Run Open-Higgsfield on Any Backend — LivePair and Beyond
LivePair AI ·
Open-Higgsfield is the community-built open studio — one prompt bar, per-model settings rails, every finished run in one gallery — and it's quietly one of the better generation UIs on GitHub. Its best feature isn't in the README headline: `HF_API_BASE_URL` decouples the frontend from the platform, so the studio runs on any API that answers the same wire format. Here's how to point it at LivePair, and what that actually changes under the hood.
What Open-Higgsfield actually is
A Next.js studio that mirrors the Higgsfield workflow: 30+ model entries (Kling, Seedance, Flux, Soul-style, Qwen, Wan, Ideogram, Recraft and more), each with its own settings rail, all submitting through one API layer. It's BYOK — you bring credentials, it handles the UX. The upstream default is the platform API, but nothing in the code hard-binds it.
That last part matters. `HF_API_BASE_URL` is a first-class env var, which means the studio's real dependency is the wire protocol — `POST /{model-path}` to submit, `GET /requests/{id}/status` to poll, `Authorization: Key id:secret` to authenticate — not any particular company.
Pointing it at LivePair
LivePair's /hf endpoint answers that exact wire format, backed by its own model catalog and per-result billing. Two changes move the whole studio:
| Setting | Default | LivePair |
|---|---|---|
| HF_API_BASE_URL | the platform origin | https://livepairai.com/hf |
| Credential | id:secret platform pair | key:lp_… — your lp_ key fills the secret slot |
What changes under the hood
Model paths translate by surface rather than by name. The studio's video entries (Kling, Seedance, Veo-style paths) route to Wan 3.0 — image-to-video when a source frame rides along, text-to-video otherwise. Image paths with an input image go to Seedream edit; Ideogram keeps its name; Qwen maps to qwen-image-3; everything else image-side lands on Seedream 5.0. LivePair model ids pass straight through, so the catalog isn't a straitjacket.
Parameters carry where they map — prompt, input images, aspect ratio, duration, resolution, audio flags — and Higgsfield-only knobs (elements, soul ids, motion presets, batch size) drop instead of silently corrupting the run. Features with no equivalent — Soul ID training, the elements system, motion presets, webhooks — return a clear 501 with a pointer to the native API.
Why run it this way
Three concrete reasons. Billing: per-result from a prepaid balance starting at $5 — no subscription, credits don't expire. Privacy: LivePair's posture is no public feed by default and renders auto-delete after 48 hours, which matters if the studio is generating client work or personas you don't want scraped. Catalog: the private model line and the native /v1/agent surface (x402 wallet payment, OpenAI-shaped image/video endpoints, MCP) sit behind the same key.
The honest caveat: if your studio sessions depend on Soul characters or motion presets specifically, those features live on the real platform — this setup is for the workflow, the models, and the economics, not a clone of the product surface.
Same trick, other clients
Open-Higgsfield isn't the only client with a base-URL seam. The official SDK family takes `baseURL` / `base_url` overrides in their client configs — the v1/SDK wire (`POST /v1/{task}/{model}`, `GET /v1/job-sets/{id}`, `hf-api-key` + `hf-secret` headers) is the second format /hf answers. Any tool that lets you set the API origin is a candidate; the endpoint, mapping rules, and limits live at https://livepairai.com/docs/hf-compat.