This is a proof of concept that:

• takes an input of files in a zip weights
• where those files should end up on disk path_base
• the name of an image to build ontop of base_image
• and where you want to push dest_image

On a CPU, without downloading the base_image, this uses a small go binary built on google/go-containerregistry to create a new image layer, and push that layer / image metadata to dest_image

This uses a prototype go binary built from my r8im tool.

There are "weights" and there are "models". Whenever I get new weights (either through dreambooth, AIT, some new model shared on HF/civit/reddit), I want to be able to use them efficiently.

This builds a new image in around a minute, without any GPU or time spent waiting for cog to build/push.

yeah, this is a prototype, do not use in production

please experiment and share feedback with replicate team in discord