Banano
/

banchan-anything-v3-0

+# Banano Chan - Anything V3 Guide
+## What is it ?
+[Banano Chan - Anything v3.0](https://huggingface.co/pbuyle/banchan-anything-v3-0) is a [Stable Diffusion](https://en.wikipedia.org/wiki/Stable_Diffusion) model (aka. checkpoint) infused with potassium and trained to the likeness of [Banano Chan](https://twitter.com/Banano_Chan/). It can be used to generate an infinite amount of pictures of the waifu embodiment of [Banano](https://www.banano.cc/), a feeless and super fast meme cryptocurrency.
+## Stable Diffusion Web UI
+### Google Colab
+The easiest but limited way to run the model is to use [TheLastBen's Google Colab Notebook](https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb). A Colab Notebook is some fancy way to run commands and code on Google's GPUs. This one contains all what is needed to run [Stable Diffusion web UI](https://github.com/AUTOMATIC1111/stable-diffusion-webui), a rich and powerfull browser interface based for Stable Diffusion. With this, there is nothing to download or install, you can get started right away in less than 10 minutes. The only downside is that it will only work for a few hours before you run out of _free credits_ on Colab. Afterwhich you will habe to wait a day or two to be able to use it again.
+Simply open [the notebook](https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast_stable_diffusion_AUTOMATIC1111.ipynb) and follow the instructions, running the steps in order with the ▶️ button.
+Once you get to the **Model Download/Load** section, select 1.5 as **Model_Version** and set the **Link_CKPT** to https://huggingface.co/pbuyle/banchan-anything-v3-0/resolve/main/banchan-anything-v3.0.ckpt.
+![model-download-load.png](model-download-load.png)
+Once the model is downloaded, the last step is to actually run the web UI. Wait a bit and you should see a link which looks like **https://\<random characters\>.gradio.app/**. That's the address of your web UI, open it and jump to the next section of this guide (ie. skip the _Running Locally_ part).
+### Running Locally
+If you are running Windows or Linux and have a supported GPU (video card), you can run [Stable Diffusion web UI](https://github.com/AUTOMATIC1111/stable-diffusion-webui) locally on your computer. Most NVidia GPU with 8GB of RAM or more should be supported. AMD support is harder, you will need to run Linux and have a GPU supported by ROCm 5.x. Other GPU (eg. Intel) aren't supporter. Head to [the install instuctions](https://github.com/AUTOMATIC1111/stable-diffusion-webui#installation-and-running) if you want to give it a try and come bacj once you got it running.
+**TODO**: Instruction to download the model and place it at the right place.
+## Your first Banano Chan pictures
+Now that you have the web UI running, open it in your browser. The interface is quite complex. But to generate your first picture, simply enter `banchan` in the first text box you see (the _Prompt_ one) and smash the _Generate_ button on the right.
+![screenshot-1.png](screenshot-1.png)
+Hopefully, after a couple of seconds you should see the generated picture right below the button.
+Nice, but the picture isn't that great. The quality is low and it doesn't look very polished.
+## Getting better pictures
+To get better pictures you need to tell Stable Diffusion that it is what you want. You do this by adding more information in the _Prompt_ but also by using the _Negative prompt_ (righ below). Simply using  `banchan, (masterpiece:1.2), (best quality:1.2)` as prompt and `lowres, low quality, normal quality, worst quality, cross eyed` as negative prompt should already provide decent results.
+![screenshot-1.png](screenshot-2.png)
+Other options to get slightly better picture is to enable the _Restore faces_ option and increase the _Sampling steps_. Between 20 and 30 steps is usually fine, but feel free to experiment with higher values.
+## Getting more specific pictures
+The prompt describe what the image should be. The negative prompt described what you don't want in the picture. You can use plain English sentence there. For instance `banchan eating a pizza, (masterpiece:1.2), (best quality:1.2)`.
+![00018-4110522853.png](00018-4110522853.png)
+```
+Prompt: banchan eating a pizza, (masterpiece:1.2), (best quality:1.2)
+Negative prompt: lowres, low quality, normal quality, worst quality, cross eyed
+Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 4110522853, Face restoration: GFPGAN, Size: 512x512
+```
+In addition, you can also use Danbooru tags. For instance `portrait of banchan, (masterpiece:1.3), (best quality:1.3), yellow dress, outside, garden scene`.
+![00014-1826284240.png](00014-1826284240.png)
+```
+Prompt: portrait of banchan, (masterpiece:1.3), (best quality:1.3), yellow dress, outside, garden scene
+Negative prompt: lowres, low quality, normal quality, worst quality, cross eyed
+Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 1826284240, Face restoration: GFPGAN, Size: 512x512,
+```
+When adding more to the prompt it sometimes become necessary to shift stable diffusion attention back to Banano Chan using attention weight, like so `(banchan:1.1)` (you can go as high as `1.3` before its start decreasing picture quality). Similarly, you can use some or all of `(green eyes:0.8), (green hair:0.8), (yellow hair:0.8), (gradient hair:0.3)` (and play with the attention weights to increase or decrease the impact of each tags) to get the model back on track. The model seems to like very long hair, so use  `short hair` and/or `blunt hair` to keep them closer to the original.
+![00013-1826284240.png](00013-1826284240.png)
+```
+Prompt: portrait of (banchan:1.1), (masterpiece:1.3), (best quality:1.3), yellow dress, outside, garden scene
+Negative prompt: lowres, low quality, normal quality, worst quality, cross eyed
+Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 1826284240, Face restoration: GFPGAN, Size: 512x512
+```
+Checkout [Safebooru tags](https://safebooru.org/index.php?page=tags&s=list) to find tags to compose your picture (warning: this site should be mostly SFW but contains suggestive pictures). And remember that you can also use a plain English description mixed with tags.
+## Eww it's lewd
+![00020-3836769494.png](00020-3836769494.png)
+```
+Prompt: portrait of (angry:0.3) (banchan:1.28), (disgusted:1.1), surprised,
+(masterpiece:1.2), (best quality:1.2), (green eyes:0.8), (green
+hair:0.6), (yellow hair:0.8)
+Negative prompt: lowres, low quality, normal quality, worst quality, cross eyed, nsfw, blush
+Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 3836769494, Face restoration: GFPGAN, Size: 512x512,
+```
+The Banano Chan model wasn't build with the purpose of generating NSFW or even suggestive pictures of Banano Chan. However, the base model on which it is build was trained on images from the Danbooru image board. Where lots of NSFW anime pictures are shared. This is what makes the model so good at anime and easy to control with tags. But it also means the model is well versed on lewd and even a bit biased toward it. If you want to avoid suggestive of explicit pictures, adding `safe` to the prompt and `nsfw` to the negative prompt should help.
+## References
+  - [Safebooru tags](https://safebooru.org/index.php?page=tags&s=list), mostly SFW but with suggestive pictures
+  - [Danbooru tags](https://danbooru.donmai.us/tags), NSFW
+  - [6 Steps to get your ANIME GIRL image from NovelAI!](https://lunarmimi.net/freebies/novelai-anime-girl-prompt-guide/), a NovelAI prompt guide which shows how to combine tags to get the waifu you want. You need to replace the `{}` with `()`.
+  - [NovelAI Anime: Prompt Technique Master List](https://aituts.com/novelai-anime-prompt-techniques/).
+---
+This work is licensed under a <a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/">Creative Commons Attribution-ShareAlike 4.0 International License</a>. Share it, update it, make something awesome from it.