WanVideo Add OneToAll Pose Embeds

This node adds pose control from One-to-All, a character-animation model whose pitch is right in its name: it's "alignment-free" character animation and image pose transfer. The usual pose-transfer headache is alignment - your reference character and your driving pose have to be roughly the same size, framing and proportions, or the result warps. One-to-All is built to relax that, so you can drive a character with a pose sequence that wasn't captured to match it. This node folds that pose control into your Wan embeds chain.

Like the other Add…Embeds nodes in the wrapper, it's a link in a pipeline: it takes the running image_embeds, mixes in a pose_images sequence, and returns embeds with the pose drive added - WANVIDIMAGE_EMBEDS in, WANVIDIMAGE_EMBEDS out. There's a companion reference-embeds node for the character identity; this one is the motion half.

How it works

You supply a sequence of pose_images describing the motion. The node conditions the generation to follow that pose over a slice of the denoising schedule, at a strength you choose, and hands back the updated embeds. The "alignment-free" part is the model's doing, not a knob here - you just provide the pose and the character reference separately and let One-to-All reconcile them. Two extras give you finer control: a pose_prefix_image to seed the pose sequence, and a dedicated pose_cfg_scale so the pose can be guided independently of your main CFG.

The inputs and outputs that matter

• embeds (WANVIDIMAGE_EMBEDS) - the running embeds from earlier in the chain.
• pose_images (IMAGE) - the pose sequence that drives the motion. The core input.
• strength (default 1) - how firmly the pose binds. Main dial: down if identity is getting distorted to match the pose, up if the pose is being ignored.
• start_percent / end_percent (0–1) - the schedule window pose control is active over. Guide early, release before the end if the output looks over-constrained.
• pose_prefix_image (optional IMAGE) - a lead-in frame for the pose sequence, useful for setting the starting pose cleanly.
• pose_cfg_scale (default 1.5) - guidance strength specifically for the pose signal, separate from the sampler's CFG. Raise it to make the pose more authoritative.

Output: image_embeds (WANVIDIMAGE_EMBEDS) → the next Add node or WanVideoSampler.

How to install it

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. You'll need the One-to-All weights on disk (Kijai mirrors Wan-adjacent models under huggingface.co/Kijai/WanVideo_comfy) plus your base Wan model, VAE and text encoder. You also need a source of pose_images - a pose estimator like NLF or DWPose over your driving video, or hand-authored pose frames.

Common issues & troubleshooting

Character identity distorts to fit the pose. Ease strength down. Alignment-free doesn't mean distortion-free - push the pose too hard and the character stretches to match it. Also make sure your reference-embeds (the identity half) are wired in so the model has something to hold.

Pose isn't followed. Raise strength and/or pose_cfg_scale, and confirm your pose_images cover the full frame range you're rendering. A pose sequence shorter than the clip leaves the tail undriven.

Jittery start. Feed a pose_prefix_image to establish the opening pose cleanly rather than letting the first frames guess.

Chain wiring. These Add nodes compose in order and interact - keep the reference and pose paths straight, and confirm you're passing genuine WANVIDIMAGE_EMBEDS from the correct base node. Node missing after install means update the pack and restart.