Stabilizer
Name is misleading. This is NOT a LoRA that can magically fix your overfitted merged base model.
This is a finetuned model from pretrained model. But trained as a LoRA.
11k images. No bias, no default style.
Zero smooth plastic glossy AI image in dataset. Glossy Al images are polluting the world, but not on my watch, I handpicked every single image.
Natural language captions from Gemini, rather than tags in random order with high FPR.
Standard noise scheduler. Inpainting/img2img friendly.
Improvements:
You can get the style exactly as it should be (as long as the model knows).
Comparisons with pretrained base model:
https://civarchive.com/images/84145167 (general styles, you get the style exactly as it should be)
https://civarchive.com/images/84256995 (artist styles, no style shifting because this model does not have bias)
See more xy plots in cover images.
Why LoRA?
This LoRA is a DoRA (from Nvidia), which is more efficient than traditional LoRA.
Enough for thousands of training images.
What? You prefer to download and store a 7GiB checkpoint than a 80Mib LoRA?
Share merges using this model is prohibited. FYI, there are hidden trigger words to print invisible watermark. I coded the watermark and detector myself. I don't want to use it, but I can.
This model only published on Civitiai and TensorArt. If you see "me" and this sentence in other platforms, all those are fake and the platform you are using is a pirate platform.
How to use
It is highly recommended that you use pretrained base model.
And load this LoRA with strength 1.
Versions:
cknb (ChenkinNoob-XL).
nbvp10 (NoobAI v-pred v1.0). FYI: you don't need CFG hacks (RescaleCFG etc.).
nbep10 (NoobAI eps v1.0).
illus01 (Illustrious v0.1).
Load this LoRA first
This LoRA uses a new arch called DoRA from Nvidia, more efficient than traditional LoRA. But the patch weight is dynamically calculated based on the currently loaded base model weights (which will be changed when you loading other LoRAs). To avoid unexpected changes, load this LoRA first.
Specify styles in prompt
This model does not have an strong default style and is very creative. You must specify the style you want in the prompt.
If you want to use it on finetuned/merged base models:
I personally disagree this. This is not the model's original intention. But this is a LoRA after all.
Be aware:
This LoRA can't remove glossy shiny plastic AI style. Although the dataset is "AI image free". It can't make AI image polluted 1girl overfitted 50 versions of Nova furry 3D anime WAI or whatever look better. What the model learned is learned. Use a pretrained base model if you want to get rid of overfitted AI style.
What is "overfitted AI style"? This is what Craft Lawrence (from spice and wolf) should be, if you've seen the anime: img. This is what those AI style polluted 1girl overfitted model generated: img
Some base models already merged this model. If you got deformed images with this LoRA even at low strength (e.g. <0.5). Your base model has already merged this LoRA (and you merged it twice). And the model weights got multiplied (because how the DoRA works) and collapsed.
Beware of fake base model creators, aka. thieves. Some "creators" never do the training, they only grab other people's models, merge them, wipe all metadata and credits, and sell it as their own base model.
Update log
Moved here
Description
FAQ
Comments (25)
so the tldr is that we slap this lora first and the rest of our prompt looks good af?
yes, simply put this LoRA at the first,
So when you change the combination of rest LoRAs, the patch weight from this LoRA won't change.
Note: Most popular "illustrious" anime base models are based on (or closest to) NoobAI, not illustrious, even if the author claims it is a "illustrious" model. And you should use NoobAI LoRAs (all LoRAs, not just this LoRA) for best effect.
Noobai has a strict license, so many authors deliberately conceal the fact that their models are based on NoobAI. Many authors (mainly for merged models) are unaware that source models they used are from NoobAI. This is their problem, and I won't comment on it.
To verify whether the model you are using is NoobAI (or mainly): You can calculate the model's cosine similarity.
There is a plugin (supermerger) for a1111 webui that has this function.
Focusing on the similarities of big deep layers (IN04~08, M00, OUT00~05).
For ref: similarity between IL0.1 / NAI is ~0.96.
The base number of similarity is big because all those models are trained from SDXL, they shared the same starting point.
If the similarity between your base model and NAI is higher than ILv0.1, then it is closer to NAI, and vise versa. Same as all other base models, e.g, you want to test the similarity between your base model and ILv2.0.
OMG🙉🙉我的模型是NOOB，我一直以为是ILL
我更换成NOOBAI lora，才发现原来这个lora的效果这么好😭😭😭之前一直觉得有点效果但不明显
daaaamn! truthnuke
i would guess they are mixing sdxl/pony/illus/noob
and if model works fine with any of these then logically the model with the most influence over the overall structure of the model would be what the they would label it (you hope)
In general, the licensing of Noob AI is questionable because Illustrious is licensed under the 'Fair-AI-Public-License' and Noob AI is based on Illustrious. That being said, Noob AI inherits the licence. As far as I know, you can make slight changes that comply with the official licence, but you can't simply change 'We do not restrict any upload or spread of the generation results' to 'Users must comply with sharing work details, such as synthesis formulas, prompts and workflows'.
