A new take on an old MIO and follow-up to a more recent MIO2.
Problem: As I see it, decent antrho renders are very prompt heavy. Without using 3-5 negative embeddings, sometimes 2 or more positive lora's and embeddings, it's very hard to get a realistic anthro that's more than just a human with an animal head. Oftentimes, even with all the tools the prompt gets extremely bloated by the time you get all the prompts required to combat this. Even when using a realistic anthro checkpoint this can happen.
Solution: I did what I have done in the past. Render hundreds (I guess thousands now) of images the way I want them and train them into a lora thus increasing the image quality without the need for so many of the prompts we add by default, (masterpiece, best quality, 8K, HDR, etc, etc.) and in this lora also eliminates the need for so many references to "furry body, furry face, animal features, etc."
Sample images are purposely rendered with minimalistic prompting (typically under 30/75 for positive and 15/75 for negative) to illustrate how little is required to get close to the desired result. The images used for training were heavily fantasy influenced so armors and peasant style clothing should render easily with other styles depending heavily on the checkpoint used.
Potential improvements: No more planned at this time. See "About this version" under MIO-Furry4.
Description
Horses, koalas, kangaroos, pangolins, dragons, and snakes?
This update includes a few more furry friends, and some not-so-furry critters. The "anthro" tag is still trained into the Lora, but as a result of the now 5822 training images used, it is no longer required. Simply prompt the desired animal, add the lora, and sit back and watch the fur fly.
So far in my testing, I have not run into an untrained animal that it can't get at least close unless the checkpoint used is not trained on it. However, I will list the new additions here rather than in the tags list as it would be quite long with the existing animals and the new ones.
First, the new furry critters. Kangaroo, koala, mouse, otter, porcupine, rat, squirrel, horse, and pony.
Now the non furry. Alligator, axolotl, frog, gecko, iguana, Komodo dragon, Naga, pangolin, dragon, and winged dragon.
With the exception of horses, only female variants of these animals were trained but it can handle males quite well also.
With specific animal tags no longer needed, I will limit the trigger list to specialized triggers needed to help with things like Naga anatomy.
FAQ
Comments (9)
This is phenomenal and solves an issue that I've always run into with furry models -- like you said, human-shaped bodies with animal heads were mostly what I'd get unless I prompted the living heck out of it. They'd usually look more like caricatures rather than what I was aiming for. With some merged checkpoints that I have made, I could at least get them to be realistically textured fairly easily, but they always came out like caricatures.
Minutes after installing this LoRa, and changing my prompt a little bit to remove some magnitudes, DEAR LORD the quality bump in the results were insane...not only are they coming out way more detailed now, further improving the photoreality I mention above, but they're actually proportional bodies now, and I flippin love it.
Amazing job, and thank you so much.
Love to hear success stories. Thank you.
hey have take request lora you creatinng plz?????
Generally not, but if it aligns with something I'm planning or interested in I might. Currently I'm considering improving 1 or 2 older ones.
@Avenka have creating lora anthro version 5 dragon aspect for sdxl aka pony
@Avenka when maybe lora character ysera and alextrasza for world of warcraft
@fog1 I "might" look into training it for pony. MIO-Furry4 took 2-3 hours due to the magnitude of training data so XL would likely take much more. Not to mention I might want to improve the training images for it.
are you going to consider adding birds and insects?
Just in response to your description, one thing I've found works amazing wonders to that end, is to use controlnet with a furry image, and a decent realistic furry checkpoint, like indigo, or another I can't remember the name of.. anyways, a decent controlnet model works wonders. You just need an image of a character that looks close to what you want, in the position that you want, then enter the prompt of what you actually want. I just made the cutest realistic tigerboy yesterday using an image of Wilykat(2011). Though it may be a bit slow if you've got under 16gb vram, cuz it'd likely have to offload to system ram.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.