🤖 System Prompt Config

If you caption images with a local vision model more than twice, you've hit the tedium: every analysis node wants a system prompt, and it's the same one every time. The System Prompt Config node is the pack's answer - a library of ready-made personas that outputs both the prompt and an on/off flag, so you can wire consistent behavior across the text and vision nodes without pasting paragraphs into widgets.

What's in the box

One dropdown (preset) with nine options: custom plus eight built-ins, each with a purpose. The useful spread:

• default - the baseline helpful assistant.
• detailed and creative - longer, image-description-flavored variants for richer captions.
• caption - a short, accurate image-captioning persona, ideal when you're feeding captions back into a training pipeline (this is where the KB's lora-training doc's "garbage captions in, garbage model out" rule bites).
• explicit - an uncensored, extremely detailed NSFW describer. It's exactly what it sounds like; the author's own label calls it adult content.
• technical, professional, and accessibility - composition/lighting analysis, neutral professional tone, and alt-text-style descriptions for visually impaired users.

Pick custom and the custom_prompt multiline box activates so you can write your own. The enable_system_prompt toggle lets you keep a preset in place but neuter it without deleting it - handy when you're A/B testing whether a persona actually changes output.

What it outputs and where it wires in

Two outputs: system_prompt (STRING) and enabled (BOOLEAN). Feed the string into a model loader's optional system_prompt input or an analysis node's system_prompt field. The enabled flag is the interesting one - since the pack's analysis nodes take the system prompt as a string, a False there means you should conditionally route an empty string instead, which is more wiring than most people bother with. For 95% of use, you'll just drag from system_prompt to wherever the prompt belongs.

Setup

It's the lightest node in the pack - no llama-cpp, no model files, no extra deps beyond the pack itself. Install the pack once and it's there:

cd ComfyUI/custom_nodes
git clone https://github.com/walke2019/ComfyUI-GGUF-VLM.git
cd ComfyUI-GGUF-VLM
pip install -r requirements.txt

Then restart ComfyUI. One honest caveat: this node is basically a glorified string constant with a nice dropdown. If you're the type who keeps your system prompts in a text file or a shared string node, you may not need it - but for anyone exploring what the various captioning personas do to output, having them one click apart is genuinely handy, and the explicit preset is the fastest uncensored-describer path in the pack.