2.5 KiB
2.5 KiB
FABRIC Plugin for Stable Diffusion WebUI
Alpha version of a plugin for automatic1111/stable-diffusion-webui. Expect bugs and rough edges and feel free to contribute if you know how fix them.
📜 Paper: https://arxiv.org/abs/2307.10159
🎨 Project page: https://sd-fabric.github.io
Releases
- [29.08.2023] 🏎️ v0.6.0: Up to 2x faster and 4x less memory usage thanks to Token Merging(tested with 16 feedback images and a batch size of 4), moderate gains for fewer feedback images (10% speedup for 2 images, 30% for 8 images). Make sure to enable the Token Merging option to take advantage of this.
- [22.08.2023] 🗃️ v0.5.0: Adds support for presets; makes generated images using FABRIC more reproducible by loading the correct (previously used) feedback images when using "send to text2img/img2img".
Installation
- Open the "Extensions" tab
- Open the "Install from URL" tab
- Copy-paste
https://github.com/dvruette/sd-webui-fabric.gitinto "URL for extension's git repository" and press "Install" - Switch to the "Installed" tab and press "Apply and restart UI"
- (optional) Since FABRIC is quite VRAM intensive, using
--opt-split-attentionis recommended.
Compatibility Notes
- SDXL is currently not supported (PRs welcome!)
- Compatibility with other plugins is largely untested. If you experience errors with other plugins enabled, please disable all other plugins for the best chance for FABRIC to work. If you can figure out which plugin is incompatible, please open an issue.
- The plugin is INCOMPATIBLE with
referencemode in the ControlNet plugin. Instead of using a reference image, simply add it as a liked image. If you accidentally enable FABRIC andreferencemode at the same time, you will have to restart the WebUI to fix it. - Some attention processors are not supported. In particular,
--opt-sub-quad-attentionand--opt-split-attention-v1are not supported at the moment.
How-to and Examples
Coming soon. Feel free to share examples with us if you have found something that works well and we'll add it here :)
Citation
@misc{vonrutte2023fabric,
title={FABRIC: Personalizing Diffusion Models with Iterative Feedback},
author={Dimitri von Rütte and Elisabetta Fedele and Jonathan Thomm and Lukas Wolf},
year={2023},
eprint={2307.10159},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
