First of all, thank you to those who always make good models.
Do not share this model on all other sites.
Especially not on fucking [PixAI].
I didn't want to write something like this here, but I have no choice. I create and share models based on other people's work. Not that I haven't put in the effort, but it's not entirely my own creation, so I never intend to profit from this model. I share it here in the hope that any individual user can freely use it.
However, I believe that sharing a model itself is very different from providing an image generation service—especially when the latter is used to make money as the main business.
I don't know how many of you will read this, but please be aware: [PixAI] is stealing models and LoRA from CivitAI without permission. In fact, they even, disgustingly, officially support a feature that steals models.
I confirmed that my models were being used without my permission and requested their deletion, but the process was handled incredibly slowly, taking a full week. I asked them to take care to ensure that such incidents wouldn't happen again, yet after I uploaded this v10 version, it was stolen in just three days.
If any of you are also creating and sharing models or LoRA, please check to see if those petty thieves have stolen your valuable work.
[v12]
It’s been a while—we meet again at version 13. As the clock has come full circle, I took extra time to present something even better.
I merged for slightly more aesthetic outputs while preserving each art style.
Until now, the focus was on consistency of style and strong impact (contrast or saturation). This time, I’ve aimed for softer, more natural results and tried to emphasize detail. There are similarities to previous versions, but many things have changed.
If you find that my model is uploaded on other sites (especially PixAI), please contact me at any time. It wasn’t uploaded by me—it was stolen without my permission.
This description was translated using ai and may not be perfect.
[Intro]
Hi there.
This model was created based on NoobAI, with the UNet from other SDXL models merged and some LoRA training added.
Ultimately, my goal was to maintain the core performance of model while achieving a slight quality improvement. However, as an average person without special techniques or high-end hardware, I aimed for “good” rather than “perfect.” As a result, some outcomes may lean in a particular direction, but I believe they’re still enjoyable.
The improvements I focused on are:
Slightly better hands
More consistent quality
More natural compositions
Balanced, non-overwhelming color tones
Aesthetic adjustments, and more.
Recommended Settings
This model is a v-pred model. It is available in the
A1111 webui dev version
reforge
comfyui
Parameters:
CFG: 4~7
Steps: 25~32
Sampling & Schedule:
Generally, the combination of [Euler a] and [SGM Uniform] is a safe choice.
Personally, the combination of [Euler Negative] and [Align Your Steps] provides more diverse results and effective noise handling, allowing clean results even in standard resolutions.
(Other settings are generally similar to those for NoobAI.)
License
This model adheres to the following [license] and is available for non-commercial use only. Commercial use is prohibited.
Thank you.
Description
This version is quite different from the previous one, so calling it just a 'version upgrade' might not fully do it justice. Nevertheless, I hope you give it a chance.
I’ve been aware that my model tends to produce brighter results. Drawing inspiration from the composition formula of the previous v3.0 version, I identified problematic unet layers and made adjustments. This led to surprisingly effective improvements, but I kept questioning whether it was enough.
So, I made a significant change to the values I had been adjusting conservatively until now. As you may have noticed from the sample images, 'darkness' is now more pronounced. The contrast has increased, though not to an excessive degree.
Of course, there may occasionally be unintended 'gray tones' or 'vignetting,' but I believe this is an acceptable trade-off.
Now, cleaner images are generated even at standard resolutions. For hires, the results are even better.
The chances of generating images that differ from previous versions have increased. While I’ve tried to maintain consistency, I ask for your understanding regarding the difficulties in achieving perfect control.
FAQ
Comments (11)
6.0 is insanely good, it’s basically perfect!
I haven't had the chance to try 6.0 yet, but this is my favorite Illustrious model. It captures the aesthetic of the original with added control that the main checkpoint seemed to be lacking.
Great stuff.
I found that this model easily generates six fingers and six toes, and it does so very accurately—almost 80% of the time, there are six
If you are using prompts related to hands, lowering the weight might be a potential solution.
@NaGaRi I only used prompts like "holding," and just drew over a dozen images. Almost 100% of them have six fingers, mostly after high-resolution restoration when they turned into six fingers
@1263496472532 Same here, I usually fix it by redrawing in Photoshop 😅
@CCCino That’s so frustrating😥
@1263496472532 😅😅😅
I know roughly which part will be modified, so I'll reflect it in the next version.
but Don't expect too much. I'm not sure if it'll work well
If you are using an artist combination, you may be able to line up if you place it in front of you in the order of many learning data.
@NaGaRi That’s great
@NaGaRi When it starts with 4 toes, it becomes 5 after restoration. If it starts with 5, it turns into 6 or even more after restoration. It always splits the toes during restoration. If I generate a 1024×1024 image directly, it usually creates 6 toes or more right away😣