Anima version is available!
Nova Furry XL
A 2d/2.5d furry checkpoint model that can have great details on any type of furs, scales and feathers, which aims to be XL-fied Peaki Furry
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 for best performance
Steps: 20-50 steps
(Pony)
CFG Scale: 6-7
Prompt: score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_furry, BREAK, furry, anthro, detailed face eyes and fur, {Animal Type}
Negative Prompt: blurry, jpeg artifacts, username, watermark, signature, normal quality, worst quality, large head, low quality, text, error, missing fingers, extra digits, fewer digits, bad eye
(Illustrious)
CFG Scale: 3-5
Prompt: masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed, absurdres, newest, scenery, furry, anthro, {Prompt}, BREAK, depth of field, detailed fluffy fur, volumetric lighting
Negative Prompt: human, multiple tails, modern, recent, old, oldest, graphic, cartoon, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra digits, fewer digits, cropped, very displeasing, (worst quality, bad quality:1.2), bad anatomy, sketch, jpeg artifacts, signature, watermark, username, simple background, conjoined, bad ai-generated
Description
Support with e621/rule34 tags as well (probably) while improving details and posing
NoobAI EPS v1.1 + Illustrious v2.0 stable DARE applied
FAQ
Comments (59)
That was pretty quick :0
Whats new on v11? Thanks!
Support with some e621 tags and rule34 tags is the biggest part
@Crody Amazing!!! Tag adherence is always welcome, thanks ❤️
@Crody Hey just wanted to give you a heads up now that you started using the R34 dataset.
Be very careful when merging it.
The best way to use it is by finding a good block wieght to merge it into the set. but DONT ever set any block higher than 0.2.
Use 0.01 ~ 0.2.
Btw, some users might find issues with DPM++2 Karras when you use this lora.
using anything outside of Euler A might create the same affect as turning the CFG up to like 100.
@LatteLeopard That is very true
First I did with the strength of 0.4
At first, I satisfied with the results but as I continued creation, I noticed particles and noises appearing everywhere
So I just adjust the value and got this one
@Crody Oh yeah, its very potent. V11 look and feels fantastic by the way.
You know with that R34 dataset, the creator claims he has no idea how it works or what it even does.
But after months of using I realized what it does is re-enforce overlapping booru tags. It also adds some of R34's more verbose tags which e6/booru doesnt use.
@LatteLeopard Do you think its better to use R34 tags instead e621? been using the last all this time with pretty amazing results
@SXCD2719 no worries, just prompt like you normally would with e621 tags or booru tags. They should be better at understanding these tags better than his previous releases.
What Crody and I were discussing relates to fine-tuning how the text encoder and model weights interact with the addition of R34 data.
@LatteLeopard alright! thanks ❤
Hey @Crody is there not gonna be anymore A and B versions of the model? Mainly wondering cuz of how strong the 2.5D /3D style can be in versions like V5, V6, V7, 8A, 9A etc. Those versions tend to fight certain 2D style loras. Just wondering if V10 and beyond can do both 2D/Flat color stuff without needing a separate model anymore.
Thanks.
V10 and beyond probably can do A's style as well but it's not strong
I'd say if the model is either not good at volumetric or flat, I'll separate them
@Crody Ah gotcha. thanks for the clarification.
V11 seems a lot noisier than V10 tbh.
It's because of rule34 datasets
Some danbooru tags and rule34 tags conflicts thus the noise keeps happening
Try adding 'particles' to the negative prompt
Does this work with all SDXL base LoRAs?
Illustrious and NoobAI: yes
Pony and SDXL 1.0: weak/noisy but yes (not recommend)
From my experience with V11, there are definitely fewer mistakes and the results feel more consistent, which is really great. The background quality in particular has improved a lot, even with minimal prompting. However, compared to V10, I noticed some additional noise — almost like the image is painted with brush strokes.
I tried adding (particles:1.7) to the negative prompt as suggested in the comments, and that seemed to help a bit. Overall, the improvements are significant, but unfortunately I still can’t reach the same character quality I was able to get with V10 using my usual settings. Maybe I’m being a bit too picky, but I’ve also noticed that I’m not the only one experiencing this issue.
I sincerely hope that in V12 everything will be fixed and it will become the most groundbreaking version among all previous models. In any case, wishing you the best of luck — and may even “picky” users like me end up leaving only positive comments!
Amazing model!
When I run locally, I can't past the noise generation. The final result is just a noisy image. How to fix this? Hope you can help me
Can I get the setting?
