Stabilizer IL/NAI

Latest updates:(4/17/2025): illus01 v1.121.Rolled back to illustrious v0.1. illustrious v1.0 and above were trained with AI images deliberately (maybe 30% of its dataset or even more?). Which caused lots problems. Didn't notice until I read its paper.Lower character style effect. Back to v1.23 level. So characters will have less details effect from this LoRA, comparing with v1.93 and later versions. But should have better compatibility. This is a trade-off.Added ~100 images that have English text (on sign, board, paper...). Which may increase the probability of successful text generation by ... 5%. Or maybe it's just my illusion. But definitely less random japanese characters. (wait... wtf who would use AI to generate text in image.)Other things just same as below (v1.113).For advance users: if you only want to improve details, don't want to change character's style a single bit. Use the "Touching Grass" LoRA. This LoRA has zero human in dataset, only objects and environments. Has much stronger effects. And you need to balance it with other style LoRA.Guys. I can't test everything before dropping a release. Remember to spam the comment section if weird things happen. And most importantly, remember to attach at least some images or settings. An image with metadata is worth than thousands words. Otherwise I have no idea what are you talking about...(4/10/2025): illus11 v1.113.Trained on Illus v1.1.New dataset "Touching Grass" added. ~1K real world high quality photographs. Very diverse and creative. Same as previous, no human in dataset, just objects and environment. Old dataset was replaced.Full natural language captions from LLM.Bring back pixel level natural texture to overfitted models.Better lighting and depth of field effect. (Photography level)Better background structural stability. Less deformed background, like deformed rooms, buildings..........See more in the update log section..........(3/2/2025): You can find the REAL me at TensorArt now. I've reclaimed all my models that were duped and faked by other accounts.StabilizerJust a personal fun coding project. The name "stabilizer" is a bait now. There's no such definition of "stable". The name used to mean "a better and stable default art style for vanilla NoobAI", back to NoobAI was released. For now this is just a LoRA that trained on a diverse dataset. And can improve character's overall details, backgrounds, natural lighting, brightness, contrast.....The advantage of doing this all-in-one thing is that you don't have to stack tons of different LoRAs for each of those aspects. So there is no conflict that might burn or blow up your output. All conflicts in the dataset will be well averaged out to zero in one training. This is how AI works.Cover images might look mediocre or bad, because they are the direct outputs from the vanilla (not finetuned) base model in a1111-sd-webui with default settings, simplest prompts, no pipeline inpaint fixes, even no negative prompt. They demonstrate the effect of the LoRA, not clickbait.Share merges using this LoRA is prohibited. FYI, there are hidden trigger words to print invisible watermark. It works well even if the merge strength is 0.05. I coded the watermark and detector myself. I don't want to use it, but I can.How to useNo trigger words needed.You don't have to set the patch strength for text encoder (comfyui, etc.). This LoRA does not patch it.Recommended -30% of your current CFG scale. This LoRA will increase contrast so you can lower CFG scale for more details.Version prefix:illus01/11 = illustriousXL v0.1/v1.1nbep11 = NoobAI e-pred v1.1 (compatible with v-pred)DatasetTldr: Only normal good looking things. No crazy art style. Not small (2k images, I won't brag it's big, there are many gigachads who like to finetune their models with millions of images).Every image is hand-picked by me. No AI images, we have enough of those, which will cause overfit and wipe out all high frequency details (textures) instantly.Below is a summery of the latest dataset.2D/anime dataset:~1k images.Character-focus. (The characters take up most of the image. I also cropped and rotated many images if necessary.)Clear characters. (no blur effect, etc.)Natural poses. Natural body proportions. (No exaggerated art, chibi, jojo pose, etc.)Real world photographs dataset:~1k images.Contains nature, indoors, animals, buildings...many things, except humans.Better background, lighting, overall effect.Brings back pixel level details, diversity and creativity. ("fix" overfitted