Hugging Face rebuilds AUTOMATIC1111’s Stable Diffusion WebUI as a Gradio Workflow with 73 nodes

Hugging Face published a blog post and accompanying Space demonstrating Workflow1111, a Gradio Workflow graph that replicates most features of AUTOMATIC1111’s Stable Diffusion WebUI. The graph consists of eleven media pipelines connected by seventy-three nodes.

Workflow1111 integrates state-of-the-art models for text-to-image, high-resolution fix, image-to-image, prompt-matrix grids, visual-language-model interrogation, detection-to-inpaint masks, ControlNet-style annotators, background removal, PNG metadata handling, and image-to-video. Users can run any pipeline by signing in with a Hugging Face account or providing an access token; model calls consume the user’s own inference quota.

The canvas uses four operator kinds: Python functions, model calls via InferenceClient, calls to other Gradio Spaces, and rows from Hub datasets. For example, the text-to-image pipeline includes a prompt-builder function node that appends style presets and cleans text, a model node that invokes a checkpoint through Inference Providers, and a post-process function that writes generation parameters into PNG metadata.

The high-resolution fix pipeline sends the text-to-image output into a FLUX.1-Kontext model node with a refine instruction, returning a sharper, larger image. The image-to-image pipeline reuses the same Kontext node: users upload an image and describe changes, and the node returns the edited image. A separate pipeline uses a Qwen3-4B model node to expand a rough prompt into a structured list of tags, which can then feed any diffusion model node.

A visual-language model pipeline reads an image and generates a prompt that could have produced it, while a ViT classifier node labels the same image in parallel. A detection-to-inpaint pipeline uses DETR to find objects and branches into two local branches: one draws detection boxes on the original image and another creates an inpaint mask. The prompt-matrix pipeline combines a base prompt with multiple suffixes, sending each variant to its own text-to-image node; since the nodes share the same dependency depth, they run in parallel.

Upscaling and background removal are implemented as both local function nodes (e.g., Lanczos resampling) and Space nodes that call dedicated Hub Spaces (e.g., AuraSR ×4 and BRIA RMBG-2.0). ControlNet-style annotators like Canny and sketch are implemented as local NumPy function nodes that run in about half a second on CPU. Roughly two-thirds of the canvas can continue running without a network connection.

The PNG Info pipeline reads back generation metadata stored in the image’s PNG chunk, and the image-to-video pipeline animates the same image using a Wan 2.2 I2V A14B node. The authors note that a single reference node can feed multiple downstream pipelines, so one upload can both read metadata and generate a video on the same canvas.