OpenHiggsfield AI: a self-hosted studio for 32 image and video models

What OpenHiggsfield AI actually replaces

The README frames this as "the free, open-source alternative to Higgsfield AI" and lists the usual complaints: no studio subscription, no vendor lock-in, your own key. The concrete problem is narrower than that pitch. Anyone generating images and video across several providers ends up with several tabs, several settings panels and several download folders. OpenHiggsfield AI collapses that into one composer and one gallery. The model you pick decides whether the output is an image or a video, and the README states the catalog holds 32 models, 8 image and 24 video, including Soul 2, Soul Cinema, Kling 3, Wan, Flux, Ideogram, Recraft, LTX, MiniMax, PixVerse, Grok and Qwen. That is a studio, not a model. It ships no weights and runs no inference. It is for people who already have a platform key and want a single surface over it. The README says the studio is built for "people who work in long sessions, iterating on prompts, inputs and settings", which is an honest description of the audience: prompt-heavy work, not one-off generations.

One generate object, resolved against the catalog

The architecture section is the most useful part of the repository. Each generate is one object: { model, prompt, media, settings }. The UI builds that object and hands it to a server action. The action resolves it against the catalog and maps it to the generation API's own fields, which the README names as image_urls and aspect_ratio. Server actions are the only caller, so the browser never talks to the generation API directly. Submit is POST /{model}; status is GET /requests/{id}/status; auth is the header Authorization: Key <api_key>. The catalog in src/generation/catalog/ is the source of truth. Each model declares its own settings allow-list and media roles, and the studio renders exactly what the model declares rather than a parallel hardcoded list. That single decision explains most of the product: per-model aspect ratio, resolution, duration, output format, audio, batch size and prompt enhancement all come from the entry, and adding a model means adding an entry, not editing the composer. State sits in five small Zustand stores covering the shared image and video prompts, the shared image and video media, settings[modelId] and a small active store.

Installing OpenHiggsfield AI and running a first generation

The README gives a two-command start. The package manager is pnpm, and the dev server runs on port 3000.

pnpm install
pnpm dev            # http://localhost:3000

Open http://localhost:3000 and you get the full-viewport studio, which the README says is the only page. Before anything generates you need two environment variables, copied from .env.example. HF_API_BASE_URL is the generation API origin and is server only. OPEN_HIGGSFIELD_READ_WRITE_TOKEN is a Vercel Blob read-write token, used because uploads go client-direct to Vercel Blob through /api/blob, which issues scoped tokens.

HF_API_BASE_URL=                      # generation API origin, server only
OPEN_HIGGSFIELD_READ_WRITE_TOKEN=     # Vercel Blob read-write token

With the server running, press Add key in the studio and paste your platform key in id:secret form. The README states the key is stored by a server action in an httpOnly cookie, and that a missing key opens the modal rather than failing silently. Pick a model in the searchable picker, type a prompt, and submit with Cmd/Ctrl + Enter. Skeletons open in the grid on submit, the request is polled every 4 seconds until a terminal status with a 10-minute deadline, and each finished result appears on its own clock. If you would rather not install anything, the README points to a hosted version at openhiggsfield.ai where you add the same platform key in the browser.

Where the studio gets in your way

History lives in IndexedDB in the browser, capped at 60 records, with favorites kept past the cap. The README is explicit that result URLs belong to the generation platform, so old history can outlive its CDN lifetime and show gaps. In practice that means the gallery is a working surface, not an archive: if you need durable output, download it. The README's bulk download is sequential and reports any files the CDN refused, which tells you the authors expect failures at scale. Batch is capped at up to 4 results per press, and for models without a native count setting the studio submits once per result, so a batch of 4 is 4 requests against your key. Failed, NSFW and canceled runs are recorded as failed tiles with the reason and a retry, which is better than silence but still means your gallery fills with non-results. Finally, this is a front end for somebody else's generation API. If that API changes a field name or a status value, the mapping in the server action is what breaks, and nothing in the repository suggests a compatibility layer.

OpenHiggsfield AI versus a plain provider dashboard

The obvious alternative is not another open source studio; it is the dashboard each provider already gives you. A provider dashboard is authoritative, always current, and needs no deployment. It also gives you exactly one model family, one settings vocabulary, and one place to look when you want to compare a Flux result against a Kling result. OpenHiggsfield AI's entire value is that comparison happening in one grid with one prompt bar and a Reuse action that restores model, settings and prompt so a run can be re-rendered rather than re-typed. The trade is real: you take on a Next.js 16 App Router deployment, a Vercel Blob token, and a catalog that only knows the models someone has written entries for. If you generate with one model and one provider, the dashboard wins and this project adds a deployment to maintain. If your work is prompt iteration across several models, the shared composer and the four gallery scopes (Image, Video, Assets, Favorites) are the reason to run it.

Maintenance, licence and upgrade cost

The last push to the default branch was on 2026-09-02, and the repository is not archived. The dependency set is small and current: Next.js 16, React 19, Zustand 5, @vercel/blob 2 and @tanstack/react-virtual 3, with TypeScript 5.9. There are no releases in the repository, so upgrades mean tracking main. That is a low surface area to maintain, but it also means there is no changelog to read before you pull. Two upgrade costs stand out. First, the catalog is the source of truth, so a model or API change is a catalog edit plus whatever the server action mapping needs, and you are the one who notices. Second, the repository has no licence file at the top level. The README calls the project free and open source, but without a licence file the terms you actually get are unclear, and that matters if you plan to modify and redistribute it or run it inside a company. Check the licence before you build on it. Nothing here is legal advice.

Editorial conclusion

Adopt OpenHiggsfield AI if you want one self-hosted composer over many image and video models and you already hold a platform key. Skip it if you need a hosted service with an account, billing and support, or if you expect the studio to supply model access on its own. Before you commit, check the catalog entry for the specific model you care about, confirm that the deployment target supports Vercel Blob, and read the licence file, which the repository does not include at the top level.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. wide-trace/open-higgsfield on GitHub