Instruct Prompt Presets (Legacy)

The instruction-editing models - the ones that take "make the jacket leather" and actually edit the image - are the least forgiving about prompt phrasing. They want an instruction, complete with preservation constraints: keep the pose, keep the angle, keep the unedited areas in focus. Instruct Prompt Presets (Legacy) hands you those instruction blocks pre-written. Pick a style, get a paragraph that tells the model both what to change and what not to touch.

Mechanically it's the same shelf-and-escape design as the other preset nodes: a lookup into the pack's collection of instruction templates, one per style (STYLE_3D, STYLE_ANIME, STYLE_DISNEY, STYLE_GHIBLI, and dozens more), with the escape_parentheses guard for downstream weight parsers. The returned text is written in instruction voice - "Modify any subjects' appearance to look like… Keep pose. Keep angle." - because that's the register these editing models were trained on. It's not a tag list and it's not a description; it's a directive with guardrails baked in.

The inputs that matter

• preset - the instruction shelf. Same style names as the bonus presets, different register.
• escape_parentheses - off by default; flip it on if a downstream parser chokes on literal parens.

The output

instruct_prompt - a STRING meant to be the primary instruction for an edit pass, not a supplement. Where the bonus presets append flavor, this is the actual command block. Wire it into the prompt input of your instruction-edit encoder (Krea2-style edit, Qwen edit, whatever the pack's companion edit nodes feed).

Installing it

Part of silveroxides/ComfyUI-UtilsCollection. Install via ComfyUI Manager (search ComfyUI-UtilsCollection) or:

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

Restart after. Dependencies: opencv-python, typing-extensions; the templates ship in the pack. "(Legacy)" marks the compat alias for the pack's canonical UC_InstructPromptPresets.

Where people get burned

The instruction blocks are intentionally verbose - they're spelling out preservation constraints ("keep subject position and their pose the same as the reference") because the editing models drift without them. Resist the urge to trim them into a short command; the "keep" language is doing real work. Also note these are written for editing models, not for plain text-to-image. Using an instruct preset as your whole T2I prompt gives you a prompt that's mostly about not changing things, which is a weird thing to feed a text-to-image sampler. Match the preset family to the job: instruct for edits, bonus for style.