· Baka-Diffusion ·

· MBW Resources ·

Baka-Diffusion is a series of finetuned / U-Net Block merges made to push the limits of SD1.5 based models. Our models uses the Danbooru Tagging system.

⚠️ Disclaimer

You are responsible for the content you generate, whether it is NSFW or SFW. The AI models do not contain explicit visual content that can be accessed easily.

⚒️ Model synopsis

Due to the nature of U-Net Block merging , some models function differently in CFG Scales. It is generally preferred to use the models in this repository with CFG 3-9. above this range are usable, but you may encounter artifacting, visible leftover noise, or color burn.

• Baka-Diffusion[Base]
  • Baka-DiffusionV4[Base] was created to improve the original V1 model by utilizing block weights. After thorough testing and merging, it became the basis for all the models that came after it.
  • As of now the base model isn’t planned to be released. although it is very flexible it’s lacking in quality.

• Baka-Diffusion[A]
  • Baka-DiffusionV4[A] is still actively WIP! It’s a direct improvement from the Base model with much more detail and quality. Releasing soon!
  • The model is highly compatible with low rank networks such as LoRA / LyCORIS models.

• Baka-Diffusion[S3D]
  • Baka-DiffusionV4[S3D] was created to achieve a soft 3D effect and replicate realistic lighting. It differs greatly from the regular Baka-Diffusion models, where the lighting has a very anime style.
  • This particular model is suited for higher resolutions. Instead of 512x768 , use 600x896 .
  • As of now this is the only model released .
  • The model is highly compatible with low rank networks such as LoRA / LyCORIS models.

To replicate my results and use the model as intended, please use the following default negative: (worst quality, low quality:1.2), lowres, bad anatomy, bad hands, nsfw . This negative served as a base when merging the model. Please avoid negative textual inversions, as they are designed for 2D anime models and using them on the [S3D] model will result in a style offset.

• Baka-DiffusionV4[S]
  • Baka-DiffusionV4[S] is based on the [BASE] model. The [S] stands for Stylization. Currently WIP.
  • The model is not as compatible with low rank networks such as LoRA / LyCORIS models.

• Baka-DiffusionV4[Experimental]
  • We can transfer concepts between two models using ERASING and ADD DIFFERENCE.
  • A+(B-C) merge technique. The recipe is as follows:
  • Erased[A] + (ORIGINAL[B] - ERASED[B]) . This allows us to transfer the chosen concept that was erased from the models. In this case, we will be taking the concept "Vivid color" from the ColorBOX model. Very much WIP not expected to be released but who knows.

? Recommended Settings

• Default negative: (worst quality, low quality:1.2), lowres, bad anatomy, bad hands, nsfw
• CFG 3-9
• Clip skip 2 (Required!)

? Note! The images in this repository are subject to change, and more may be added or removed with each update. Baka-Diffusion is a self-made project that I am solely responsible for, and I only have so much time to spend before experiencing burnout.

? License

This repository is released under the CC BY-NC 2.0 license . The license specifies the following terms:

? Kudos!

• Erasing : https://github.com/rohitgandikota/erasing
• Runtime Block Merge: https://github.com/ashen-sensored/sd-webui-runtime-block-merge
• Super Merger: https://github.com/hako-mikan/sd-webui-supermerger
• KL-F8 Anime2 VAE: https://huggingface.co/hakurei/waifu-diffusion-v1-4/blob/main/vae/kl-f8-anime2.ckpt