XB_WanSCAILToVideo

If you want to try Wan SCAIL - the pose-driven character animation method - without committing to XB's whole ParamBus + relay scaffolding, this is the node. It's the plain single-segment SCAIL-to-video builder: you hand it conditioning, a pose video, a reference image, and optionally a CLIP-Vision encoding, and it hands back the conditioning plus a latent for your sampler to fill in. One shot, one segment, no offset bookkeeping, no replacement mode. Think of it as "SCAIL for people who just want one clip."

How it works

SCAIL's trick is that it's an image-animation model wearing a video model's clothes. The node builds an empty latent of the right shape ([batch, 16, (length-1)//4+1, h/8, w/8]), then stuffs two things into the conditioning. First, the pose video: it VAE-encodes it at half resolution, multiplies by pose_strength, and attaches it as pose_video_latent with a timestep range of pose_start to pose_end - so the pose only drives the generation during that window of the sampling schedule. Second, the identity: your reference_image gets VAE-encoded into reference_latents, and a clip_vision_output (from encoding the reference with CLIP-Vision) gets attached too. The sampler then produces frames that follow the pose while keeping the character looking like the reference.

The inputs that matter

• positive / negative - your CONDITIONING from a text encoder.
• vae - your Wan video VAE.
• width / height / length / batch_size - dimensions and frame count. Length defaults to 81, SCAIL's native 4N+1 section size.
• pose_strength - how hard the pose drives the motion (1.0 = follow it closely).
• pose_start / pose_end - the fraction of the sampling schedule where the pose applies.
• vae_tile_size - tiled encode/decode to keep VRAM down.
• reference_image / pose_video / clip_vision_output (optional) - the identity, the motion, and the pre-encoded identity if you've already encoded it elsewhere.

Outputs: positive, negative (conditioning with the pose/identity injected) and latent - wire those into a KSampler, then decode.

Install

Standard XB_BOX install - ComfyUI Manager search XB_ToolBox, or clone into custom_nodes and restart. You also need a SCAIL-capable Wan model in diffusion_models (the Wan 2.1 14B SCAIL fp8 from kijai's WanVideo_comfy_fp8_scaled repo is the usual choice for consumer cards). Unlike the relay nodes this one doesn't need the ParamBus.

Common issues

Because there's no extension logic here, you're capped at one segment per run - that's the tradeoff for the simplicity. If your character drifts from the reference, you skipped clip_vision_output or reference_image. And if the pose barely influences the output, check pose_strength isn't set low and pose_start/pose_end aren't pinching the window shut. For multi-segment or replacement-mode work, this node's big sibling XB_WanSCAILToVideoPro is the one to reach for.