Acly/comfyui-tooling-nodes

What it solves

This project provides a set of custom nodes and API extensions for ComfyUI, specifically designed to make ComfyUI easier to use as a backend for external software tools. It addresses the friction of exchanging images and masks via the filesystem, simplifies the process of regional prompt application, and provides tools for tiling large images to save VRAM.

How it works

The toolkit extends ComfyUI's functionality through several specialized modules:

• Image/Mask Handling: It introduces nodes to load images and masks via Base64 strings embedded in prompts or via a temporary RAM cache, and sends output images back to clients via WebSockets or HTTP GET requests.
• Regional Prompting: It implements attention masking, allowing users to define a background region and multiple specific foreground regions, each with its own prompt, to create more natural compositions.
• Tiling: It provides tools to split images into tiles with configurable padding and blending, process them individually, and then merge them back together seamlessly.
• Miscellaneous Tools: It includes a local translation node (powered by Argos Translate) and an NSFW filter (using the Stable Diffusion safety checker).
• API Extensions: It adds HTTP endpoints for inspecting model metadata (base model, quantization type), translating text, and uploading models directly to the ComfyUI directory.

Who it’s for

Developers building external applications, UIs, or scripts that use ComfyUI as a headless generation engine, as well as ComfyUI users who need advanced control over image regions and memory-efficient tiling.

Highlights

• Direct Image Transfer: Bypasses disk I/O by using Base64, WebSockets, and RAM caching for faster image exchange.
• Attention Masking: Enables precise, natural-looking regional prompting without the harshness of condition masking.
• Flexible Tiling: Decouples tile extraction and merging from the sampling pipeline for maximum workflow flexibility.
• Model Inspection API: Allows external tools to automatically categorize and filter models by base architecture (e.g., SDXL, Flux) and quantization (e.g., GGUF).
• Local Translation: Integrated translation of non-English prompts into English locally.