DucHaiten-NoobAI

If you like my content, support me at:
https://ko-fi.com/duchaiten

Patreon is here (but there's nothing in it):
https://linktr.ee/Duc_Haiten

I created my Discord server. I'm still quite new to Discord server management, feel free to share your opinions.
https://discord.gg/vKEW6jqa49

PLS Read this article to learn how to use the Snapvu model.

Update version 1.4.37:

1. Adjusted lighting and slightly reduced contrast to make the realistic style look more natural (high contrast suits 2D and 3D styles better, but in realistic style it can make the image look harsh, overly bright or dark, and lacking natural depth and volume).

2. Tweaked the influence level of training data so that the model doesn't overfit to the training set. This helps it better understand keywords already present in the base model's prior knowledge. It may slightly reduce the strength of trained keywords, but not to a degree that causes major issues.

3. Improved fine details — minor elements are now rendered more accurately. For example, fingers and toes now look better, along with other small features.

4. When using the 2.5D style, adding the keyword "style anime" can enhance aesthetic quality.
   Example: "2.5d, style anime".

Although I was really excited when I heard people say that V-pred is much better than EPS, and how powerful NoobAI is, I spent an entire week researching how to train it in the best possible way. I tried training both EPS and V-pred using the same base model, which was the V-pred version of NoobAI, and I suddenly realized that training it with EPS actually gave much better results than continuing with V-pred. This discovery was quite disappointing. I came to realize that what actually made it perform better was the base model itself, not the fact that it was trained with V-pred. In reality, training with V-pred wasn’t as effective as training with EPS. With the same base model and the same dataset, anyone can compare them, and just by looking at the results, it’s easy to see which one is better.

This model was not only created for research purposes but also to supplement NoobAI with additional 2.5D, 3D, and realistic data. The dataset used for training the 2.5D style is the same as Pony-No-Score, while the 3D and realistic styles come from two separate datasets that were split from Pony Real.

The activation keywords are very simple:

• 2.5d is 2.5d (you can add "style anime" for more emphasis)

• 3d is 3d (you can add "unreal engine" for more emphasis)

• realistic is realistic (you can add "photography" for more emphasis)

Note: When the face is far away, you should use hires fix, but for very close faces, hires fix will make the skin too smooth, losing its realism.

Negative prompt:

(worst quality:1.5), (low quality:1.5), (normal quality:1.5), lowres, bad anatomy, bad hands, multiple eyebrow, cropped, extra limb, missing limbs, deformed hands, long neck, long body, signature, username, artist name, conjoined fingers, deformed fingers, ugly eyes, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, painting by bad-artist

Model EPS

Model V-pred