Pose Retarget Prompt Helper

Every other node in this pack outputs images or skeletons. This one outputs sentences. Pose Retarget Prompt Helper looks at your detected pose and writes a text prompt that tells an image model - it was built for Flux Kontext, per the author - what pose to put a character in. It's the odd one out, and if you're not doing the specific thing it's for, you can skip it.

The thing it's for: pose retargeting that starts from a static character image. Before you can map a dancer's motion onto a reference character, that character needs to be in a clean, standardized pose - facing forward, arms readable, hands and feet visible. The prompt helper generates the text that gets the character there, based on what it can actually see in the template pose.

How it works

It inspects the body keypoints in the pose_data you feed it and checks whether the arms and legs are visible and confidently detected (confidence ≥ 0.75 in normalized coordinates). Arms live at keypoints 3/4/6/7, legs at 9/10/12/13. Then it picks from a small set of canned prompts:

• Both arms and legs visible → the full prompt: a standard T-pose (landscape) or arms-down forward stance (portrait), with "Feet and Hands are visible in the image."
• Only arms visible → a shorter variant without the feet claim.
• Nothing confidently visible → just "Change the person to face forward."

Whether the character is wide or tall (landscape vs. portrait frame) decides which phrasing gets used.

The two outputs

• prompt - for the template pose (the driving side).
• retarget_prompt - for the reference image.

You'd wire both into whatever text encoder your Kontext-style workflow uses for the two character images. If the pose data has no reference retargeting behind it (no refer_pose_meta), both outputs fall back to the same "Change the person to face forward." string - which is the honest floor of what it can do.

Where it fits

This node makes the most sense in the Flux-Kontext-era retargeting experiments, where the trick was: render the reference character into a standard pose with an image model, then drive that clean pose with the template's motion. If you're following the mainstream path - Pose and Face Detection → Draw ViT Pose → Wan Animate - you don't need it at all. Reach for it only when your retargeting pipeline includes an image-generation step that needs a natural-language standing pose.

Gotchas

It's not a general prompt writer. It emits one of a handful of fixed strings based on joint visibility - there's no customization, no style. If you need "arms raised, confident stance, dramatic lighting," this isn't the tool; write the prompt yourself.

The outputs only make sense with a reference. With bare pose_data (no retargeting), both outputs are identical and near-useless. The helper is designed for the branch of the workflow where a reference character is being prepared, so make sure you're feeding it pose_data that came from a detection pass that used retarget_image.