OpenHiggsfield is a self-hosted image and video studio for 38 models, and it has 3,329 stars in 25 days
Image: GitHub
Why it mattersA team that already pays a per-model generation API can now put its own studio in front of it, with the composer, gallery and per-model settings picking up automatically from a catalogue rather than a hardcoded list.
OpenHiggsfield, released as wide-trace/open-higgsfield on 26 August, is a Next.js 16 studio that composes prompts, runs generations and holds finished results in one gallery for 38 image and video models. The project describes itself as "the free, open-source alternative to Higgsfield AI" and had 3,329 stars and 742 forks 25 days after release, with the last commit on 17 September. A hosted version runs at openhiggsfield.ai for anyone who wants to try it without installing.
One composer for 38 models
A single prompt bar drives image and video work: the model chosen decides which one, and Command-Enter or Control-Enter submits. The catalogue holds 8 image models and 30 video models, including Soul 2, Soul Cinema, Seedance 2.5 (Edit and Extend), Seedance 2.0 (Fast and Mini), five Kling 3 variants, Wan, Flux, Ideogram, Recraft, LTX, MiniMax, PixVerse, Grok and Qwen.
Each model declares its own allow-list for aspect ratio, resolution, duration, output format, audio, batch size and prompt enhancement, and the studio renders exactly what the model declares rather than a parallel hardcoded list of settings. Media inputs are typed by role: start frame, end frame, references, video and audio, each with the per-role file cap the model declares. Files upload directly to Vercel Blob through a token endpoint and become public URLs the generate request can carry.
The gallery is where the work lives
Runs open into the grid on submit as skeleton tiles at the model's true aspect ratio, and the request is polled every 4 seconds until a terminal status or a 10-minute deadline. Every finished result blooms into place on its own clock. The gallery has four scopes as an arrow-key tab rail: Image, Video, Assets and Favorites. Selection mode with shift-click supports bulk download, bulk favourite and bulk delete. Deletion is reversible for 6 seconds through a bar with a draining hairline. History persists in IndexedDB in the browser at 60 records, and favourites stay out of that cap.
The studio uses the reader's own platform key, entered in a modal and stored server-side in an httpOnly cookie. Server actions are the only caller for the generation API; the browser never talks to it directly.
What the traction says, and where the licence goes
3,329 stars and 742 forks in 25 days is real activity, and 6 open issues against that many stars says a small maintainer team is fielding a lot of installs. There is a catch on adoption: the repository lists no licence, which is the GitHub default of "all rights reserved" until an OSI licence is added. The LICENSE field on the API is null and the README does not name one. For a project marketing itself as open source that is a gap a team should confirm with the maintainer before running it in production.
The value proposition on the page is that the studio itself is free and works with any platform key the reader already has, so the team's cost is only the per-generation cost of whichever provider serves the models. That is a specific alternative to buying a subscription that bundles the studio and the generation together, and the studio is the part that ships with the source.
Source
Source: OpenHiggsfield AI
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.