Portrait Pose Prompt

"Dynamic pose" is a prompt phrase that produces nothing in particular, because the model has to invent what dynamic means. Portrait Pose Prompt takes the other approach: it breaks a pose down into body parts and gives you a dropdown for each. Standing, walking, arms crossed, head tilted, one leg lifted - you assemble the stance from 350 poses, 531 actions, and per-body-part micro-actions, and the node turns it into a weighted prompt. It's from ComfyUI-PromptWrapper, and it pairs naturally with the pack's Portrait Prompt, which handles the face while this handles the body language.

How it works

Same machinery as the other builders in this pack: each dropdown reads from a bundled JSONL dataset in either Chinese or English, random 🎲 picks via seed % dataset_length, and weight sliders (0–2) set emphasis - weight 1 is the bare phrase, anything else wraps it as (phrase:1.3). Parts left at -- are simply omitted from the joined comma string.

The inputs that matter

• pose - the overall stance, and the one you'll set first. 350 options.
• action - what the subject is doing on top of the pose. 531 options.
• head_action, arms_action, hands_action, legs_action - the detail sliders when you want something specific like "hands behind back" or "head tilted". Each has a weight, so you can make the hands thing loud and everything else quiet.
• chest_action, waist_action, feet_action - the fine-grained stuff for when you're chasing a specific look.
• seed - makes any random pick reproducible.
• language, enable, preset_prompt, optional input_prompt - the standard pack controls; input_prompt gets prepended to whatever pose text is generated.

One text output, straight into a CLIP Text Encode (or into a Combine Prompt alongside a Portrait Prompt result).

Install

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

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

Datasets ship with the pack; nothing to download.

Where people get burned

Pose words are an area where models differ wildly. A weighted phrase like (hands behind back:1.2) is respected by SDXL-lineage checkpoints, but LLM-encoded 2026 models drop the weights, and some simply don't carry the pose vocabulary the datasets assume - so test against your checkpoint before you blame the node. Also, this is text-only posing. If you need a pose to be enforced geometrically - a real skeleton your image must follow - you want a ControlNet OpenPose workflow, and this node becomes a source of good caption text instead. For most portrait work, though, it beats typing "dynamic pose" and praying.