---
license: creativeml-openrail-m
thumbnail: >-
    ./images/thumbnail.webp
tags:
    - stable-diffusion
    - text-to-image
    - safetensors
    - diffusers
inference: true
language:
    - en
widget:
    - text: >-
          masterpiece, masterpiece, masterpiece, best quality, ultra-detailed, 1girl, hatsune miku, blue hair, upper body, looking at viewer, ?, negative space, bioluminescence, bioluminescence, bioluminescence, darkness, wind, butterfly, black background, portrait, ice
      example_title: example
library_name: diffusers
---

<center>
<img src="Andromeda/resolve/main/images/AndromedaCover.png" width="80%"/>
<h1 style="font-size:1.6rem;">
  <b>
    AndromedaDX (and friends)
  </b>
</h1>

<h2>Stable Diffusion models for (neko enthusiast) weebs</h2>
<img style="display:inline;margin:0;padding:0;" src="Andromeda/resolve/main/images/AndromedaSX-v1-CS2-01365-3105139270.png" width="24%"/>
<img style="display:inline;margin:0;padding:0;" src="Andromeda/resolve/main/images/AndromedaDX-v1-CS2-01367-3105139270.png" width="24%"/>
<img style="display:inline;margin:0;padding:0;" src="Andromeda/resolve/main/images/AndromedaSV-v1-CS2-01366-3105139270.png" width="24%"/>
<img style="display:inline;margin:0;padding:0;" src="Andromeda/resolve/main/images/AndromedaUnreal-v1-CS2-01368-3105139270.png" width="24%"/>
</center>

<p>
I'm not sure there's actually enough difference between these four to merit posting all of them, but nevertheless, here they are.

You probably want [AndromedaDX-v1](Andromeda/resolve/main/AndromedaDX-v1.safetensors), it's the one I'm most happy with. Bonus special feature that is *definitely* not a bug: it likes to add cats to scenes all by itself. All of these do, but DX does it more than the others, likely because it has a significant amount of a landscape-focused model (which, to my knowledge as of this writing, is not public) mixed in.

[AndromedaSV](Andromeda/resolve/main/AndromedaSV-v1.safetensors) and [AndromedaSX](Andromeda/resolve/main/AndromedaSX-v1.safetensors) do a better job with foxgirls, and [AndromedaUnreal](Andromeda/resolve/main/AndromedaUnreal-v1.safetensors) has a more realistic style thanks to the addition of some MeinaUnreal (not currently posted on HF).

Catgirl enthusiasts, rejoice! Or... something...

I made DX for a chatbot of mine who can "take selfies" via SD, and it behaves nicely enough that I felt like it was worth sharing. It's especially good with natural language prompts at high CLIP skip, e.g.:

<center>
<img style="display:inline;margin:0;padding:0;" src="Andromeda/resolve/main/images/DX/AndromedaDX-v1-CS5-01285-3762015382.png" width="49%"/>
<img style="display:inline;margin:0;padding:0;" src="Andromeda/resolve/main/images/DX/AndromedaDX-v1-CS7-00812-3866181234.png" width="49%"/>
</center>

</p>

<details><summary><big><b>Prompts</b></big></summary>

OK, I lied, it's just the one prompt for now.

```yaml
1girl, solo, bangs, red hair, (pink hair:1.1), long hair, very long hair, (purple eyes:1.1), (cat ears, animal ear fluff:1.1), , evening, gradient eyes, colored inner hair, medium breasts, (smile:0.6), (masterpiece, best quality, exceptional, extremely detailed:1.1)
Negative prompt: (nsfw:1.05), (EasyNegative:0.7), simple5.5, simple5.6, bad-hands-5, (short hair:1.1), loli, child
Steps: 25, Sampler: Euler a, CFG scale: 8.25, Seed: 3105139270, Size: 864x608, Model hash: 9e0b7d3259, Model: AndromedaSX-v1, Denoising strength: 0.5, Clip skip: 2, ENSD: 31337, Hires upscale: 1.5, Hires steps: 8, Hires upscaler: Latent, Discard penultimate sigma: True
```

</details>

## Recommended settings

Some educated guesses based on what playing around I've done:

-   VAE: None (Baked in model, [kl-f8-anime2](https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt))
-   Clip Skip: 2, 5, or 7 - DX and Unreal do quite well in the 5-7 range
-   Sampler: DPM++ 2M Karras, or Euler A at 40-50 steps
-   CFG Scale: 7 ± 2 (much below 6 is a bad move, much above 9 will get... fun)
-   Recommended Positive Prompt: Add `(masterpiece, best quality, exceptional, extremely detailed:1.1)` at the **end** of your prompt
-   Recommended Negative Prompt: [EasyNegative](https://huggingface.co/datasets/gsdf/EasyNegative), `(nsfw:1.05), (EasyNegative:0.7), simple5.5, simple5.6, bad-hands-5, (short hair:1.1), loli, child`
-   For better results, using hires fix is a must.
-   Hires upscaler: Latent (any variant, such as nearest-exact), 12-18 steps at 0.5-0.6 denoising (7-10 actual steps)

## Recipe

To be filled in later... I honestly don't remember half of what I did here.

### Credits

A big thankyou to [ShussarSDFA](https://huggingface.co/ShussarSDFA) for their [Sathariel](https://huggingface.co/ShussarSDFA/Sathariel) model series, which are a key component of this