WanVideo Add SteadyDancer Embeds

SteadyDancer is a pose-transfer model (from MCG-NJU, a 14B model built on Wan) that drives a character's motion from a pose sequence - the community pitch when it landed was that it's better than Wan Animate at pose-driven image-to-video. This node is how you wire SteadyDancer's pose conditioning into a WanVideoWrapper graph: it takes pose latents and folds them onto an existing image-embeds bundle, with separate strength dials for spatial and temporal pose influence. If your goal is "make my character do the motion in this pose video," this is the injection point.

It sits in the same family as the SCAIL and Lynx embed nodes - they all take a WANVIDIMAGE_EMBEDS bundle and add one specific kind of conditioning to it. SteadyDancer's flavor is pose.

How it works

You feed it pose latents - the driving motion, encoded to latent space - and it adds them to the embeds the sampler will use, so the generated video follows that pose sequence. The interesting part of the interface is the split between spatial and temporal pose strength: how strongly each frame's pose is enforced, versus how strongly the motion between frames is enforced. That separation lets you keep tight per-frame posing without over-constraining the flow, or vice versa.

The inputs and outputs that matter

• embeds (WANVIDIMAGE_EMBEDS, required) - the bundle you're adding pose conditioning to.
• pose_latents_positive (LATENT) - the driving pose sequence. This is the motion you want your character to perform.
• pose_strength_spatial (default 1) - how hard each individual frame's pose is enforced. Raise for stricter posing, lower if the character looks pinned or stiff.
• pose_strength_temporal (default 1) - how hard the motion across frames is enforced. This is your lever for smooth-vs-literal motion.
• start_percent / end_percent (0 / 1) - the denoise window the pose conditioning is active over.

Optional inputs: pose_latents_negative (a pose to steer away from) and clip_vision_embeds (WANVIDIMAGE_CLIPEMBEDS, appearance guidance). Output is the updated image_embeds for the sampler.

How to install it

Ships in the WanVideoWrapper. ComfyUI Manager: search ComfyUI-WanVideoWrapper, install, restart. Manual:

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

then restart. On top of base Wan you'll need the SteadyDancer weights in your models folder - it's a separate checkpoint. Heads up from the community: SteadyDancer's text encoder is a chunky ~11GB, so budget disk and RAM for it. As with other newer features in this pack, if the node doesn't appear after an update, check whether it's on the dev branch (git switch dev).

Common issues & troubleshooting

Motion is too literal / jittery. Ease off pose_strength_temporal so the model smooths the motion between poses instead of snapping to every frame. If individual poses are wrong, that's pose_strength_spatial instead.

Character ignores the pose. Both strengths too low, or your pose latents aren't clean. Make sure the pose sequence encoded correctly and bump the spatial strength.

It's a Wan-tier compute bill. SteadyDancer is a 14B pose model on Wan - expect the same minutes-per-clip reality as the rest of the ecosystem, not a fast lane. It's a quality-of-motion play, not a speed one.