WanVideo Add SCAIL Reference Embeds

SCAIL-2 is Z.ai's motion-transfer model built on Wan 2.1, and its headline trick is that it dropped the stick figure. Where Wan Animate and most pose-transfer tools convert the driving video into a DWPose skeleton first - throwing away depth and contact in the process - SCAIL-2 takes the reference image and driving video more directly, which is what unlocked animal driving, multiple characters, and cross-identity replacement. This node is the piece that injects your reference character into a SCAIL-2 run: it adds the reference embeds onto an existing image-embeds bundle so the motion gets mapped onto your subject.

If you're doing "take this dance video, put my character in it, and don't stretch them to match the dancer's proportions," this is a node in that chain. Proportion preservation is the specific thing SCAIL beat Wan Animate on.

How it works

You give it your reference image plus the Wan VAE, and it encodes and folds the reference into the WANVIDIMAGE_EMBEDS object flowing to the sampler. SCAIL-2 then transfers the motion from its driving input onto that reference identity. The strength and percent controls decide how strongly, and over what part of the denoise, the reference asserts itself.

The inputs and outputs that matter

• embeds (WANVIDIMAGE_EMBEDS, required) - the existing embeds bundle you're adding the reference to. This node augments a pipeline rather than starting one.
• ref_image (IMAGE) - your subject: the character the motion gets applied to.
• vae (WANVAE) - the Wan VAE, needed to encode the reference into latent space.
• strength (default 1) - how hard the reference is pushed. 1 is the sane default; lower it if the reference is fighting the motion, raise it if identity is slipping.
• start_percent / end_percent (0 / 1) - the denoise window the reference conditioning covers.

Optional clip_embeds (WANVIDIMAGE_CLIPEMBEDS) adds CLIP-vision guidance from the reference. Output is the updated image_embeds, which continues to the sampler.

How to install it

Comes with 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. You'll also need the SCAIL-2 weights. Grab Comfy-Org's scaled/MXFP8 weights (Comfy-Org/SCAIL-2) or a community GGUF (realrebelai/SCAIL-2_GGUF) rather than the raw zai-org repo, which is a 65.6GB download. fp8 runs comfortably on 16GB; Q4 GGUF gets you onto 8GB at low resolution.

Common issues & troubleshooting

The face keeps changing across the clip. This is SCAIL-2's known weak spot, and the community answer is consistent: drop the LightX2V speed LoRAs, raise steps, raise CFG. If it still drifts, that's a model limitation, not a setting you're missing - SCAIL-2's native context is 81 frames and identity wanders across longer, chunked clips.

It's brutally slow. Expected. SCAIL-2's cost is the recurring complaint - think ~5-6 minutes for 81 frames on a 4070 Ti, ~19 minutes for a 14-second clip on a 5060 Ti. Nothing about this node changes that; it's the model.

Body morphing on fast movement. Persists even in good workflows. Keep the driving motion reasonable and don't expect perfection on complex, high-speed choreography.