VLM System Query Raw Presets

This is the quiet sibling of UC_VLMSysQueryAddPresets, and it's easy to skim past. Both are preset libraries for vision-language model captioning, but they answer different questions. The "Add" version takes your text and wraps it in a preset's prefix and suffix. This one, Raw, takes just a preset and hands you the instruction block itself - no text input, no wrapper syntax, nothing added. One combo box in, one string out.

"Raw" is the operative word. In this pack's VLM stack, query strings normally travel inside a request envelope - the advanced instruction node wraps things in {system_query} = { ... } braces before they reach the model. This node skips all of that and outputs the preset content verbatim. That's exactly what you want when your downstream node supplies its own envelope, or when you're building the string yourself with a concat node and don't want duplicate framing around the middle.

The preset list overlaps heavily with the Add variant - character_transfer_gemma, character_transfer_qwen, text2image, image2image, video_basic, ideogram_4, and the h3_* MiniMax H3 family. The same vocabulary warning applies: the Gemma presets speak in [image 1] / [image 2], the Qwen ones in <Picture 1> / <Picture 2>, and which one you pick has to match the picture-token convention of the model you're actually calling. The preset is a block of English instruction either way - the syntax inside it is the part that has to line up.

Inputs and output

• preset (combo) - the only input. Pick your instruction block.
• system_query (STRING) - the single output: the raw preset text, ready to feed into a VLM node's system-query slot or to concatenate with your own prompt.

It's a trivial mechanism under the hood - a dictionary lookup returning one string - which is exactly why it's nice to have as a node. It keeps a stable, searchable source of VLM guidance in your graph instead of burying a 500-word instruction in a text box where it can't be shared or versioned.

Installing it

Part of ComfyUI-UtilsCollection. Install via ComfyUI Manager (search ComfyUI-UtilsCollection), or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection

Restart ComfyUI and you're done - the pack's deps are just opencv-python and typing-extensions, and none of these text nodes need model downloads.

When to reach for this over the Add variant

If you're composing your own query string and just want the preset content spliced in, use Raw. If you want your own words injected between a preset's prefix and suffix, use UC_VLMSysQueryAddPresets. Both are worth having on the same graph - the Raw output feeds a concat node, the Add output goes straight to a VLM call. And if you loaded an old workflow that references VLMSysQueryAddPresets but not this node, that's expected: this node is the newer canonical ID, while the legacy family lives on under "(Legacy)" names for backward compatibility.