hearmeneigh/e621-rising-v3

NSFW

This model is not suitable for use by minors. The model can and will produce X-rated/NFSW content.

Quickstart

Downloads

⤓ Checkpoint (fp32 | fp16 | bf16)
⤓ Tag Autocomplete CSV
⤓ ComfyUI Workflow

Reference

• Installation instructions
• What's new in v3?
• Prompt examples
• Prompt guide
• Tag list
• Tag autocomplete guide

E621 Rising V3 (SDXL)

• Furry / anthro base model trained with images (mainly) from E621
• Guaranteed NSFW or your money back
• Stable Diffusion XL 1.0 base model:
  • 1024x1024px
  • Trained with 11 epochs of 280,000 images each
  • Finetuned with 23 epochs of 40,000 images each
• Compatible with:
  • 🤗 diffusers
  • stable-diffusion-webui
  • ComfyUI
• Fully open source crawl, dataset, curation, and training process:
  • Use these tools to train your own version with your own dataset!
  • Configuration
  • Toolchain
  • Dataset

Examples

For more examples, continue here.

For more examples, continue here.

Training Procedure

Training legend

• 160 images per batch (epoch variant)
• 1024x1024px image size
• Adam optimizer
  • Beta1 = 0.9
  • Beta2 = 0.999
  • Weight decay = 1e-2
  • Epsilon = 1e-08
• Constant learning rate 4e-6
• fp16 mixed precision
• SNR gamma set to 5.0
• Noise offset set to 0.07
• cosine_with_restarts scheduler
• 11 epochs of V3 curated dataset samples resized to < 1024x1024px (maintain aspect ratio)
• 16 epochs of V3 finetuner dataset samples resized to < 1024x1024px (maintain aspect ratio)
• 6 epochs of V3 finetuner dataset samples resized to < 1024x1024px (maintain aspect ratio, randomly drop 70% of tags)
• 1 epoch of V3 finetuner dataset samples resized to < 1024x1024px (maintain aspect ratio, randomly drop 50% of tags) and learning rate set to 4e-5
• Tags for each sample are shuffled for each epoch