Nodes/ComfyUI-UtilsCollection/VLM System Instruction Advanced Presets Experimental
ComfyUI Node

VLM System Instruction Advanced Presets Experimental

The advanced captioning presets with the bleeding-edge video ones

By silveroxides·Created 2 months ago·Updated about 7 hours ago· 24
VLM System Instruction Advanced Presets Experimental
    • system_instruction
    presetvideo_timeline_minimax_h3_base_system_instruction
    jailbreakfalse
    system_query
    user_query

    Experimental, in pack-speak, means "the newest stuff ships here first." UC_VLMSysInstrAdvPresetsExperimental is the advanced preset node pointed at the experimental preset library - and that library's defining feature is the video-timeline system instructions, including the MiniMax H3 family built for the H3 video model's chat-style prompting. If you're captioning video frames or building prompts for the timeline-structured models, this is the variant you want.

    What it is

    Structurally it's a clone of UC_VLMSysInstrAdvPresets - same preset dropdown, same jailbreak toggle, same system_query / user_query injection - but it pulls from system_instructions_vlm_experimental instead of the stable library. The default preset is the tell: video_timeline_minimax_h3_base_system_instruction.

    What the experimental library contains

    The experimental set is dominated by video-timeline instructions:

    • video_timeline_system_instruction and _crude - the base timeline captioning prompts.
    • video_timeline_minimax_h3_base_system_instruction - the default, built for MiniMax H3's timeline-driven prompting.
    • video_timeline_minimax_h3_reference_system_instruction - the reference-image variant.

    These matter because MiniMax H3 (the open 33B video model) and the newer Qwen-VL-era pipelines prompt with ordered <Picture N> timeline blocks rather than a single flat caption, and the system instruction has to teach the VLM that structure. The base H3 preset does exactly that: it defines how each timeline entry maps to a target timestamp and how the model should describe progression. That's the kind of detail you'd otherwise have to reverse-engineer from a working workflow.

    Since it's the experimental branch, expect these to drift - presets get renamed, split, or folded into the stable library as the underlying models settle.

    The inputs and output

    • preset - the experimental dropdown.
    • jailbreak - wraps the result in extra framing.
    • system_query / user_query - injected into the preset's query slots.
    • Output: system_instruction (STRING).

    Installing it

    Part of ComfyUI-UtilsCollection:

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

    Restart ComfyUI (or use Manager, search "ComfyUI-UtilsCollection"). No models, minimal deps, needs a current ComfyUI.

    Common issues

    Treat the "Experimental" label as a warning label: the preset list can change between updates, and a workflow saved against a renamed preset will silently fall back or error on load. If you're on the stable captioning path, the non-experimental advanced node is the safer default; reach for this one when your model needs the H3/timeline vocabulary. And remember the injection semantics carry over - system_query overrides preset rules, user_query steers, and jailbreak changes refusal behavior, so use it the same way you would on the stable sibling.

    Categoryadvanced/text

    Inputs (4)

    NameTypeDefaultDescription
    presetCOMBOvideo_timeline_minimax_h3_base_system_instruction4 options: video_timeline_minimax_h3_base_system_instruction, video_timeline_minimax_h3_reference_system_instruction, video_timeline_system_instruction, video_timeline_system_instruction_crude
    jailbreakBOOLEANfalse
    system_querySTRING
    user_querySTRING

    Outputs (1)

    NameTypeDescription
    system_instructionSTRING