Anime on Furry Models

Or: How I learned to stop worrying and take the knot

NEWS:

  • a9 is out now "Tagging is closer to normal booru tags, no more "humanoid, not furry". Colors should be better. Add "bright" to negatives if it's too bright."
  • Furception vae released
  • ZSNR noise scheduler added to dev branch but it has issues with karras and working under 30 steps

Base Requirements

  • Get a copy of easyfluff
    • latest is 11.2
    • if you are using HLL LyCO, it is trained on fluffyrock so other derivatives work and there may be pros and cons to other models
    • fun and funner editions allegedly help with color bleeding, offsetting cfg rescale, and being better for natural prompting
  • Download the .yaml and make sure its next to the checkpoint with the same name
    • 11.2 yaml right here for you
    • If you got nothing but brown; something went wrong at this step
  • CFG rescale
    • Auto1111
    • comfyui
    • CFG 7 = Rescale 0.7
    • Higher rescale can make images duller,
    • Adjust rescale depending on the prompt, you can often get away with below 0.7
    • DDIM and UniPC do not function with rescale on auto1111 currently

If you are on Comfyui

  • Connect your checkpoint/model to Model sampling Discrete then RescaleCFG
    • sampling: V_prediction
      You can also take the image I have below and install missing nodes using ComfyUI Manager

Oh, so you want anime?

  • Download this Lyco and just add it the way you add LoRA
  • Check below for artists wildcards and .csv files in the grid below
  • CSVs on Auto can be installed as such...
    • place csv in stable-diffusion-webui\extensions\a1111-sd-webui-tagcomplete\tags
    • Settings > Tag Autocomplete > Extra filename

Model Artist Wildcards Artist Examples Tags Changes
a4 Wildcards, Uncounted Examples, Combos HLL CSV tags Vtubers, artists, and anime, all on furry models
a5a Wildcards, Uncounted Examples, LCM Grid HLL CSV tags 206 New artists; slightly better color
a5b ⬆️ LCM Grid, WIP Rentry ⬆️ Better look even without tagging artist
a6 Wildcards, Uncounted x HLL CSV tags 432 New artists
a7 Wildcards, Uncounted - HLL CSV tags 129 Artists, AI-generated tags
a8 Wildcards, Uncounted - HLL CSV tags 109 new artists
a9 Wildcards, Uncounted - HLL CSV tags 68 new artists, better colors,

What settings?

  • Clip skip:
    • -1
    • -2 is fine
  • Resolution:
    • 578 ~ 1088
    • It's pretty generous about handling lower and higher resolutions.
    • 512x768 still viable but I mainly only do that for batch tests
  • Tagging system:
    • HLL lets you use Danbooru/e621
  • VAE
    • You can use whatever vae you already prefer
    • Furception vae is an e621 finetune you can try

Prompting:

  • Prompting artists is generally essential
    • by x
    • by x, by y, by z or by [x|y|z] for artist mixes
  • EF base recommends prompting natural language and refining with tags
    • Anecdotally I prefer the other way around
    • Depending on your base EF model some have different ratios of understanding tags vs natural language
  • The LyCO has quality tags
    • best quality, high quality
    • normal quality
    • worst quality, low quality
  • Some optional tags to keep in mind. They are not essential.
    • Human, not furry, 5 fingers
    • Anthro, furry, fluff, tuft, paws, slit pupils

Drag and drop starter workflows

※ Make sure to check model/VAE/LoRA/Upscaler
Featured is the A9 version with a 640x960 Resolution into a 2x upscale

Auto1111 ComfyUI  

You don't want to use the method above?

EPS versions
LS_Kerberos model

A mix of EasyFluff, HLL, and LS Vividus.

  • It's all baked in so you can skip the above steps
  • Has some issues with hands unless specified
  • A more standard 1.5 base resolution than EF has as vPRED

Based68 model

Read the civitai description for more info


Raw furry model wrangling

  • Toggle refiner and set the checkpoint you want
  • Adetailer
    • "Inpainting" > "Use separate checkpoint" > "Use separate CLIP skip"
    • leave adetailer on the furry model for anything it handles better
  • Prompt
    • Positive: human
    • Negative: anthro, tuft, fluff, fur, slit pupils
  • Optionally put this LoRA into the prompt and set it to around -.6
  • Quality tags aren't trained in base EF. If you feel you need them then this LoRA exists
  • Base EasyFluff artist examples | rentry
Example  

Optional and Additional

Negative embeddings

Embeddings made for furry models specifically.

This doesn't mean you can't use non-furry embeddings

Embedding Effect
Boring e621 Easynegative
bwu blurry, watermark, unrealistic
dfc dull, flat, color
ubbp unbelievably bad body parts
updn ultra-saturated, painting, drawing,

Misc tip section

  • On Auto with lora block weights you can do something like <lora:HLL:1:start=5> in order to get more EF composition
    • <lora:LCM:1:stop=5> being put in the mix especially bakes in the EF and lets the rest of the steps work on the style

Fixing washed out artists

Allegedly the CFG rescale extension does not work with DDIM in Auto but if you like your results then you may want to consider dropping down to 0 on euler or any other sampler as an alternative.
In all honesty while I do in this guide tell you what tools and settings are standard in base requirements, depending on the checkpoint/prompt/artist CFG doesn't need to be at 0.7 for every image as long as you get results you like.

1
2
3
4
5
6
7
8
human, not furry, by kantoku, by mignon, by wamudraws,
BREAK
masterpiece, scenery, lake, tree, landscape, colorful, saturated vibrant,
upper body, 1girl, looking at viewer,
(best quality, high quality:1.4)
<lora:a5a:1>
Negative prompt: (worst quality, low quality, normal quality:1.4)
Steps: 20, Sampler: Euler, CFG scale: 7, Seed: 2082381253, Size: 512x768
  rescale
Auto1111
Comfy
Auto fix:1
Script:1
Auto:.7/Script:.7
DDIM

2D~3D~Real tagging

Version comparison

Consider this a mild wip since I don't have all the models downloaded and I want to decide which versions I should grid along with a few ideal and fair prompts.

1girl, human (no quality tags)

1girl, human, amazing background, woman looking over at lanterns (no quality tags)

LoRA

Nai Trained LoRA

It's worth just shoving your existing LoRA into EasyFluff and seeing what works. Some even do better by adding by [artist]


Training on furry models

mega with guide and model
Annotated powershell script from the furry diffusion discord

Edit
Pub: 01 Dec 2023 22:30 UTC
Edit: 26 Jan 2024 20:35 UTC
Views: 32168