IL v2.3 : New version, more flexibility. Better than IL v1
This version don't have a baked in VAE. I recommend to use SDXL VAE
Please consider supporting me so I can continue to make more models --> https://www.patreon.com/BB95FurryMix
Don't forget to join the Furry Diffusion discord server --> https://discord.gg/furrydiffusion
Since v3, this model uses e621 tags.
This model is a mix of various furry models.
It's doing well on generating photorealistic male and female anthro, SFW and NSFW.
I HIGHLY recommend to use Hires Fix to have better results
Below is an example prompt for the v7/v6/v5/v4/v3/v2.
Positive:
anthro (white wolf), male, adult, muscular, veiny muscles, shorts, tail, (realistic fur, detailed fur texture:1.2), detailed background, outside background, photorealistic, hyperrealistic, ultradetailed,
Negative:
I recommend to use deformityv6 and boring_e621_fluffyrock_v4, you can add bad-hands-v5 if you want
Settings 1.5 :
Steps: 30-150
Sampler: DPM++ 3M Karras, DDIM, UniPC or Euler A
CFG scale: 7-14
Size: from 512x512 to 750x750 (only v4/v5/v6/v7/v8/v9), up to 900x900 for v10/v11/v12
Denoising strength: 0.6
Clip skip: 1
Hires upscale: 2
Hires steps: 30-150
Hires upscaler: Latent (nearest)
Settings XL :
Steps: Below 30
Sampler: Euler A
CFG scale: 3-6
Size: 1024x1024
Denoising strength: 0.2
Clip skip: 1
Hires upscale: 2
Hires steps: Below 30
Hires upscaler: Remacri
Description
This is a Special Edition of BB95. It's trained on a small, curated dataset of very high quality images, with the hope of improving the composition quality.
Most of the time, it will be better than v11. But in some cases, it will be worse than v11. It's not good enough to be a full, proper v12 with the kind of quality that's expected. It's a side-grade from v11.
In testing, it did better at higher resolutions, and the composition was better. Give it a try, see if you like it, and leave a review! If a lot of people say that it's an improvement, maybe the training can continue on a larger dataset.
Thanks to Dogarrowtype for the idea, and for helping me collect the dataset.
FAQ
Comments (13)
I'm not going to leave a review for this version just yet, but I thought I'd say I'm not really noticing any obvious differences between the V110 and V110SE, nothing that would make it stand out as superior or inferior at any rate. I suppose I would have to make some comparisons at higher resolutions.
Yeah but it's much harder to find an evolution in actual models
I hope soon to have SDXL furry models
I noticed a definite change between the two, which is a certain game-changer for anyone working with color balance and intensity... 11.0 SE is MUCH more vibrant in things than the original. If you're looking for a more matte finish on things, this is something to note! I made a collage of a batch done the same way under each one, but I don't know how to present it. Still, if you make your own batches, I'm sure you'll quickly see what I mean.
Put simply, SE has a detrimental effect on some attempts at a realistic effect... the more matte result usually looks more believable... but if you're going for a vibrant eye-catching gloss and shine, SE is absolutely the ticket! I can totally see both 11 and 11SE being useful for different scenes.
Just want to say amazing effort on this model, having some good results with 11 SE so far. Thanks for your hard work!
Your model is great, doesn't take too many prompts to generate real cool stuff. Definately not requiring some over-engineered prompts or plenty of LORAs. Big thumbs up and keep up the great work!
This is THE best model hands down for standard fur, more experimental, and human. I cannot thank you enough for a checkpoint that is flexible, polished, and doesn't force you into a specific style, gender or species. From anthro, to slimes, monsters to aliens, and everything in between. Seriously, THANK YOU.
What sampler settings is everyone using for this model? I get fantastic high detail images with DDIM at 120-150 steps (at both 512px and hires) but this obviously uses a lot of time.
DPM and UniPC are also suggested in the model description and I see people using those here. From my understanding with those samplers, after a certain number of steps there isnt any significant convergence or detail, so what is the ideal step range for those samplers?
Would love to hear what other people are using and the results they get. Ive done some experiments with DPM++ 2M at around 80 steps with decent results, although it looks a lot more anime/cartoon like than my usual DDIM 120-150 process, if that makes sense.
Actually I'm using DPM++ 3M SDE Karras at 75 steps
@Bumblebee95 Thanks! Ill give it a try
So use this as a refiner mostly?
and Inpaint as well :D img2img
@kikoshi399 you can use it as a base model too
This... thing... is dangerous. I've barely left my house in months. This could mean the end of furry porn sites as we know them.
I predict that this model will be used as a weapon of cultural subversion in international warfare, with the potential to cripple the reproductive and economic output of opposing countries, if it hasn't already.
Looking forward to the next version.