- Fixed gallery app adding new images to the last page instead of the first page when refreshing (7bf3b04)
- Added output commit ID to metadata if GitPython is installed (19eeb30)
- Added cutout augmentations to VQGAN-CLIP (b3a7ab1) and guided diffusion (9651bc1)
- Fixed random seed not saved to output if unspecified in guided diffusion (8972a5f)
- Added feature to enable generating scrolling/zooming images to VQGAN-CLIP (tnwei#11)
The Python environment defined in environment.yml has been updated to enable generating scrolling/zooming images. Although we only need to add opencv, conda's package resolution required updating Pytorch as well to find a compatible version of opencv.
Therefore existing users need to do either of the following:
# Remove the current Python env and recreate
conda env remove -n vqgan-clip-app
conda env create -f environment.yml
# Directly update the Python environment in-place
conda activate vqgan-clip-app
conda env update -f environment.yml --pruneSince starting this repo in Jul 2021 as a personal project, I believe the codebase is now sufficiently feature-complete and stable to call it a 1.0 release.
- Added support for CLIP guided diffusion (tnwei#8)
- Improvements to gallery viewer: added pagination (tnwei#3), improved webpage responsiveness (tnwei#9)
- Minor tweaks to VQGAN-CLIP app, added options to control MSE regularization (tnwei#4) and TV loss (https://github.com/tnwei/vqgan-clip-app/5)
- Reorganized documentation in README.md and
docs/
Update to 1.0 by running git pull from your local copy of this repo. No breaking changes are expected, run results from older versions of the codebase should still show up in the gallery viewer.
However, some new packages are needed to support CLIP guided diffusion. You can follow these steps below instead of setting up the Python environment from scratch:
- In the repo directory, run
git clone https://github.com/crowsonkb/guided-diffusion pip install ./guided-diffusionpip install lpips- Download diffusion model checkpoints in
download-diffusion-weights.sh
VQGAN-CLIP app and basic gallery viewer implemented.