Gradio Releases Workflow1111, Replicating Key A1111 Features

The Gradio team has released Workflow1111, which replicates the core features of AUTOMATIC1111 in a single workflow canvas. This project consists of 73 nodes and 11 media pipelines, supporting text-to-image generation, high-resolution refinement, image editing, prompt matrices, VLM-based inverse prompting, ControlNet style annotators, background removal, image-to-video conversion, and more.

Architecture and Execution

The workflow is composed of 4 operators (fn, model, space, dataset). Of the 36 total operator nodes, 32 are fn nodes, and 22 of them run in-process without network connections. This allows approximately 2/3 of the functionality to remain available even when disconnected. fn nodes can load models on a local GPU or run in a ZeroGPU environment and can be tested directly as Python functions.

Key Feature Implementations

• High-Resolution Refinement: Uses a 2-node approach applying refinement instructions via the FLUX.1-Kontext model node instead of A1111's upscaling.
• Prompt Generation: Converts rough prompts into a list of up to 40 tags using the Qwen3-4B model node.
• Image Inverse Prompting: Runs Qwen2.5-VL and a ViT classifier in parallel to generate prompts and labels simultaneously.
• Inpaint Masks: Detects objects with DETR and processes them locally using Pillow/NumPy.
• Annotators: Canny, line art, etc., are implemented as NumPy-based fn nodes (approx. 0.5 seconds on CPU).

API and MCP Integration

All output nodes automatically generate REST endpoints without manual routing, exposing 9 endpoints such as /image, /edited_image, and /generated_prompt. When mcp_server=True is set, output nodes are exposed as MCP tools that can be called by AI assistants like Claude Code and Cursor. Authentication uses the X-HF-Token header, and the Space does not store tokens.

Comparison with ComfyUI

gr.Workflow covers a similar domain to ComfyUI based on node graphs, but offers flexibility by allowing custom nodes to be written as Python functions. It also automatically generates zero-code REST/MCP endpoints and can run on Hugging Face Space without its own GPU. With OAuth enabled, visitors can use it with their own accounts without installation.