Instruct Prompt Video Presets

The most boring kind of node is sometimes the one that saves you real time, and that's this one. UC_InstructPromptVideoPresets is a dropdown that hands you a finished, carefully worded style-instruction prompt for video generation. Pick "3D" or "photorealism" or "anime" and out comes a long, structured instruction that tells the model to keep the subject's pose, keep the lighting plausible, keep the camera angle - the stuff you'd write by hand and get subtly wrong half the time.

It's the video sibling of UC_InstructPromptPresets (same presets, different wrapper). The output is a plain string, so you wire instruct_prompt into whatever video text-encode node you're using - this pack also ships a UC_TextEncodeLtxv2SystemPrompt and friends if you want the full LTX-Video treatment, but it feeds any encoder that accepts a raw prompt string.

How it works

Under the hood it's embarrassingly simple, which is the point. The pack keeps a collection of INSTRUCT_PROMPT_STYLE_* templates in presets_collection.py - things like STYLE_PHOTOGRAPHY_PHOTOREALISM, STYLE_ANIME, STYLE_3D, STYLE_CYBERPUNK, STYLE_PIXAR. Each one is a multi-sentence instruction that repeats the important constraints ("Keep subject position and their pose the same as the reference") because video models genuinely respond better to redundant phrasing. The node wraps your pick as an instruction-role prompt (to_video_prompt(..., role="instruction")) before handing it to you, which is the one thing that makes this distinct from just pasting text into a string node.

Two inputs, both obvious:

• preset - the style dropdown. STYLE_3D is the default.
• escape_parentheses - if your preset text would contain literal ( or ) and whatever's downstream chokes on them, flip this. It escapes them so a style phrase like "(vivid colors)" can't break your conditioning parser.

That's it. One output: instruct_prompt, a STRING.

Install and gotchas

This is part of the silveroxides/ComfyUI-UtilsCollection pack, a broad utility collection. Install once, get ~150 nodes:

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

then restart ComfyUI (or use ComfyUI Manager and search "ComfyUI-UtilsCollection"). Dependencies are light - just opencv-python and typing-extensions - so nothing heavy to babysit. The pack is AGPL-3.0, fine for local use.

Where people trip: these presets are instruction prose, not tags. They work best when the model was trained to follow that kind of instruction text (the Qwen-VL / LTX-Video style encoders), and can look like word salad if you dump them straight into a Krea 2 or SDXL prompt box. Also, don't expect magic - a preset won't fix a bad base prompt, it just removes the "make it look 3D" half of the writing from your plate.

The honest take: if you're already happy hand-writing style instructions, skip it. If you batch out a lot of videos and want consistency across a run, this is a nicer way to guarantee every frame block gets the same carefully-written instruction than trusting yourself to retype it at 2am.