Improved installation and inference from python function (rather than CLI)

[rb-synth]

• Moved code inside src, so that pip install -e . can be used (by updating pyproject.toml).
• Moved the inference script inside the package so that it can be imported from elsewhere, e.g., from ltx_video.inference import infer.
• give def infer signature the same defaults as argparse so that it can be used as a function from elsewhere.
• models_dir argument exposed (so can be set to None and use hf default)
• move configs inside src/ltx_video so that it will be included with pip install. Path defaults to relative to inference.py.

Local install:

uv venv --python 3.12
uv sync --all-extras --inexact --python 3.12 --index-strategy unsafe-best-match

Or add as dependency to another package (in e.g., requirements.txt):

>>> requirements.txt
git+https://github.com/rb-synth/LTX-Video.git@refactor#egg=ltx_video[inference-script]

And run with:

from ltx_video.inference import infer
infer(prompt="something")

[ybitterman]

Thanks for contributing to the package!
General comment:
Please use commit messages that match the regex pattern: .: [A-Z].

[ybitterman]

Please squash commit bd56660 into the first commit of the PR

[ybitterman]

The default should be ltxv-13b-0.9.7-distilled.yaml

[ybitterman]

Please fix the Readme to point to the new location of inference.py