How to Use FROM scratch Docker Images as an Asset Library for Large ML Assets such as Models, Corpora

MJT + Stable Diffusion 2022

I’m starting to deal with large models and other ML-related assets, and need a solution for how to handle this inside the context of Docker development.

One might be tempted towards volumes. Unfortunately volumes are not portable between machines.

One can simply bundle the assets directly into the image via as simple COPY. That works fine up to a point.

The problem is that one will likely want to be efficient, and take as much advantage of Docker Image Layers as possible.

Docker Images as Asset Library

Docker multi-stage building to the rescue.

The trick is that one can use tagged images as a predefined ‘stage’ of the build

To create an ‘asset image’ first create a Dockerfile for the image

FROM scratch
COPY <assets> /

Replace <assets> with the assets one wishes to preserve

Next build the image

$ build -t 'registry_host/asset/foo:0.1' .

I’d suggest sticking with a naming convention that will make it obvious what it is being used for. I use /asset you do you.

Effectively the image tag becomes the URI of the asset, which makes it easy both computationally and cognitively to deal with.

Importantly this tag is universal and can be pulled/pushed and referenced across machines using the existing Docker registry/repository infrastructure.

Use Assets in Images

COPY --from=registry_host/asset/foo:0.1 / /assets/

For bonus points one can also use the — link command to COPY if/when available.


Using tagged images in combination with multi-stage builds, leads to quicker builds and deployments and more efficient use of disk space.

Originally published at



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store