Workflow1111 Rebuilds AUTOMATIC1111 in Gradio With 73 Nodes
Workflow1111 rebuilds most AUTOMATIC1111 features in Gradio with 11 media pipelines, 73 nodes, nine REST endpoints, MCP tools and flexible GPU support.
Summary
Published September 10, 2026, Workflow1111 rebuilds most AUTOMATIC1111 features as 11 media pipelines on a 73-node Gradio Workflow canvas. Four operator types, fn, model, space and dataset, combine text-to-image controls, FLUX.1-Kontext refinement and editing, Qwen3-4B prompt expansion capped at 40 tags, Qwen2.5-VL interrogation, ViT classification, DETR detection-to-inpaint masks, prompt matrices, local Lanczos and AuraSR ×4 upscaling, BRIA RMBG-2.0 background removal, five NumPy ControlNet-style annotators, PNG metadata and Wan 2.2 I2V A14B animation. In demonstrations, ViT labels a night-market image restaurant 51.9%, tobacco shop 15.6% and toyshop 9.1%, while DETR finds six objects, three people, a dog, bicycle and car. Parallel branches run both interrogators together and generate four prompt-matrix images simultaneously. Of 36 operator nodes, 32 are fn nodes, 22 run in-process, roughly two-thirds survive disconnection, and five annotators take about 0.5 seconds each on CPU.
Users sign in with Hugging Face or provide a token, with model calls consuming their quota. Remote models run through Inference Providers or Spaces, while fn nodes can load local checkpoints. FastVideo/fastvideo-fasth3-preview demonstrates FastH3, a four-step MiniMax-H3 distillation producing video with sound on ZeroGPU, which allocates and releases a GPU per call. Nine generated REST endpoints expose images, edits, prompts, detection, grids, upscaling, annotators and PNG data. Setting mcp_server=True converts outputs into MCP tools for Claude Code, Cursor and other clients, using caller-owned X-HF-Token credentials that the Space does not retain. Developers can duplicate the Space, bind Python functions, rewire nodes and publish with gradio deploy.
Positives
- Eleven media pipelines place image generation, editing, analysis, upscaling, background removal, metadata and video animation on one browser-accessible canvas.
- Twenty-two of 36 operator nodes run in-process, leaving roughly two-thirds of Workflow1111 functional without an internet connection.
- Parallel execution runs Qwen2.5-VL and ViT interrogation together and starts all four prompt-matrix generations simultaneously.
- Nine output nodes automatically become typed REST endpoints and can also operate as MCP tools without hand-written routing code.
- ZeroGPU and locally loaded checkpoints let developers run GPU models without changing how gr.Workflow calls fn nodes.
- Five NumPy annotators process the example facade in about 0.5 seconds each on CPU without model calls.
Risks & concerns
- Hugging Face sign-in or an access token is required, and remote model calls consume each user’s own quota.
- gr.Workflow lacks a loop operator, forcing four separate text-to-image nodes for the four prompt-matrix variants.
- Network-backed model and Space nodes stop working without connectivity, although roughly two-thirds of the canvas remains available.
- DETR detection leaves the local machine, even though subsequent box drawing and mask creation stay local with Pillow and NumPy.
- Running Workflow1111 without a local GPU depends on external Inference Providers, Spaces or ZeroGPU capacity.