hearmeneigh/sd21-e621-rising-v2

Outdated!

This model has been superseded by:

E621 Rising V3 (SDXL)

Warning: This model is NOT suitable for use by minors. The model can/will generate X-rated/NFSW content.

Ckpt | Safetensors | Yaml | Tags | Example Prompts | WebUI How-To

E621 Rising V2: A Stable Diffusion 2.1 Model [epoch 29]

• Guaranteed NSFW or your money back
• Fine-tuned from Stable Diffusion v2-1-base
• Training continued from E621 Rising V1
• 10 additional epochs of 250,000 images each, collected from E621
• Trained with 6,246 tags
• 512x512px
• Compatible with 🤗 diffusers
• Compatible with stable-diffusion-webui
• Compatible with anything that accepts .ckpt and .yaml files

Getting Started

• Stable Diffusion WebUI How-To
• Python How-To
• Prompt Guide

Examples

More examples and prompts here

Versions

Precision	CKPT	Safetensors	YAML	Notes
FP16	Download	Download	Download	Use this by default
FP32	Download	Download	Download
BF16	Download	Download	Download

Changes From E621

See a complete list of tags here.

• Symbols have been prefixed with symbol:, e.g. symbol:<3
• All categories except general have been prefixed with the category name, e.g. copyright:somename. The categories are:
  • artist
  • copyright
  • character
  • species
  • invalid
  • meta
  • lore
• Tag names are all lowercase and only contain a-z, 0-9, /, and _ letters
• : is used to separate the category name from the tag

Additional Tags

• Image rating
  • rating:explicit
  • rating:questionable
  • rating:safe

Training Procedure

Training legend

• 204-272 images per batch (epoch variant)
• 512x512px image size
• Adam optimizer
  • Beta1 = 0.9
  • Beta2 = 0.999
  • Weight decay = 1e-2
  • Epsilon = 1e-08
• Constant learning rate 4e-6
• bf16 mixed precision
• 8 epochs of V1 dataset samples stretched to 512x512px (ignore aspect ratio)
• 9 epochs of V1 dataset samples resized to 512xH or Wx512px with center crop (maintain aspect ratio)
• 2 epochs of V1 dataset samples resized to < 512x512px (maintain aspect ratio)
• 10 epochs of V2 dataset samples resized to < 512x512px (maintain aspect ratio)
• Tags for each sample are shuffled for each epoch, starting from epoch 16