Portrait Prompt

Describing a face is where prompts get vague. "Attractive woman" - what does the model do with that? Portrait Prompt is the opposite of vague: it hands you forty dropdowns covering lens angle, nationality, gender, age, body type, face shape, eyes, brows, lashes, nose, hair, mouth, lips, and beard, and assembles your choices into a weighted portrait prompt. It's the flagship of ComfyUI-PromptWrapper, a mostly Chinese-audience pack that flies under the radar in English discussions but puts real work into its prompt datasets.

How it works

Every dropdown is backed by a JSONL file in the pack's assets/ folder, in your choice of Chinese or English (the language input swaps the whole set). Most fields follow the same pattern: -- omits it, random 🎲 picks one via seed % dataset_length, and the weight slider (0–2) controls how strongly it lands - weight 1 is bare text, anything else becomes (word:1.3). A few fields are weight-only, like pretty_face_weight, ordinary_face_weight, and ugly_face_weight, which pull fixed phrases and let you dial realism up or down.

The one genuinely clever input is nationality: nationality_1 and nationality_2 with nationality_mix_weight. Pick two and the node emits the [a:b:0.5] blend syntax, so you can mix ethnicities rather than pick one.

The inputs that matter

You can't set forty fields every run. The ones that carry the image:

• gender, age, hair_style, hair_color - the big four. Everything else at -- still gives you a recognizable portrait.
• lens_angle - 128 options, and it's the highest-leverage input for shot composition.
• seed - reproduces random picks deterministically.
• preset_prompt + optional input_prompt - both prepend your own text, so you can say "cinematic, 8k" without losing the generated part.

The single text output goes straight into a CLIP Text Encode.

Install

ComfyUI Manager → Custom Nodes Manager → search ComfyUI-PromptWrapper → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/clouddreamfly/ComfyUI-PromptWrapper

All the datasets ship with the pack - nothing to download.

Where people get burned

The weighting caveat applies loudly here: (word:1.3) and [a:b:0.5] are CLIP-encoder syntax. On SDXL-lineage checkpoints (Illustrious, Pony, SD 1.5) they work as intended. On 2026's LLM-encoded models - Flux's T5, Z-Image, Anima - those weights are silently discarded, so the node degrades to a phrase picker and the mix-weight feature does nothing. Also default language is Chinese; flip it to English for clean output on Western models. And a sanity note: this builds the description, not the pose - for stance and movement you'll want the pack's Portrait Pose Prompt wired into the same prompt.