Any Custom Frontend with Gradio's Backend

Hugging Face has been pushing Gradio past its comfort zone lately, and the latest move is arguably the biggest one yet: gradio.Server. The pitch is simple. Keep Gradio's backend machinery—queuing, concurrency control, ZeroGPU allocation, gradio_client compatibility—but throw out the constraint that your UI has to be built from Gradio components.

To prove the point, the team built Text Behind Image, a browser tool that removes the background from a photo using the BiRefNet segmentation model, then lets you slide stylized text between the subject and the background, complete with 3D extrusion, perspective tilt, drop shadows, and 25-plus fonts. That's the kind of layered, drag-and-drop interface that simply can't be expressed with Gradio's built-in sliders and blocks. So instead of forcing it, they wrote a roughly 1,300-line vanilla HTML/CSS/JS frontend with zero React, zero build step, and paired it with a backend of about 50 lines of Python.

The trick sits in gradio.Server itself, which extends FastAPI rather than replacing it. You get normal FastAPI routes for serving static files—an @app.get("/") handler just returns the index.html—alongside a new @app.api() decorator that wraps any function in Gradio's queuing engine. That matters because GPU endpoints without concurrency control fall over the moment two people hit them simultaneously. With @app.api(), requests queue up properly, @spaces.GPU handles ZeroGPU allocation automatically, and the same endpoint becomes callable from gradio_client, so other scripts or apps can hit your Space's API without touching the frontend at all.

On the client side, the frontend doesn't use a plain fetch() call to reach the backend. It goes through the Gradio JS client, connecting to the Space's own origin and calling the /remove_background endpoint directly. That one choice is what keeps everything inside Gradio's queue rather than bypassing it, which means GPU requests don't collide even under load, and there's room to surface things like queue position down the line.

Hugging Face frames this as removing a false choice: previously you either took Gradio's infrastructure and its component-based UI, or you left for a custom frontend and built your own backend plumbing from scratch. gradio.Server says you don't have to pick. The live demo, ysharma/text-behind-image, is up on Spaces now, and the team has already teased follow-ups covering MCP tool registration, SSE streaming, and multi-page apps with shared state.