|
--- |
|
license: mit |
|
tags: |
|
- not-for-all-audiences |
|
--- |
|
|
|
A finetune of lodestones/fluffyrock-1088-minsnr-zsnr-vpred-ema-pytorch with the following changes: |
|
|
|
- Caption dropout of 10% to improve classifier-free guidance |
|
- Timesteps weighted by 1/(snr+1) to improve output quality |
|
- A higher virtual batch size of 64 |
|
- Includes RedRocket/furception_vae for convenience (licensed separately, please refer to its repository for more information) |
|
|
|
Like its predecessor, this is a v-prediction, zero terminal SNR model. Please use the provided .yaml file (or do whatever your preferred frontend requires you to do to load a model as v-prediction and zero terminal SNR) so the model can load correctly, or you will have a lot of trouble generating correct outputs! |
|
|
|
**NEW Feb 11 2024**: Fluffyrock Unleashed Refiner has been released. Use it as a refiner model for the last 20% of generations for better fine details at no additional compute cost. For best results, use the default noise schedule. |
|
|
|
The VAE decoder is licensed under Furception's license (as of writing CC-BY-NC-SA 4.0 though due to change). All other model components are licensed under the MIT license. |
|
|
|
Special thanks to @RedHotTensors for general assistance. |
