WanAnimatePlus SCAIL_2 Embeds

SCAIL-2 is Z.ai's skeleton-free motion transfer model on a Wan 2.1 base - the one that dropped the stick figure entirely, which is why it can do animals, multi-character and cross-identity replacement. It ships in ComfyUI core via Kijai's PR, but core gives you a bare-bones path. This node is the fork's wrapper-native version: it prepares the reference image, driving pose, colored masks, and optional prefix/transition context into one image_embeds bundle that the WanAnimatePlus Sampler consumes - and it layers on the pack's loop-seam and prefix machinery that the stock graph doesn't have.

The core conditioning. ref_image is your character (only the first frame is used if you feed a sequence). pose_images is the driving video, encoded at half resolution for the SCAIL pose stream, with pose_strength as your control multiplier. pose_image_mask and reference_image_mask are the colored per-identity masks - and replacement_mode is the master switch: off means animation mode (pose mask on black background, reference mask on white), on means replacement mode (backgrounds flipped). Get this wrong and the model composites the wrong way. ref_strength scales the reference stream.

The modes that took work. single_frame_prefix_encoding (on by default) encodes prefix_frames as individual full-resolution reference latents instead of expanding the canvas - the modern path, and it doesn't stretch your output. Turn it off and you get the legacy 37-frame front-canvas layout where prefixes hard-freeze the start of the generation. bg_image occupies one reference slot in animation mode (ignored in replacement mode), and preserve_main_ref_background / prefix_alpha_crop control whether backgrounds are kept or alpha-cropped from the masks. transition_video hard-freezes frames at the front for seamless joins, and transition_colormatch offers auto_drift - a light seam fix that compares RGB means between the last 5 frames of one segment and the first 5 of the next, plus the full mkl/reinhard/hm family. loop_colormatch_reference picks what later chunks match against for loop mode.

The window knob. frame_window_size (default 81) is the SCAIL chunk length, auto-rounded to 4n+1. Setting it differently from num_frames enables built-in loop generation with a 5-frame handoff - this is the pack's answer to the manual 81-frame chaining that community threads spent June 2026 automating. The WanAnimatePlus Easy Sampler is the recommended companion, since it keeps the full feature path while exposing only common controls.

The honest grounding. SCAIL-2's strengths and limits come from the model, not this node: proportions survive motion transfer, but faces wander (drop LightX2V LoRAs, raise steps/CFG), it's compute-hungry (minutes per clip), and identity drift across chunks is the persistent failure mode. This node bundles the conditioning so you fight fewer moving parts. Install via ComfyUI Manager (search "WanAnimatePlus") or git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus into custom_nodes, restart, and keep the chain WanAnimatePlus end to end - including the fork's sampler, since that's where this embeds object is consumed.