Just tested the "illustrious" merge i've been using and the watermark showed up clear as day. The style wasn't really the same but I'm guessing this means it's secretly using NAI. What I dont understand is how the NAI guys can create a merge using IL which has a fair use license then apply a restrictive license on their derivative product. That's probably not enforceable but the AI world is still the wild wild west.
@GasparGames In short: They can't.
Note: I am not a legal expert and the following is merely my interpretation of those licences.
In long:
The Fair AI Public License 1.0-SD in general is compliant to the 'parent license' CreativeML Open RAIL++-M License. Within the 'parent license' is stated 'at minimum - the same use-based restrictions as the ones in the original license (this license)', 'The restrictions set forth in Attachment A are considered Use-based restrictions.' and it also states that 'The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model.'
Fair AI Public License 1.0-SD clarifies this further: 'The output of this software is not covered by this license, and no contributor claims any rights to it.'
That being said: You are free to do what you want with the outputs you generate, as long as they comply with the 'Prohibited Uses' provisions.
To close this circle: Derivatives have the right to add 'Your own copyright statement [...] respecting paragraph 4.a', BUT you can't just restrict what was passed down differently by the parent license - that's even non-compliant to 'Share Alike'.
@Pietraczlab I asked a friend who is a contract lawyer and she laughed. She says this kind of thing happens all the time in the open source software world. Apparently, anything in a "downstream" license that contradicts the "upstream" or parent license is totally unenforceable because that license itself is in breach of contract to begin with. Her best guess as to why they do this sheds a lot of light on the situation IMO, "If NoobAI was fine-tuned with proprietary or copyrighted data (e.g., specific artist styles or licensed images), the creators might add restrictions to avoid potential legal issues, even if the base model’s license doesn’t require it." This makes a lot of sense when you consider what @reakaakasky says about using gammainks as a litmus test for NoobAI.
@GasparGames It's great that you have someone to ask these kinds of questions. It pretty much underlines what I figured out last night. The only thing that could possibly be valid is the compliant nature of the transition from SDXL licence to the Free AI license. But the artists' styles and training datasets debate is a huge grey area. I only heard about this a few weeks ago. Super interesting!
Would you consider bringing this statement to my recently published article on this topic?
(What the License?! https://civitai.com/articles/18619)
@Pietraczlab Great article. Added my two cents.
To add two cents to the debate... i just tried your test prompt on a clean just downloaded from the website IllustriousXL V2.0 "Stable" and got a clean twitter watermark.
https://civitai.com/posts/21990363
So either Illustrious V2.0 took a shortcut and used NoobAI somewhere, or they have included this artist in the meantime (I was trying this since i have rebased most of my models using IXL V2.0 and was wondering if it could have an influence).
@n_Arno Now that's interesting. It seem the premise here may be incorrect. Thanks for sharing your findings.
@n_Arno Try using the gammainks prompt with the very first Illustrious 0.1. You'll still see the twitter logo and signature most of the time. The artist was already in base Illustrious.
@n_Arno Thanks for sharing this, just tested illustrious v0.1, v1.0, seems it was already trained on e621 (or maybe other dataset contains this artist), but not much. There are watermarks, but 90% watermarks are only text and no twitter logo (often wrong logo). Does not match the one that the author was using. Unlike illustrious v2 and NoobAI you will always get the exact watermark.
So my simple testing has big problem and is misleading. If you got the exact watermark, your base model maybe illustrious v2 or NoobAI.
I thought illustrious was never trained on e621, because I was trying to prompt e621 tags long time ago, and it did not response to those tags at all....
The reliable way is calculating the similarity. There is a "supermerger" plugin in sd-webui has this function, if you want a GUI.
But the "most popular "illustrious" base models are based on NoobAI" still true. I downloaded and tested the similarity of some most popular based models, wai, hassaku, nova..., they all are closest to NoobAI.
I've updated the testing method using similarity.
@reakaakasky I'll test that too anyway with my checkpoint, i am curious of the result
I went ahead and did a more complete test on my end with my model with a tool: https://civitai.com/articles/19225
Interesting. Vpred version, when used on my colorfixed vpred + detailer lora at 1.0 strength becomes what author described - a realism lora. Add some tags like realistic to promote base non anime faces to increase effect. https://civitai.com/images/95661031
Took some getting used to but this is really helpful. It actually helped bring some three-dimensionality to one of my favorite models that is hard to get more 2.5d visuals from
Hi, have you been able to generate consistent images? I've kind of discovered a "sweet spot" for generating through the website since I don't have a powerful machine for it. I've been using 0.5 for some work, but sometimes I notice some noise.
Hello, I've a question, I could use this Illustrious Lora v1.198 on Illustrious v2 just fine or I should just stick with Illustrious v0.1?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.