WanVideo Add SCAIL Pose Embeds

This node adds a pose guide into a SCAIL motion-transfer workflow. SCAIL-2 is Z.ai's motion-transfer model built on Wan - you give it a reference character and a driving video, and it makes the character perform the driving motion. This node is one of the control layers you stack onto the embeds before sampling: it takes the running image_embeds, folds in a sequence of pose_images, and returns the embeds with that pose control baked in.

It fits the wrapper's "embeds chain" pattern. You start from a base image-embeds node, then pipe through a series of Add…Embeds nodes - reference here, pose there - each one taking WANVIDIMAGE_EMBEDS in and handing WANVIDIMAGE_EMBEDS out, until the final embeds carry every control signal you want. This node is the SCAIL pose link in that chain.

How it works

SCAIL-2's headline is that it dropped the stick-figure skeleton that Wan Animate depends on - the KB is emphatic that DWPose throws away depth and contact before the model ever sees the motion, and SCAIL keeps those. So think of this pose path as an explicit control you can add when you want to steer the pose directly, on top of what SCAIL infers from the driving video. The node encodes your pose_images through the Wan VAE and injects them as conditioning over a chosen slice of the denoising schedule, at a chosen strength.

The inputs and outputs that matter

• embeds (WANVIDIMAGE_EMBEDS) - the running embeds from earlier in the chain. Required in, and the same type comes out.
• pose_images (IMAGE) - your pose sequence, one per frame. This is the drive.
• vae (WANVAE) - needed to encode the pose frames to latent space.
• strength (default 1) - how hard the pose binds. The main dial. Ease it down if the output looks locked to the pose and loses the reference character's identity; push up if the pose is being ignored.
• start_percent / end_percent (0–1) - the schedule window where pose control is active. Guiding early and releasing before the end often gives the model room to clean up detail. If motion looks over-constrained, try ending around 0.6–0.8.

Output: image_embeds (WANVIDIMAGE_EMBEDS) → the next Add node, or straight to 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 SCAIL weights on disk (ComfyUI's own scaled build at Comfy-Org/SCAIL-2 is the sane download - the raw zai-org repo is enormous) plus a Wan model, VAE and text encoder. Heads up from the KB on cost: SCAIL is not fast - figure several minutes per 81-frame clip even on strong cards, more at higher resolution.

Common issues & troubleshooting

Identity wanders / face keeps changing. This is a known SCAIL trait, and heavy pose forcing makes it worse. Drop strength, and if you're stacking speed LoRAs, back them off - the community fix for SCAIL face consistency is fewer LoRAs, more steps, and not over-driving control.

Pose is ignored. strength too low, or your pose_images don't align frame-for-frame with the intended output length. Make sure the pose sequence matches the frame count you're rendering.

Order in the chain matters. These Add nodes compose, and the reference-embeds and pose-embeds layers interact. If results are off, check that you fed genuine WANVIDIMAGE_EMBEDS from the correct base node and didn't cross wires between the reference and pose paths.

Node missing after update. Update WanVideoWrapper and restart - SCAIL nodes ship with the wrapper and land via updates.