Skip to content

v0.5.0 dev 005
Compare
Choose a tag to compare
@talosh talosh released this 20 Nov 15:16
· 996 commits to dev since this release
v0.5.0_dev_005
0b8bac1

Newer release for Linux, MacOS versions will be coming soon.

  • Added brightness compression so values above 1 and below 0 are preserved.
  • Added new faster model that has been trained for rational timewarps and does not need iterations to reach ratios different from 0.5 This faster model tend to perform a bit worse in terms of a quality / artefacts though. Flow resolution currently does not have an effect on that mode.
  • Added Slower model that uses RAFT model to pre-calculate initial flow. This might help in certain scenarious but it currenlyt might also introduce artifacts on the edges of an image.

To install copy the whole folder into flame python path. Linux version should contain numpy and pytorch 12.1 bult for cuda 10.2 for both flame 2023 (python3.9) and flame 2024 (python 3.10). Those packages are located in hidden folder - flameTimewarpML/packages/.lib/python{version}/site-packages This is done for Flame not to scan those packages and attempt to import it on startup.

For some newer graphics cards you might need newer version of pytorch. For example pytorch 12.1 does not work with RTX 3060 card.
In order to install it just for TimewarpML without altering main intstallation one can do virtual env with /opt/Autodesk/python/{version}/bin/python -m venv myenv
and then install it to myenv folder with ./myenv/bin/ python -m pip install torch==1.13.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116

the copy the contents of myenv/lib/{version}/site-packages to corresponding folder in flameTimewarpML/packages/lib/

Assets 3