Welcome to the Wild Safari!!!
This model is designed for variety. It depicts humans, anthros, ferals, and semi-humans together or individually, with the widest variety possible of types and styles while keeping backgrounds as detailed as possible.
I'm aiming for the highest quality possible with the fewest AI-related artifacts, so images don't need fixing or require minor inpainting.
It does incredibly good with western-styled digital art, pixel art, cartoon, 3D, and anime, but it can also work with realism. It's not designed for it, but it can generate images, albeit not as well as dedicated realistic models.
Recommended settings:
Steps: 20-30
CFG scale: 3.5
Sampler: Euler a (it will not work properly with others)
Description
Increased dataset to 14GB
FAQ
Comments (28)
Will there be the ~6.5GB variant of V2.0 or will it remain at the ~13GB?
Is there an issue with 13gb? Speaking from ignorance. Does it have any problems with certain lora or UIs? From testing, I couldn't find any difference in speed between the 6gb version and the 13gb version. I do have a 6gb version.
@thebiologist92137 Storage space, and VRAM use that I'm aware? For me it's more storage space. I have a 16GB VRAM GPU so that's not much an issue, but I also have way too many other checkpoints.
@Kaladae I guess I'll generate some example images with the 6gb model and upload it as a variant.
@thebiologist92137 That would be cool. Don't upload it if you don't want, but I know a lot people stick to the ~6GB versions cause that's what they can run on their hardware. And the fewer that remain on SD1.5 because potato.
@thebiologist92137 4070 и 32 ОЗУ запустилось.
Why fp32?
The overall quality is much better, from what I could test. I had a 6gb V2, but after increasing it to 13gb, it was outputting far more consistent and high-quality images, especially shadows, backgrounds, and text.
That said, I have a fp16 version as well.
I'm upset this one went past my radar for as long as it did. This is really good 👍
Thank you for your comment and the images provided. They're good.
Underrated model. I saw it but didn't mind testing until recently, good thing i did. Its pretty consistent and has a good range of knowledge thanks to its size, both in anatomy, characters and artists styles.
If you can run it, it's definitely a good idea to try. Eager to see updates 👍.
Thank you for your comment.
Previous model worked great, this version seems to render grayed/washed out images. I'm not entirely sure why. I'm using ComfyUI.
Try a different VAE? Or too long negative prompts can sometimes do that as well. I find if there's a word in the positive/negative that the model doesn't like or confuses it, it can break.
I know exactly why. You need to set a VAE if you're getting that.
The model does not have a baked-in VAE. You need to add one. I wanted to allow people to choose one.
Thank you, manually setting the VAE helped. I am getting another curious issue though, trying to set values such as (this:0.7) for reducing the strength of a token seems to cause the images to spiral out into a blobby mess of colors. It works normally on most other models and I'm not quite sure why it's happening.
@antiman9 I think brackets like [this] also have a similar effect. Try that maybe? If I remember, that could be a setting in your UI, or it could also be the keyword itself may be confusing the checkpoint. Some checkpoints don't like certain keywords. Misspellings can also cause that sometimes, from my experience. Also, NoobAI models seems to be more sensitive to keywords and weights than SDXL/PONY/Illustrious. That might be part to do with it. Noob models can create some great results, but they do seem a touch more finnicky to prompting.
Struggles with hybrid cocks. Not a surprise as the majority do. Doing enough piecemeal Loras causes it to deepfry. I like this model enough that I'm finally going to try and figure out how to do a Lora lol
Went and did it 😤
could be the lora or artist style. I get them to generate without issue or fail with "horsecock, knotted horsecock, knot". It also does cetacean penises, simple tapering penises, feline penises, dragon penises, and other various hybrids pretty well. I have published some examples,
@thebiologist92137 hmm Could be something I'm overlooking. I also got in the habit of not using "horsecock" back in the day because that would almost guarantee a horse character, which is likely an outdated practice
@nenten Yeah, horsecock, horse penis, and equine penis will NOT create a horse character in this model if you specify another character. In my experience, if the type of penis is not being properly created, it means either the artist or the art style prompted have little to no examples, or lora are interfering because they are trained with other types of dicks.
@thebiologist92137 Yeah. Your species adherence is great. Definitely a prompting habit/Lora reliance problem on my side.
Can you please share a workflow or something. cuz I'm not getting anything close to the sample images. I'm using the V2. I am using comfy btw.
Wait, i fixed it. I guess it does not include a VAE. I used the sdxl VAE and it is working fine. Can you suggest any other VAEs?
I use the standard SDXL VAE, but I did not bake in any base, so all should work fine.