Instruct Prompt Video Presets (Legacy)

Editing a video with an instruction model is harder than editing an image for one simple reason: the instruction has to survive across frames. Instruct Prompt Video Presets (Legacy) is the video version of the instruct-preset shelf - same style dropdown, same preservation-heavy instruction voice, but the text is passed through the pack's video-prompt rewrite so it reads as video guidance rather than a still-frame edit.

Mechanically it takes the same instruction templates as the image version (STYLE_3D, STYLE_ANIME, etc.), runs them through the same "to-video-prompt" transform used across the pack, and appends the role-specific video guidance for the "instruction" role. What changes is the register: the output leans on keeping the subject, pose, and scene stable across time, which is what a video edit model needs to hear. The escape_parentheses guard behaves exactly as in the other preset nodes.

The inputs that matter

• preset - the style shelf, same names as the image version.
• escape_parentheses - off by default; on if a downstream parser objects to literal parens.

The output

instruct_prompt - a STRING for the primary instruction slot of a video-edit encoder. Because the transform appends video-stability guidance, this block is doing double duty: it commands the edit and reminds the model the scene should keep moving and staying consistent. Wire it into your video-edit prompt path alongside the subject/action text.

Installing it

Part of silveroxides/ComfyUI-UtilsCollection. ComfyUI Manager → search ComfyUI-UtilsCollection → install → restart, or:

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

Restart after. Dependencies: opencv-python, typing-extensions. No downloads; templates ship in the pack. "(Legacy)" marks the compat alias for the pack's canonical UC_InstructPromptVideoPresets.

Where people get burned

The frame-consistency language makes these read oddly if you're used to image instructions - "keep subject position" appears next to motion guidance, and that's not a mistake, it's the point. The bigger trap is mixing families: feed a video-instruct preset into an image edit model and you're wasting half the prompt on temporal language the image encoder can't act on. And as always with the instruct family, don't shorten the blocks - the preservation constraints are the part doing the work, and video models drift harder than image models when you take them out.