Animal Pose Retarget Prompt Helper

This is the odd one out in ComfyUI-WanAnimalPreprocess. The other nodes crunch pixels; this one writes a sentence. It looks at the detected animal pose and tells you what to put in your prompt - specifically, it describes whether the animal's legs are visible, which is the kind of thing you'd never think to prompt for and that makes the difference between a clean pose transfer and a dog with a phantom fourth leg.

Here's why it exists. In the retarget workflow - template video pose transferred onto a reference animal - Wan Animate is much happier when the prompt says something about limb visibility. If the template animal is standing facing you with all four legs and paws visible, your prompt should say so. If it's a side profile with the far legs hidden, the prompt should reflect that instead of confidently asserting legs that aren't there. Getting that text wrong is how you end up with generative anatomy that no amount of seed-flipping fixes.

The mechanism is simple and easy to trust: it reads the keypoints from the template pose and checks confidence values on the front-leg indices (shoulder/elbow/front paw) and back-leg indices (hip/knee/back paw). If a keypoint passes the visibility threshold, that leg set counts as visible. Then it produces one of a small set of canned prompts: "All four legs and paws are visible," or a front-legs-only variant, or a minimal "Change the animal to face forward."

Inputs and outputs

One input: pose_data, straight from Animal Pose and Detection. Two outputs, both plain STRING:

• prompt - the sampler prompt describing the template animal's visible legs.
• retarget_prompt - the same description in the retarget convention, for the reference animal.

Wire the first into your positive prompt on the sampler, the second into whatever your retarget workflow passes for the reference. If no reference pose exists in the pose data, both outputs fall back to "Change the animal to face forward." - which is a fine default and also a hint that you probably wanted to pass a retarget_image upstream.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-WanAnimalPreprocess.git
pip install -r requirements.txt

Or search "ComfyUI-WanAnimalPreprocess" in ComfyUI Manager. No models, no extra dependencies - it's pure logic on the pose data.

Be honest about what this is: it's a helper, not a magic prompt engineer. You'll get a handful of legibility-based sentences, and you still need to write the actual prompt (subject, style, scene) around them. And it only helps in retarget workflows - for plain pose-to-video generation where you're not transferring onto a reference animal, the prompt output is close to noise. It's the one node in this pack you can safely skip if you're not doing retargeting, but when you are, it saves you the exact guess that produces the weirdest failures.