papercute / README.md
dorothyisnotarobot's picture
Update README.md
36e6aaf
|
raw
history blame
856 Bytes
metadata
license: bigscience-openrail-m

Papercute

Papercute is an intentionally overfitted fine-tuning checkpoint for Stable Diffusion 1.x that lets you to create cute papercut images using either txt2img or img2img prompts. The Papercute-100k checkpoint was trained using on 141 curated and captioned images over 100,000 steps.

Training took approximately 11hrs on an Nvidia 3090ti.

Tips for best results

  • No keyword is necessary to engage the model. Just load the checkpoint and enter your prompt.
  • Works best at around 50 steps

Known Limitations:

  • Does not work well with PLMS sampling
  • Does not do well with complicated scenes
  • Big cats tend to look like doilies. This is due to a particular training image being overempahasized. However the bug/feature has been popular so I've left it in.

Sample images from Papercute-100k: