Best Dataiku schema for a human-review dataset of AI-generated product images?

MartynFoster735

Context

Our creative team has started pulling early visual drafts from AI image tools for concepting, including product mockups, storyboard frames, and social graphics. One source is Nano Banana 2 Lite (https://nanobanana2lite.tools/), a third-party independent site not affiliated with Google or DeepMind. It supports prompt-based generation and object-reference workflows. It is useful for fast iteration, but we now need a proper review pipeline before anything moves downstream.

The problem

I want to build a Dataiku dataset that lets human reviewers score and annotate drafts while keeping full lineage back to the prompt and reference images. Fields under consideration:

- image_path: pointer to the stored file (managed folder or external storage)
- prompt_text: full prompt used to generate the image
- reference_set_id: links to object-reference images
- reviewer_score: numeric or categorical scale
- rejection_reason: free text or controlled vocabulary
- version_lineage: parent/child relationship for regenerated or edited drafts

Questions for the community

1. Is a managed folder plus dataset-with-path-column pattern right for images, or is there a better Dataiku storage strategy?
2. For version lineage, has anyone modeled iterative generations with a self-referencing key, or is there a cleaner recipe-based approach?
3. For reviewer scoring, would you recommend a Dataiku app, a webapp, or a plugin-based review flow?

I am not looking for a finished pipeline, just advice on how others structure similar review datasets before committing to a schema. Happy to share what we land on once tested.

Turribeach

1. I would use a managed folder using local storage for fast reading for images that need to be reviewed. Then I would move them to a different managed folder under a cloud storage bucket so you can benefit for cheaper storage rates and PAYG rates rather than provisioning with overhead in local storage.
2. Use a surrogate key and add versioning to it. You could use the file timestamp for versioning.
3. I would use a Dataiku Webapp. If you are clever about it you could expose the images in the managed folder in your webapp using a symlink so that you can display them in the webapp without having to copy them to the static files folder for webapps. See link below:

https://developer.dataiku.com/latest/tutorials/webapps/standard/custom-static-files-kb/index.html#gsc.tab=0