StableMondAI-SDG
This model is an attempt to stabilize noobaiXL towards better anatomy and a solid builtin style while still keeping it malleable to follow artist tags and art styles. It also better understands the art styles that were present in the SDXL 1.0 base model but got lost in the lineage of style merges. StableMondAI excels at toony styles, photorealism is not a strong suit, use RealMondAI-SDG or 3WolfMondAI-LastSDG for that instead.
Disclaimer
I am not the creator of this model but are allowed to host it here on civitai.
If you have any questions, suggestions or want to leave some praise about this model head over here
Prompt Recommendations
Start with a basic prompt and work from there. This model works better with leaner prompts, us as many tags as necessary and as few as possible.
Prompts need to use danbooru/e621 tags, natural language is not supported. Use quality tags at your own discretion, they have some style influence.
Starter Positive prompt:
masterpiece, best quality, newest,Starter Negative prompt:
worst quality,As with most noobaiXL merges it defaults to anime, so use the following tags for anthro characters
anthro, furry,Generation Recommendations
The following settings are recommended starting points:
Sampler: Euler A
Scheduler: SGM Uniform
Steps: 24-32
CFG: 4-5
RescaleCFG of 0.5-0-7 is recommended to prevent the colors from oversaturating.
This model is in V-PREDICTION mode.
Initial tests show that other samplers work as well but need appropriate settings, especially SDE samplers work only with specific schedulers.
More Information
Check the HF for sampler comparisons and style grids
Description
Initial version
FAQ
Comments (23)
can you please tell me if I have some kind of interference with the generation, and I can't figure out the reason
What kind of issue by chance? Are you using comfy or reforge as a UI?
@Odano441 I'm probably not very knowledgeable, I use AUTOMATIC1111
This is a VPRED (V-Prediction) model and I think only the dev branch of A1111 supports these types of models. A1111 is kinda outdated, if you want to keep the UI look then its recommended to switch to reForge UI
Does this have illustrious 1.0?
No, its based in noobaiXL, which is based on IllustriousXL 0.1. There is no real noobaiXL model based on IllustriousXL 1.0 (yet),
I guess i should finally write a (much too long again) review for this, because it certainly deserves it. This model so far is great, and my now mainly used one already.
This is for me, what i hoped i could get Noob v-pred to look like. It seems to have most, if not almost all, improvements of Noob, compared to Illustrious merges. Like the expanded training (especially furry), less bias (things like overpowering style focus, human girls, etc.), better prompt understanding (even with natural language, despite this and many other models screaming at you that they cant do it. Somehow, they can understand even longer sentences for me), and thanks to v-pred the lighting ofc, which is fantastic and can do pure black and darkness (for the most part) like the base Noob v-pred can, but also daytime, much easier and not burned as it often looks in Noob for me.
The colors in general look much more natural but still strong by default, then in the base Noob v-pred for me (and ive tried probably all the settings there i could have by now). BUT it can still give the extreme contrast and bright colors of Noob, if you want it. The overall visual quality is also much cleaner and closer to a good Illustrious finetune, but with most of the good stuff Noob can offer.
All in all, it has a few things people would still pick Noob over, which is mainly because of the (probably) trained in loras. I did some tests, and while it can get a perfectly black background, the character seems to always get a glowing outline with it, no matter what ive tried. Im pretty sure that for instance is from a lora (if intended or not), and i understand that some people would see this as having less control over what exactly you want to get.
For me personally tho, that is a tradeoff i can live with. The trained loras and maybe other resources, are probably the main reason why the visual quality of this is so much more consistent and clean then base Noob imo. Im not at a point yet where i need, or could reach, complete 100% control over what i want a model to do. For those people Noob might still be preferred. But i love that this model gives me a much easier to achieve cleaner quality, and the prompt understanding and knowledge of Noob, some concepts, scenes and styles it does in fact grasp much better then even IL tunes.
As most times, at least for me it works best with its recommended settings, but has some freedom of trying. Ive settled on Euler Ancestral - SGM Uniform, 32 steps, cfg 5, 896x1152, and so far didnt have to fiddle around with it anymore (which i did anyways, just to return to this every time). While ive tried it with cfg rescale 0.7, as i HAD to use in Noob v-pred before, as the model claims, i didnt need it. However you can still use it, it will mute the colors quite a bit, which depending on what you want can still work better then just lowering the cfg itself.
I do recommend hires fix, while this model can look very nice already at base, especially in this model hires fix seems to work absolutely amazing, and makes images SO much cleaner and detailed, its crazy. I personally use it most times at the same step count as my main generation, but if that takes too long for you, half the steps work just fine too, sometimes even better depending on style. I never use more then 0.2 denoise, at 2x upscale, but if you upscale at lower, then you can prob get away with slightly higher denoise.
What is the specialty of V-PREDICTION? It doesn't work in Comfy / Stable as I understand it. And it doesn't work with Lora. So what is its advantage?
@Twostep Comfy and (re)Forge support it just fine if you use the current versions. Illustrious/noobaiXL EPS based Loras also work fine if they are well made.
@Xeno443 (re)Forge is what? Is it a new version? I use Forge and Comfy.
And what is the trick or difference between V-PREDICTION and others?
@Twostep Forge current version also works automatically with vpred. I cant explain the mathematics, its a slightly different method that the image is generated which for example allows for more color contrast and rage. Honestly if you use a current UI it doesnt really matter from the user perspective, you can use it like any EPS checkpoint.
@Xeno443 It's very similar to the FurryToonMix and the Nova Furry XL. What should I look out for? What are some possibilities to play with colors?
I've even found a couple of guides and read what V are, but I haven't figured out what changes.
Could you import the model to Tensor? Pweaasee 🥲
done
This model is truly like a hidden gem; it would be great if everyone used it a lot.
one word for this model amazing
I'm trying to set this up in Krita. Does anyone know what the resolution the model was trained on was? There's a setting for that in Krita and I think setting it higher than the default 1024 is doing something helpful, but I'm uncertain.
It is based on sdxl so the default resolutions apply, though the noobAI finetunes often work with slightly higher megapixel count, e.g. up to 1384x1384 generates pretty coherent as well. I only use the sdxl resolutions for basegens to be sure and get to the desired resolution via (careful) upscaling.
@Xeno443 Thanks. I'll keep experimenting!
Any plans on making an update to this model? It's really great, it would be a pity if it's left behind
No new base models/finetunes have been released, so there is really no "being left behind". All the "new" models that appear on civit daily are just re-merges upon re-merges of existing stuff in the search for pleasant built-in styles, but 99% of them still only share the common noobai-XL knowledge, nothing new (concepts/chars/artists) is being added. StableMondAI was made to not have a very strong built-in style but to be very flexible and malleable with artist tags and loras, so as long as there is no noobai-XL 2.0, there is not really any (easy/cheap) way to update it with more knowledge.
if I wanted to train a lora on this model what would I use?
Use as in which training software? You can train a lora directly on a merge, but usually its recommended to train on one of the base (finetune) models instead, so either IllustriousXL v0.1 or noobAI-XL
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.