Noesis XL

Noesis XL is a merge of NoobAI and CyberRealistic. Comparing V1 examples shows that Noesis XL, in addition to retaining the complete Danbooru and e621 datasets from NoobAI, has further improved the anatomical accuracy and realistic curves thanks to CyberRealistic.

The differences between V3, V2, and V1 are as follows:

1. V1: The original foundational model, serving as the base for the Noesis XL series.

2. V2: Built upon V1. Reduces 2.5D stylistic elements and incorporates a more cartoonish aesthetic.

3. V3: Also based on V1. Introduces brighter, more vibrant color palettes and significantly enhanced detail rendering.

V3/V2 > V1

Here’s my favorite usage method:

• Stop at clip layer: -2

• Positive:

  [1girl/1boy/1other,] [safe/nsfw,] [other tags...,] masterpiece, best quality, newest, absurdres, highres

• Negative:

  [nsfw,] worst quality, old, early, low quality, lowres, signature, username, logo, bad hands, mutated hands, mammal, anthro, furry, ambiguous form, feral, semi-anthro

• Steps: 20 (for fast) | 28 (for best)

• CFG: 3 (more cartoon) | 4 (more 2.5D)

• Sampler: Euler

• Scheduler: Normal

• Resolution: 768x1152 (first) | 1152x1728 (upscale) | 1280x1920 (final)

• Denoise: 1 (first) | 0.5 (image-to-image) | 0.25 (upscale) | 0.1 (final)

Explanation for steps, resolution, and denoise:

1. Use ​20 steps, ​768x1152 resolution, and ​1 denoise to quickly generate images that meet expectations.

2. Use ​20 steps, ​768x1152 resolution, and ​0.5 denoise to refine the image to satisfaction.

3. Use ​28 steps, ​1152x1728 resolution, and ​0.25 denoise to upscale the image.

4. Use ​28 steps, ​1280x1920 resolution, and ​0.1 denoise to generate the final image.

Acknowledgments to the following projects:

NoobAI XL

CyberRealistic XL

StableDiffusion XL

StableDiffusion XL VAE FP16