Nodes/ComfyUI-UtilsCollection/VLM System Instruction Presets (Legacy)
ComfyUI Node

VLM System Instruction Presets (Legacy)

A dropdown of ready-made captioning system prompts

By silveroxides·Created 2 months ago·Updated about 7 hours ago· 24
VLM System Instruction Presets (Legacy)
    • system_instruction
    presetaction_system_instruction

    The "(Legacy)" is doing real work here: this is the old node ID for UC_VLMSysInstrPresets, kept registered so workflows from before the UC_ rename keep loading. Functionally it's the simplest node in this corner of the pack - a dropdown, a lookup, and a string out. But the strings it hands you are anything but simple.

    When you're driving a vision-language model to caption images or build prompts, the system prompt is where the quality lives. A flat "describe this image" gets you a flat caption. These presets are multi-hundred-word system instructions that wire up a whole captioning persona: how to identify subjects, how to handle multiple pictures, what vocabulary to use, what to prioritize. Pick neutral_system_instruction, action_system_instruction, photo_system_instruction, toon_system_instruction, or one of the video variants (video_basic_*, video_timeline_*, and the MiniMax H3 timeline instructions) and you get a finished, battle-tested prompt you'd never hand-type.

    Worth being straight about what this family is: these presets belong to the "uncensored captioning" school. They explicitly instruct the model to describe content without euphemism or refusal, which is why people pair them with abliterated Qwen3-VL builds for building training datasets. If that's not what you're doing, the neutral and photo presets still work fine as strong general captioning instructions - just know what you've loaded. There's no separate jailbreak toggle on this node; that's the Advanced variant's job.

    Input and output

    • preset (combo) - the only input. Roughly 29 options covering neutral/action/photo/toon captioning, Ideogram JSON instruction sets, and the video timeline family.
    • system_instruction (STRING) - the full system prompt text. Wire it into the system-prompt slot of whatever VLM node you're using, typically alongside one of the query-preset nodes from the same pack.

    Installing it

    Ships in ComfyUI-UtilsCollection - Manager search, or:

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

    Restart. Deps are opencv-python and typing-extensions; nothing to download. If ComfyUI suggests replacing this node with UC_VLMSysInstrPresets, accept - same thing, current name.

    Categoryadvanced/text

    Inputs (1)

    NameTypeDefaultDescription
    presetCOMBOaction_system_instruction29 options: action_system_instruction, action_system_instruction_crude, action_system_instruction_qwen, cinematic_dumb_intelligent, ideogram_4_json_instruction, ideogram_4_json_instruction_color, +23

    Outputs (1)

    NameTypeDescription
    system_instructionSTRING