Hugging Face published a post on September 10 showcasing Workflow1111, a project built with Gradio Workflow that reconstructs AUTOMATIC1111's feature set as a graph of 11 media pipelines and 73 nodes. The authors are Yuvraj Sharma and Abubakar Abid.
For readers who never used it, AUTOMATIC1111 was the most widely used local web UI of the Stable Diffusion era, and it more or less defined how people worked with AI image generation for a couple of years: positive and negative prompts, samplers, hires fix, image-to-image, ControlNet extensions. The community's center of gravity later shifted toward the node-based ComfyUI, but A1111's feature list has remained the de facto baseline ever since.
What Got Ported Over
The features the post lists as rebuilt include text-to-image with style presets, hires fix (upscaling plus an extra denoising pass), image-to-image editing, prompt rewriting with a large model, image captioning, mask generation for inpainting via object detection, prompt-matrix batch generation, upscaling and background removal, Canny and line-art ControlNet-style annotators, PNG metadata read/write, and image-to-video.
The models called behind the scenes include FLUX.1-Kontext, Qwen2.5-VL, DETR, a ViT classifier, BRIA RMBG-2.0, and Wan 2.2 I2V.
Gradio Workflow itself defines four types of operators: fn is a plain Python function, model calls a model through InferenceClient, space connects directly to another Gradio Space, and dataset pulls rows from a dataset on the Hub. Of these four, only fn runs entirely on your own machine — the other three all depend on compute elsewhere.
The Split With ComfyUI
The comparisons the post draws out are fairly concrete: a node can be hardware you don't own; every output automatically becomes a typed REST endpoint; OAuth is supported, so logging in through the browser is enough; different modalities can sit on the same canvas; and a custom node is just a Python function, so it can do anything Python can do.
The minimum amount of code to get started is three lines — import gradio, write a function, and call gr.Workflow(bind=[your_function]).launch(). Deployment goes through gradio deploy, which pushes to Spaces. The whole workflow also supports MCP, meaning an agent can treat it as a callable tool.
On performance, the post gives only one figure: once images are preloaded, each CPU-based annotator takes about 0.5 seconds. The license isn't stated in the post.
What It Means for Users in China
How usable this actually is depends on which side of the connection you're on. The 22 purely local nodes mean tasks like background removal, Canny detection, line-art extraction, and metadata handling can run entirely offline, unaffected by network conditions. But the model and space operators route through Hugging Face's inference endpoints, and direct connectivity to those from China has long been unreliable — especially noticeable when calling heavier models like FLUX.1-Kontext and Wan 2.2.
Wan 2.2 is a video generation model open-sourced by Alibaba, and it has a closer distribution channel in China through ModelScope. In other words, most of the models in this graph do have equivalent hosting available domestically, but the Workflow runtime is currently tied to the Hub — switching sources means modifying the fn nodes yourself to call local or domestic services instead. That's not impossible, just extra engineering work.
Looked at as a step in tool evolution, the move from A1111 to ComfyUI was a shift from forms to graphs; Workflow1111 is a shift from "the graph runs on my GPU" to "the graph can run on anyone's GPU." Once that layer of abstraction holds, local VRAM stops being a precondition for whether something can run at all and becomes a cost option instead. Whether that holds up will show in the ratio of self-hosted to rented nodes the community settles into over the next year or two.
Sources: Hugging Face official blog, Gradio documentation, CocoLoop; node counts, operator categories, and timing figures were checked against the original post, and the assessment of availability in China reflects editorial judgment.