CFG, Steps, sampler(scheduler) and prompts
Are you using samplers which add noise during steps? If no - use them. This model works only with this kind of samplers.
Some of them:
1. Euler_a OR (cfg_++) variant
2. DPM++ 2m/3m SDE
3. res_multistep_ancestral OR (cfg_++) variant
For ComfyUI:
If you use just the sampler (without cfg_++ variant of it), then set cfg to ~4.
If you use cfg_++ variant of sampler - then set cfg from 1.0 to 2.0
@ee_di_tor I use Euler-a for the generation
25 steps / CFG = 5 / Full VAE
This is my favorite model I've tried so far. But I am having an issue which seems to be with any model I try. For some reason the more zoomed out the character is, the worse the quality is. Most areas aren't so bad, but the worst is facial expressions, particularly teeth. Mouths get blurred and distorted.
I generate locally, and have tried LoRAs, VAEs, upscalers, etc. Nothing fixes these issues. Close-up results are good, but for a full portrait, I think the results for me are not great, for whatever reason. I wonder what the problem is?
thats a limitation of the ai. just use inpaint only masked on faces
Constant noise in this model. Do not have these issues elsewhere.
Am I the only one who thinks the contrast in version 11 is too high compared to version 10?
It was unintentional thing
The contrast issue will be resolved on the next version
Why am I unable to find this model in TensorArt?
I generate locally (Automatic1111)
what sampling and scheduling should I use to get a reduced noise image?
For me, read this article
https://civitai.com/articles/19107/crodys-illustrious-noobai-image-generation-tips
@Crody Euler A doesnt work on my pc, I have an AMD gpu, could that be the reason?
@Le_Angry_Dragon Euler A works on AMD gpu as well
Can I get the settings? (CFG, Steps, Image size)
curious to see what settings you use. I use EulerA, with resolution at 832x1216 with 20 steps and 4 cfg and it produces great results. Though I also use ComfyUI.
@Crody cfg 5, 20 steps, 832x1216
what schedule type do you use?
@r3aper010 basically the same just cfg 5
@Le_Angry_Dragon interesting, while waiting for a reply I tried to use it through Automatic 1111 though it had issues even loading for me xD so i used ForgeUI which is more or less the same and had no issues.
@Le_Angry_Dragon The best scheduling type would be Normal
And from the setting, I'd say it's good to go with those but I use hires.fix with 0.4 denoise, 16 steps at the end
@Le_Angry_Dragon is it by any chance related to the contents in the positive or negative prompts?
For what it's worth I don't do hi res fixing and use Karras scheduler. xD
@r3aper010 thanks for the help, I guess it could be because I am using an AMD gpu, as i had to go threw a nightmare to get it working and sometimes just randomly a model stops working, hell WAInsfw just stopped working for 2 days, and then it just continued working
@Le_Angry_Dragon np, should give ComfyUi a try if you haven't already, looks daunting at first but its amazing and very lightweight compared to automatic1111
@r3aper010 I tried but didnt work, i may try try again later on like i did with auto1111
What are the differences between version 10 and 11? I apologize, but I don't know where to find a changelog with the new features.
Huge noise in the model and hugely oversaturated images all the time ..
EDIT: Use CFG Scale smaller than 3, Euler A + Beta and smomething like 30 steps works pretty well actually
Would like to see a eventual v-pred variant if the base model get workable versions of those, as v-pred really helps on contrast and details.
Brilliant!
good
I like this, but it does not work on draft mode at all. It seems to be pixelated and renders draft images with artifacts. Can you please check this out? It would be really helpful and save a lot of buzz when testing out an image in draft mode first before making the final render.
Peak…
Absolutely fantastic.
Good shit
Love this model!
I've recently started to programmatically generate embeddings for a side project and I might have found an issue. Could someone with more knowledge or experience doublecheck me?
At least as far back as v8, I'm finding that the CLIP-L encoder is possibly corrupted, with six Layer 11 weights (Q, K, V, Out projections + MLP, FC1, FC2) showing NaN values. I looked at Nova Animal XL for comparison, and that model doesn't show the same issue. My understanding is that this situation basically kills CLIP-L output, which impacts overall output quality. Can anyone confirm? Happy to be corrected.
I love this model!
I usually get good results 10/10 😎
Hello there! New here, but is there a way to put back the previous model (10.0) for us to use? While the new one is still good, I could do better ones with the older model.
I love this model, it's definitely the best.
Details
Files
Available On (3 platforms)
Same model published on other platforms. May have additional downloads or version variants.