Hey everyone,
I've been working on a ComfyUI custom node that I thought might be useful to share. It's called Random Prompt Builder, and it's basically a way to generate detailed prompts directly in your ComfyUI workflow using local language models.
What it does:
The node takes simple inputs (like "warrior princess" or "cyberpunk detective") and expands them into full, structured prompts. It uses GGUF models running locally via llama-cpp-python, so everything stays on your machine. No API calls or subscriptions needed.
Main features:
- Works with local GGUF models (I've been using Qwen2.5-7B and Mistral-7B)
- GPU acceleration support for faster generation
- Character controls for things like age, body type, ethnicity, etc.
- Structured prompt categories (subject, action, pose, clothing, location, background)
- Multiple output formats (Pony score tags, SDXL tags, Flux natural language, etc.)
- Quality tag injection based on style presets
- BREAK separator support for SDXL/Pony workflows
- Enhancement rounds for iterative prompt refinement
Why I made this:
I originally built a desktop app for this, but I found myself copying and pasting prompts into ComfyUI constantly. This node version integrates directly into the workflow, which makes batch generation and workflow saving much more practical. It also handles the formatting differences between models automatically, so you don't have to remember whether you need score tags, comma-separated tags, or natural language.
Installation:
It's available on GitHub: https://github.com/btitkin/ComfyUI-RandomPromptBuilder
You can clone it into your custom_nodes folder and install llama-cpp-python. The README has detailed instructions for GPU support on NVIDIA, AMD, and Apple Silicon.
You'll also need a GGUF model in your ComfyUI/models/LLM folder. I recommend starting with Qwen2.5-7B-Instruct Q4_K_M (about 4.4GB).
