VLM System Instruction Presets Experimental

Same dropdown, different shelf. UC_VLMSysInstrPresetsExperimental is the base VLM system-instruction selector pointed at the pack's experimental preset library - the one where the newest prompts land before they're considered stable. Its headline content is the video-timeline instructions for MiniMax H3, which is exactly what the default preset name (video_timeline_minimax_h3_base_system_instruction) advertises.

What it is

The experimental counterpart to UC_VLMSysInstrPresets. Identical shape - one preset dropdown, one system_instruction STRING output - but it reads from system_instructions_vlm_experimental instead of the stable library. If you want the current stable captioning presets (neutral/action/photo/toon), you don't need this node; if you want the newest video-timeline prompts for the H3-era models, this is where they live.

What's in the experimental library

The set is small and focused on video:

• video_timeline_system_instruction and video_timeline_system_instruction_crude - timeline captioning prompts.
• video_timeline_minimax_h3_base_system_instruction - the default; the H3 timeline prompt.
• video_timeline_minimax_h3_reference_system_instruction - the reference-image variant.

These exist because MiniMax H3 and the newer Qwen-VL video pipelines prompt with ordered timeline entries - each <Picture N> block tied to a timestamp - and a system instruction has to teach the VLM that structure. That's genuinely useful if you're building video-captioning or video-to-prompt workflows, since it's the exact wording the pack's own H3 examples rely on.

The flip side is the "Experimental" contract: the list is volatile. Presets move into the stable library as models settle, get renamed, or get replaced. Building a workflow that hardcodes an experimental preset name is borrowing against a moving target.

The input and output

• preset - the experimental dropdown.
• 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

The only real trap is treating experimental presets as stable. Update the pack and a saved workflow can reference a renamed preset, which loads as a mismatch rather than an error - the workflow still runs, but with a different instruction than you tuned. If you're doing reproducible captioning, pin your preset text (save the system_instruction output to a text file) so a pack update can't silently change your results. And for video work specifically, pair this with the H3 timeline workflows in the pack's README - the preset is written to feed exactly that prompt shape.