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
This is v-pred LoRA and is for NoobAI v-pred ONLY, do not use it on EPS models.
Special version, "a" means anime. There is a default 2D anime style. So it should be easier to use if you don't care about style and don't want to prompt style.
Otherwise, for advanced users, v0.271 is still recommended because it doesn't have bias and is more creative.
FAQ
Comments (22)
This is a fantastic thing. I don't know how I lived without it before. It's just a shame that the Illustrious version will no longer be updated. Despite the author's statement that " most (90%) anime base models labeled as 'illustrious' nowadays are actually NoobAI," the illus01 v1.198 version gives me significantly better results with Illustrious models.
oh, I need to update that statement. NoobAI LoRA has worse compatibility.
So if you are using a heavily finetuned model, even if it is NoobAI, illustrious LoRA usually gives better effect.
"no longer be updated"
I've put my full faith (and gpus) into the Lumina 2 model.
@reakaakasky If we take a look at models and loras available here for Lumina vs IL/Noob, specially when it comes to anime, I will say you put the bet on the wrong horse.
Its not always about what models who has the most parameters, but the ones that are most understood and people manage to squeeze the most out of.
@zack39 I tried working with Lumina and Noob and didn’t quite understand their advantages. Objectively, they are several orders of magnitude worse in anime than Illustrious, even taking into account all the “new technological solutions” they include. But that’s the author’s choice. If these models are better suited for his purposes, who are we to demand anything from him? We should be grateful that he even shared his work, and we can have a decent background in Illustrious if we use his LoRA with a strength of ~0.3.
the original (pretrained) model is unusable for most users. If you ever used those models (IL/NOOBAI) without any LoRA when they were released. Fun fact, this is what the pretrained model should be. Have all the knowledge, but no strong styles so maximum finetune-bility.
What most users using nowadays are the finetuned models with different strong default styles. Which usually needs just <100 images to finetune. But needs time to let the best style emerge.
Yes, Lumina 2 has less models. but Lumina 2 released in Jan 2025 and the anime finetuned Neta lumina released just mouths ago. While Illustrious has be released for years and SDXL almost 3 years. You can't just compare model counts between those models...it's unfair. Not to mention Illustrious also had a big advertising event across platforms pushed by a company.
Anime models are trained on the same data sources, danbooru and e621. The only difference is the capability of the base model. Newer, better.
@reakaakasky Fair. I see your point. Guess I need to wait and see what the next one people will widely use :)
Hi. Any chance you could train a stabilizer for Pony v7? Your NoobAI stabilizer brings so many improvements to the model without changing the style, works like a charm.
Probably not. I just don't get the point of pony v7.
Wow
The current popularity of Lumina is still insufficient, and I'm not sure if it's compatible with SDXL's LoRA. If it's not, it would be as troublesome as moving house. For now, I still have to keep my focus on Illustrious and NoobAI.
I've tried Neta Lumina, and honestly, I'd suggest waiting a few more months—because a more powerful, better-architected MODEL will be released around the beginning of next year. This new model not only has a well-prepared dataset but also uses the more powerful Gemma3 4B LLM + Jina CLIP dual TE encoding. It adopts Next-DiT with a larger parameter count than Neta, while retaining the same 16-channel VAE.
and the name of the model is?
@krigeta The model is called Newbie and is currently undergoing large-scale training. Barring any unexpected issues with the training progress, it should be released by the start of 2026—its performance is significantly better than Neta right now.
@xChenNing any sources?
@Noxy_ https://huggingface.co/NewBie-AI/NewBie-image-v0.1-exp-model-repo
I'm just a tester. For more info, please reach out to the official team or check the official repo.
@xChenNing Thanks, been trying to find information online before but nothing came up
What is the suggested strengths for the latest versions?
I am struggling to use one for the Illustrious 1.198:
my usecase is base Illustrious, characters lora and style sketch lora for the whole image? I am not able to get the good stable results please help.
pretrained Illustrious model is not recommended.
Although illus v1.198 was trained on illustrious, it's recommended to use pretrained NoobAI model. illus versions are partial trained for better compatibility with those hyper merged base models.
If you are looking for a pretrained model for a clean start, NoobAI v-pred is recommended. Or NoobAI eps v1.1, if you need to use eps LoRAs.
"SDXL is just too ... old"
Still 100x more versatile than Flux when you know what you're doing.
For my use case Ill 1.23 gives the intended effects. New models change the styles too much. Thanks for this <3.
"Why not try new models. They have newer architecture, better performance, and can be much more efficient. For example: Lumina 2. If you just heard about Lumina 2".
If it cool if the developers of Neta Lumina will made caption model for training and simple colab notebook as with SDXL, but it's not happened and NL will forgetten....
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.