Construct LLM Prompt Extra

If Construct LLM Prompt is the "what do you want the LLM to do" node, Construct LLM Prompt Extra is the "how detailed should the description get" node. It's the companion piece that assembles the detail instructions: mood, lighting, camera details, materials, weather, time of day - the layer of specificity that turns a generic caption into something a text-to-image model can actually render.

The interface is a wall of booleans, and that's the feature. Toggle on describe_mood, describe_setting, describe_actions, include_lighting, include_camera_details, include_materials, include_motion, include_weather, include_time_of_day, include_reflections, include_symmetry, include_scale, include_context, include_anomalies, include_cultural_references, include_emphasis_on_details... each one appends a pre-written instruction line telling the LLM to cover that dimension. Plus extra_instructions, a multiline free-text field for anything the toggles don't cover. The output is a single extra string with the enabled instructions joined together.

Mechanically it's the same text-assembly pattern as its sibling: read the config in llm_prompts.json, concatenate the canned instruction for each enabled toggle, append your free text. Where it differs is division of labor: Construct LLM Prompt pulls its toggles from the "style / quality / content focus" categories, while this node draws from the "detail" category (and any others those don't cover). Together they're meant to be joined - main instruction from one, detail layer from the other - and the pack's LLM workflow usually feeds the combined result into the chat tab or an LLM text-generation node.

The same honest caveats apply as the parent node. The LLM features are the part of Sage Utils the README calls work in progress, and the pack logs connection warnings at startup when LM Studio or Ollama aren't running - harmless, silenced with SAGEUTILS_LOG_LEVEL=WARNING. But this node, like its sibling, is pure string construction: it works with zero LLM configured. You can build the instruction, copy it, and use it with any model anywhere.

One real opinion, since you're here: the detail toggles shine with vision models. If your local LLM can look at an image, feeding it a reference image plus "include lighting, describe mood, include camera details" produces the kind of rich caption that's actually useful for img2img or training data - the difference between "a woman standing in a room" and something you could feed back into a prompt and get close to the source. The toggles are the difference between the two outputs.

Install

ComfyUI Manager → search Sage Utils → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/arcum42/ComfyUI_SageUtils.git
cd ComfyUI_SageUtils && pip install -r requirements.txt

Restart ComfyUI. No extra dependencies beyond the pack's dynamicprompts. Optionally run ollama pull llama3.2 or set up LM Studio if you want to actually feed the result to a local model.