using SDNext with the exact same settings as me, I utilize FreeU and HDR Clamp and Center which both affect the final output. If you're not getting the same results as me, that is probably why.

IcedCoffeeMix

edit 06/24/23 I've taken down the model for 2.0, the more I use it the more I realize it's not up the the same level of quality as v1.0, once I've fixed those issues I'll re-release 2.0.This updates quite a bit over the base model, first I retrained the original model by tweaking the percentages slightly, then I used that to generate around 250 high quality images with a variety of subjects, compositions, and overall vibe in a specific style which I then used to train a lora that I am affectionately referring to as "creamer" (more on that later), that was then merged into the model at 75% strength. After that I pruned the model with ema-only, fp16, and fix clip settings. The style has changed a bit from v1 of the model, but I think it gives consistently better results.As for my Creamer LoRa, it produces similar results when used with my v1 model, I haven't released it but would be open to doing so if anyone is interested and prefers v1.

Files with SHA256: 40979680981e988b55ad28303a4f470646685309a9aabd57e1c95f354a47aebd