Stable WarpFusion v0.5.23 Nightly v2 / depth / patchwarp (Patreon)
Content
Notebook: https://www.patreon.com/posts/75724464
Changes:
- add v2-depth support
- add v2 support
- add v1 backwards compatibility
- add model selector
- add depth source option: prev frame or raw frame
- add fix for TIFF bug
- add torch downgrade for colab xformers
- add forward patch-warping option (beta, no consistency support yet)
- add *.mov video export thanks to cerspense#3301
Added support for v2 models: v2-512 and v2-depth
Load up a stable -> define SD + K functions, load model -> model_version
Added model selector, you can now choose between v1, v2-512, v2-depth models. You still need to specify path to checkpoint.
Add depth source option
stable-settings -> depth_source
v2-depth model uses depth map, predicted from input image. You can use previous stylized frame or raw init video frame as the source of this map.
Add forward patch-warping
stable-settings -> depth_source
A test feature. Uses forward flow to move large clusters of pixels, grouped together by motion direction and speed
warp_num_k - number of those clusters per frame
warp_forward - enable\disable this feature
Add *.mov video export thanks to cerspense#3301
Videos can now be exported as quicktime *.mov in nearly lossless quality
Xformers install fix for colab
Important!
As colab env has been updated, the prior xformers build has stopped working.
In order to make it work we need to downgrade pytorch to v1.12.1.
This is done automatically in the notebook, but the env needs to be restarted after that, otherwise xformers will not install correctly. So the Install xformers cell will run, downgrade pytorch and kill the env, so you'll have to click run all to continue.
Notebook: https://www.patreon.com/posts/75724464
Files
Previews only