The original framework for this code is based off of Github user bloc97's implementation of Prompt-to-Prompt.
EDICT (Exact Diffusion Inversion via Coupled Transformations) is an algorithm that closely mirrors a typical generative diffusion process but in an invertible fashion. We achieve this by tracking a pair of intermediate representations instead of just one. This exact invertibility enables edits that remain extremely faithful to the original image. Check out our paper for more details and don't hesitate to reach out with questions!
Paste a copy of a suitable HF Auth Token into hf_auth with no new line (to be read by the following code in edict_functions.py)
with open('hf_auth', 'r') as f:
auth_token = f.readlines()[0].strip()
Example file at ./hf_auth
abc123abc123
Run conda env create -f environment.yaml, activate that conda env (conda activate edict). Run jupyter with that conda env active
Check out this notebook for examples of how to use EDICT; including edits on randomly selected in-the-wild imagery (alternative to render in browser).
edict_functions.py has the core functionality of EDICT. my_diffusers is a very slightly changed version of the HF Diffusers repo to work in double precision and avoid floating-point errors in our inversion/reconstruction experiments. EDICT editing also works at lower precisions, but to allow all experiments in a unified setting we present the double-precision floating point version of the code.
If you find our work useful in your research, please cite:
@article{wallace2022edict,
title={EDICT: Exact Diffusion Inversion via Coupled Transformations},
author={Wallace, Bram and Gokul, Akash and Naik, Nikhil},
journal={arXiv preprint arXiv:2211.12446},
year={2022}
}
Our code is BSD-3 licensed. See LICENSE.txt for details.