ways here:

• use https://github.com/dmarx/Multi-Modal-Comparators
• implement a few simple clip variants here (mclip, chinese clip, japanese clip)
• expose a way to plug any clip-looking function here (and use it in python ?) ?
• something similar than mmc but packaged

mmc is not usable, see dmarx/Multi-Modal-Comparators#41

so I'm thinking to take this plan

1. define a simple generic interface
2. implement a few interesting clips with optional deps
3. when happy with result, move to dedicated package

I haven't touched MMC in a while and had recently been thinking about revisiting it. I think I have a solution that satisfies both my original goal and your packaging needs, but I'd also be open to starting with a fresh solution and use lessons learned. Anyway, here's the idea I had with mmc:

• trim the package down to a bare installation, maybe shipping with open-clip or clip-anytorch
• provide scripts and instructions users can use to install all models, groups of models, or individual models as needed. this could be as simple as renaming the current pyproject.toml so it isn't auto-detected by build tools and just add an additional build step to rename this file (after which the current install instructions would still apply)

Need to double check that it's ready for merge, but the changes staged in the ezmode PR adopts the openai/CLIP interface, including returning preprocessor and attaching tokenizer to clip.tokenize

Another MMC-esque generic interface to consider: https://github.com/salesforce/LAVIS

#50 is a first part in this

mostly done