models)There is a LoRA (Touching Grass) that was only trained on this dataset.Full range colorsImplemented since nbep11 v0.114.It will automatically balance the things towards "normal and good looking". Think of this as the "one-click photo auto enhance" button in most of photo editing tools.Pros: Big improvements in brightness, contrast and CFG stability. Now you can get high contrast and full color range (almost 0~255) at CFG 4~5 (v-pred is CFG 3~3.5, assuming you are using Euler and Normal Schedule) without oversaturation and deformed image.Cons: Normal == Not suitable for crazy bias effect. For example, you want 95% of the image is pure dark and only 5% is bright, instead of 50/50%.Older versionsNew version == new stuffs and new attempt != better version for you base model.You can check the "Update log" section to find old versions. It's ok to use different versions together just like mixing base models. As long as the sum of strengths does not > 1.Update log(4/17/2025): illus01 v1.121.Rolled back to illustrious v0.1. illustrious v1.0 and newer versions were trained with AI images deliberately (maybe 30% of its dataset or even more? 70%?). Which is not ideal for LoRA training. I didn't notice until I read its paper.Lower character style effect. Back to v1.23 level. Characters will have less details from this LoRA, but should have better compatibility, especially when you are using a strong style base model or stacked tons of LoRAs. This is a trade-off.Other things just same as below (v1.113).(4/10/2025): illus11 v1.113.Update: illustrious v1.0 and newer versions were trained with AI images deliberately. Which is not ideal for LoRA training. I didn't notice until I read its paper. So use this version only if you know your base model is based on Illustrious v1.1. Otherwise, use illus01 v1.121.Trained on Illustrious v1.1.New dataset "Touching Grass" added.Better natural texture, lighting and depth of field effect.Better background structural stability. Less deformed background, like deformed rooms, buildings.Full natural language captions from LLM.(3/30/2025): illus01 v1.93.v1.72 was trained too hard. So I reduced it overall strength. Should have better compatibility.(3/22/2025): nbep11 v0.160.Same stuffs in illus v1.72.(3/15/2025): illus01 v1.72Same new texture and lighting dataset as mentioned in ani40z v0.4 below. Brings RTX level natural lighting and detailed texture to 2D characters.Added a small ~100 images dataset for hand enhancement, focusing on hand(s) with different tasks, like holding a glass or cup or something. Can't say whether it's useful, as popular checkpoints already have similar enhancement. Anyways, the more, the better.Removed all "simple background" images from dataset. -200 images.Switched training tool from kohya to onetrainer. Changed LoRA architecture to DoRA.(3/4/2025) ani40z v0.4Trained on Animagine XL 4.0 ani40zero.Added ~1k dataset focusing on natural dynamic lighting and real world texture. Better RTX natural lighting and detailed texture to 2D characters.Above: Added more real world images. More natural texture and details.Below: Has detail enhancement, but not much.ani04 v0.1Init version for Animagine XL 4.0. In my test it fixed Animagine 4.0 brightness issues and brings much better contrast. Maybe too optimistic. Anyway this is a LoRA so you always can adjust the strength as you want. CFG 5 and LoRA strength 0.5 seems the best for most cases.illus01 v1.23nbep11 v0.138Added some furry/non-human/other images to balance the dataset.nbep11 v0.129bad version, effect is too weak, just ignore itnbep11 v0.114Implemented "Full range colors". See the section above.Added a little bit realistic data. More vivid details, illumination, less flat colors.Above: High contrast. Wider color range.Below: Normal anime contrast and colors.illus01 v1.7nbep11 v0.96More training images. Then finetuned again on a small "wallpaper" dataset (Real wallpapers, the highest quality I could find. ~100 images). More improvements in details (noticeable in skin, hair) and contrast.Above: Has a weak default style.Below: No default style, all data is equally weighted. Pure anime dataset.nbep11 v0.58More images. Change the training parameters as close as to NoobAI base model.illus01 v1.3nbep11 v0.30More images.nbep11 v0.11: Trained on NoobAI epsilon pred v1.1.Improved dataset tags. Improved LoRA structure and weight distribution. Should be more stable and have less impact on image composition.illus01 v1.1Trained on illustriousXL v0.1.nbep10 v0.10Trained on NoobAI epsilon pred v1.0.