Human Composer Prompt Builder

The hardest part of prompting a human character is rarely the model - it's remembering the details. You type "woman, ponytail, beach," get a generic render, and then spend five generations bolting on stray tags that never quite land. Human Composer Prompt Builder is the anti-blank-page tool: one node with roughly thirty dropdowns that hands you a finished prompt in the format your checkpoint actually wants.

The name is a lie in the best way - it doesn't call any API, needs no LLM, no key, no model files. It's pure string assembly in Python.

How it works

You pick values from dropdowns grouped by subject, appearance, hair, clothing, accessories, scene, and camera. Every field ends with a Custom value… option that unlocks a plain text input right next to it, so you're never stuck with the author's vocabulary. Fields you leave at - not set - are simply skipped - no dangling Gender: not set garbage in your prompt.

Under the hood there's exactly one renderer per model family, switched by output_format. That switch is the whole ballgame:

• SDXL – Structured - labeled Label: value blocks separated by blank lines. Readable, and easy to hand-edit in a "Preview as Text" node before it hits the encoder.
• SDXL – Comma Tags - Danbooru-style comma tags for tag-trained SDXL models.
• FLUX – Natural Prompt - proper sentences with article handling ("An" vs "A"), "The person has…", "The person is wearing…".
• Pony – Comma Tags - comma tags with score_9, score_8_up, score_7_up prepended.

This is the bit people get wrong, and it's where the KB's advice matters: tags, prose, and Pony's score system are three different prompting paradigms, and the same words behave completely differently across them. Run a Pony checkpoint through the SDXL formats and you silently lose the quality tags the model was trained on; feed a FLUX checkpoint comma tags and it reads like a grocery list instead of an instruction.

The fields that matter

Honestly, output_format decides everything - set that first. After that, the subject block (gender, age, ethnicity) and hair are where this node earns its keep: nine independent hair dimensions (length, texture, style, parting, bangs, color) that you'd otherwise forget to describe at all. And when presets aren't enough, additional_details is a multiline free-text box for anything not covered. There's no node with more field groups; the value is the completeness, not any single dropdown.

The output is a single prompt string - wire it straight into a CLIP Text Encode node's text input and queue.

Install

Clone into ComfyUI/custom_nodes/ and restart:

cd ComfyUI/custom_nodes
git clone https://github.com/max-dingsda/ComfyUI-HumanComposer

No requirements to install - the pack's requirements.txt is empty and pyproject.toml lists zero dependencies. ComfyUI Manager can grab it too, since the repo publishes to the Comfy registry; search for the pack title there.

Gotchas

The format switch is a trap only if you pick it wrong, so don't. Switching formats later doesn't hurt - your field values are unchanged, it just re-renders.

Two honest caveats. First, the Structured format with all thirty fields filled gets heavy - newline-separated labels are great for humans, but caption-trained models tolerate them less; Comma Tags is usually the safer default if you're not sure. Second, and worth stating plainly: this is a prompt-writing aid, not a consistency tool. It won't keep a character identical across generations - that's IPAdapter- and LoRA-shaped problems, not a text node's. What it does, reliably, is stop you from staring at a blank prompt box because you don't know the English word for "waist-length hair with side-swept bangs."