Gradio introduces gr.Workflow for building, running, and deploying AI pipelines

Gradio has added gr.Workflow, a visual tool for constructing AI pipelines as typed-node graphs where each node represents an input, operator, or output. Operators can call Hugging Face Inference Providers, other Gradio Spaces, or custom Python functions, and intermediate results are displayed in place during execution.

Users can drag-and-drop to connect nodes, run the pipeline interactively, and inspect outputs without leaving the canvas. The same graph automatically becomes a REST API with endpoints named after each output label, enabling programmatic calls without additional configuration.

Example workflows demonstrate fan-out patterns: one prompt can generate a base image via FLUX, remove its background via a background-removal Space, and produce a title via an LLM—all from a single canvas and exposed as separate REST endpoints.

Pipelines can also fan out to parallel analyses of a Hugging Face dataset using the Datasets Server API, producing overview cards, row previews, per-column statistics, and distribution charts independently and concurrently.

For custom compute, an fn node can run a user’s own GPU model by decorating a Python function with @spaces.GPU; ZeroGPU allocates a GPU for the call, runs the model, and releases it, enabling end-to-end pipelines without relying solely on hosted inference.

Deploying a workflow to Hugging Face Spaces is a one-command process from the same graph, and the REST endpoints are immediately available for integration via the Gradio client or plain HTTP requests.

The workflows can be duplicated from live Spaces, modified, and redeployed, or built from scratch in Python with a minimal API: import gradio as gr; gr.Workflow(bind=[your_function]).launch().