A self-hosted studio that wrangles 32 generative models into one tab

What it does

OpenHiggsfield is a browser-based studio for generating images and videos through third-party APIs. You bring your own platform keys (id:secret), pick from a catalog of 32 models, and submit prompts from a single composer that adapts its settings rail to whatever the chosen model declares. Finished runs land in a masonry gallery with reuse, bulk selection, and a six-second undo on deletion.

The interesting bit

The entire UI is driven by a model catalog that acts as the single source of truth: add a new model entry and the picker, settings, and media-role inputs render themselves automatically. The studio never maintains a parallel hardcoded list, which is the kind of architecture decision that prevents UI rot when the model count doubles again.

Key highlights

• One prompt bar handles both image and video generation; the model determines the output type.
• 32 models in the catalog (8 image, 24 video), each declaring its own allow-list of settings like aspect ratio, duration, and batch size.
• Server actions are the exclusive caller to the generation APIs; the browser never holds your key.
• History and favorites persist locally in IndexedDB (capped at 60 records), while media uploads go to Vercel Blob.
• A live polling lifecycle updates skeleton tiles every four seconds until a ten-minute deadline or a finished result.

Caveats

• You must bring your own platform keys and pay the underlying providers; the studio is only the interface.
• Old gallery entries can outlive their CDN lifetime, leaving gaps in historical runs.

Verdict

Worth a look if you juggle multiple generative APIs and want a self-hosted, unified interface with a thoughtful gallery workflow. Skip it if you are looking for a model to train or an API that abstracts billing; this is a studio, not a backend.