Loading
Du hast einen ComfyUI-Workflow? Erstelle direkt im Browser eine verständliche Workflow-Dokumentation.Workflow verständlich machen.Jetzt ausprobieren ↗
This is an experimental Anima model based on Anima base v1.0, finetuned on a very small amount of e621 and bara furry data. (Big thanks to a friend of mine who helped me train it!)
Modellquelle
Quellenbeschreibung
This is an experimental Anima model based on Anima base v1.0, finetuned on a very small amount of e621 and bara furry data. (Big thanks to a friend of mine who helped me train it!)
So, being an Anima model, the architecture is not the same as SDXL. It runs on English natural language (NL) + Danbooru-style tags, and there's NO e621 data in the base. Even with my little finetune on top, it's still not great at e621 tags and concepts, the main improvement is the aesthetic style.
Features:
Quellen
1 QuelleModellartefakte
3 ArtefakteindigoFurryMixAnima_v10_txt.safetensors
safetensors · 1,11 GB · SHA-256 cd2a512003e2…3eba · Civitai
HerunterladenindigoFurryMixAnima_v10.safetensors
safetensors · BF16 · FULL · 3,89 GB · SHA-256 2f57f2b218da…b2b8 · Civitai
HerunterladenQuellenauszüge
1 AuszugSource context: 1114 downloads · Type Checkpoint · Base model Anima
indigoFurrymix_anima_v1 keeps most of the natural language ability and the text rendering (you can still prompt it to write words in the image most time).
It also keeps most of the original Danbooru artist styles, plus I've mixed in some bara furry artist styles (no promises on accuracy or how well they actually work, though).
Compared to Anima base v1, this model seems to react more to e621 tags — but I'd still recommend sticking with Danbooru-style tags.
That said, the natural language understanding on this model is very decent. Always feel free to just "describe" any missing e621 concept in plain English NL — sometimes it works shockingly well.
It's pretty obedient overall. It'll respond to both positive and negative prompts.
Compare to Anima base v1, The aesthetics got a light tune-up, so it hits a "usable" level even without artist tags. (Still, use style prompts or artist tags — you'll get better results.)
Can generate very dark image, and the stylization is very strong. Tag weights can also go very high — SDXL caps out around 1.8, but this model takes 2.5, even 3.
Downsides:
Limb stability can be a bit shaky sometimes, especially finger and toe counts. Since it's missing e621's precise vocabulary like 4 toes / 5 toes , the model is basically just guessing. It also loves to add pawpads where you didn't ask for them.
Artist tags seem to have a pretty big impact on overall image stability, and background stability can be pretty rough sometimes, images can come out looking really muddy.
Some e621 concepts are just plain missing — it can't generate certain genital anatomy (canine sheath, genital slit, etc). And use 'testicles' instead of 'balls' (bth 'knot' will become real rope knot with anima xddd).
It can't do a lot of characters, Legoshi being one example.
Sometimes using 'anthro' without 'furry' tag will pull a full "Detroit: Become Human" on you — the character just turns human. So I'd recommend just throwing 'furry' into your positive prompt everytime.
And sometimes the aesthetic quality drops so low it makes me question my entire life. It definitely wasn't this bad when I was merging, so it's probably something I just didn't catch in testing :(
TL;DR: not having e621 data really does hit hard.
How to use:
You'll need WebUI Forge Neo: https://github.com/Haoming02/sd-webui-forge-classic/tree/neo or ComfyUI (I'm using portable): https://docs.comfy.org/installation/comfyui_portable_windows . Some of the preview images were generated with Forge Neo and some with ComfyUI, so the embedded metadata format won't be consistent across all of them.
You need to download the VAE, the CLIP/text encoder, and the diffusion model itself — all three, no skipping.
Sampler: Euler a or ER SDE. Keep steps under 30 (20–50 works, but honestly I don't think going higher does anything better). CFG 3–6 (I used 4 the whole time while testing).
Anima is great at generating high-res images natively — up to 1536px straight out of the box. I'd recommend going for 1k, no hires fix needed.
For artist tags, use the @xxx format, not by xxx . And don't mix more than two artist tags — any more and you'll start getting messy compositions or straight-up weird artifacts.
It feels 'really' different from SDXL to work with. Give yourself a bit of time to adjust anima based model.
It uses both the generic quality words like best quality, very aesthetic and a Pony-style score system. Don't forget to throw those quality tags in.
Artist tags: https://thetacursed.github.io/Anima-Style-Explorer/ .
qwen_image_vae.safetensors
safetensors · 242 MB · SHA-256 a70580f0213e…3d1f · Civitai
