Nova Furry AM
A 2d/2.5d furry checkpoint model that can have great details on any type of furs, scales and feathers, which aims to be Anima version of Nova Furry XL
Rules
You cannot use the generated images for commercial use if it's not edited (or just turning it to black and white)
Advertising is always welcome
Recommended settings:
Sampler: Euler a
Steps: 20-30
CFG Scale: 4-6
Prompt: masterpiece, best quality, score_9, score_8, score_7, year 2025, newest, highres, absurdres, very aesthetic, scenery, furry, anthro, {Prompt}
Negative Prompts: human, smooth skin, worst quality, low quality, early, old, score_1, score_2, score_3, cartoon, graphic, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, artist name, blurry, jpeg artifacts, lowres, censor
Description
Initial version
The style feels like the middle of Nova Furry XL and Nova Kemono XL
Anima-preview3 based
FAQ
Comments (19)
Good, but seems to have forgotten a lot of art styles
hope we get the realistic one!
do you need the VAE and the decoder?
artist tags have zero effect unfortunately
Artist tags seen to have weak effect, need something high like (@artist:3≤) to have better effect, Indigo Void Furry Anima / Furry Anima have the same issue...
The Anima Furry Domain Test lora seems to create very good results with NovaFurry. I suggest using the lora in combo with this if you have trouble getting a fluffier furrier look.
The problem is Anima doesnt contain the data for furry tags from sites like e621.
Yes, unfortunately that and less responsibility of artist tags are the current issues
I hope this will be resolved in future versions
@Crody I had wanted to do the exact same thing you did, but the problem is that making a LoRA of our models carries over the style and some data, but its not enough.
I have considered resolving this issue. the e621 dataset that contains all data up to 2023 is small in comparision to the danbooru one or the LAION datasets.
The only issue I see with doing this is the following:
-Anima is more natrual language, will QWEN understand tags?
-Anima is trained on about 1mp images/e621 dataset is 4mp
-Will I need to train some of danbooru to get 4mp images all around?
-I am unfimiliar with training on Anima.
-Is Anima actually going to have long term value making it worth this effort
@LatteLeopard I think we need to gather up to this day considering the e621 dataset: if it's possible, if it is I could mash up -2023 dataset and 2023-2026 to create complete one
Since Anima understands some unique tags like @artist tags, I think Qwen doesn't do the tag recognition: it's more like the placeholder/natural language holder like SDXL's CLIP-G is
I think these DiT's blocks are the responsible for it:
llm_adapter >> read the tag
blocks.*.cross_attn >> tag to image
blocks.*.mlp / self_attn >> keep the visual
adaln_modulation_* >> control the effect
What if we add 4 times or more than that of datasets to the checkpoint itself rather than the LoRA?
That way the limitation could be much higher
Unfortunately, I'm also unfamiliar with training as well just like SDXL
There are several people who succeeded on training the checkpoint/LoRA so we could ask them about it
I think it has the worth since the model size is smaller than SDXL, and smarter than SDXL
@Crody That's surprisingly good new to hear. I wasn't really aware and to be honest my understanding of SDXL/Anima architecture is basic, but I know how to train models.
Believe it or not this does not mean tags are useless. It means Anima’s tag behavior is probably more distributed than SDXL CLIP models.
For furry/e621 injection, caption consistency matters even more because you are teaching the adapter/DiT pair a new tag dialect.
Anima was trained on danbooru2023 dataset which does use tags. but they changed up some stuff like you said, to prompt and artist it requires @artist.
We may not even need the full e621 dataset. We need a proper curated set. More data is better only when it is high-signal and matches the behavior you want. And the full dataset contains alot of "low-signal" poor quality data.
Also going back to tags. We can actually teach it tags like this:
rating_explicit
rating_questionable
rating_safe
rising_masterpiece
rising_unpopular
favorites_below_X (25, 50, 100, 250, 500, 1000)
favorites_above_X (250, 500, 1000, 2000, 3000, 4000)
score_below_X (0, 25, 50, 100, 250, 500)
score_above_X (100, 250, 500, 1000, 1500, 2000)
I will take a try at it with a curated set. If it runs wells after a few epochs, I will start adding the datasets after 2024 in roll out releases. If you like I can DM you the DL for feedback. If you like it your free to use it as you like.
@LatteLeopard Keep in mind when there's additional artist tags, it should be @artist kind to contrast with the actual prompt like for example if the artist includes portrait to the name, on SDXL this would likely to create portrait image but on Anima, because we have @ before the artist tag this wouldn't happen
The example tags looks great
We can add something like high favorited, extremely famous or high score to make those additional tags much usable
That's great to hear
I'd like to have some DM for further feedbacks!
@Crody im going to try my best to convert the artist names to @artist and removes the underscores. I will contact via DM as it progresses, wish me luck, lol!
Waiting for 3D Anima fine tune sooo muchh🙏🏻
How do I use this? I get an error
What software did you use?
Does that mean Kemono and Furry are gonna be aborted? And which VAE should I use for it? Many thx
No, Nova Kemono and Furry won't be discontinued
I used Qwen VAE: https://huggingface.co/circlestone-labs/Anima/blob/main/split_files/vae/qwen_image_vae.safetensors
To save anyone interested a bit of time: right now most checkpoint loaders dont work so you need to get the correct VAE and clip. Its not hard, just extra steps. Now onto the review:
AM is the most surgical furry model I have used to date, responding extremely well to descriptive tags or to the ones that I use at the very least, your mileage may vary. Its not creative in the best way possible, if you dont ask for it, it doesnt produce it. That being said, its far from a cinematic model. If you like complexity, fine detail, or texture then you're better off looking at other models. If you're looking for that flatter, more graphic look then this is second to none in terms of clarity, line art, and responsiveness.
It ain't perfect though, some are a result of the tech it runs on and the others are likely training data issues. BF16 is a little harder to run than FP16, you'll definitely feel the difference. I think its worth it, you might not. Colors are more unpredictable, get used to including color temperature to your prompts if you want consistency. Tragically, counter shading remains a sore spot for me. Its certainly better than most other models, I get the proper design 4/10 times instead of 1/20 but its still a gripe.
TL;DR this model is probably the best for furry character work and design. Its responsive and clean at most resolutions, but runs a little slower and falls flat for cinematic work.