XB-BOX - 🆕 SCAIL 总线 (New)

SCAIL is Wan's pose-driven character animation - you feed it a character reference image and a pose/motion video, and it makes the character perform that motion. XB's SCAIL pipeline turns that into an infinite-length video by chaining relay segments, and this node ("🆕 SCAIL 总线 (New)") is the control panel for the whole thing. Everything the relay nodes need downstream - model, CLIP, VAE, pose video, reference image, all the sampling and VRAM knobs - gets bundled here into one WAN_SCAIL_BUS wire, so your downstream relay nodes only ask for a prompt and a segment length.

How it works

pack_bus() does two jobs. First it encodes your global_ref_image with the clip_vision model, scaled to the video's width/height - that's the identity conditioning that keeps the character looking like the reference across every segment. Then it bundles every input into a dict and stamps a current_offset of 0, which is where in the pose video the first segment starts. Downstream relay nodes read from the bus and advance that offset as they go.

The inputs that matter

There are a lot, but they fall into a few groups:

• model / clip / vae / clip_vision - the whole generation stack, required. clip_vision is the one people forget; without it the reference identity is weak.
• pose_video - the motion you're driving from. total_frames of 0 means "take the length from the pose video"; set it above 0 to force a cut. fps default 16, which is SCAIL's native rate.
• pose_strength / pose_start / pose_end - how hard the pose drives the generation, and over which fraction of each segment (e.g. 0.3–0.8 applies it only in the middle). This is your dial for "follow the pose exactly" vs "interpret it loosely."
• replacement_mode - off is "animation mode" (character performs the pose); on is "replacement mode," where the reference character replaces the subject, and you can isolate it with reference_image_mask so only that region is replaced.
• concat_mode - where the accumulated video lives: auto switches to CPU when VRAM gets tight (over 30% used), or force GPU/CPU.
• VAE tile / temporal knobs - vae_encode_tile_size, vae_decode_tile_size, spatial_overlap, temporal_chunk_size, temporal_overlap. These keep encode/decode from OOMing on long videos.
• steps / cfg / sampler_name / scheduler / seed - the sampling config, applied to every segment.
• cleanup - up to "卸载显存模型" (unload the GPU model) after the run.

Output: a single WAN_SCAIL_BUS socket.

Install

Standard XB_BOX install - ComfyUI Manager search XB_ToolBox, or clone into custom_nodes, restart. Then you need a SCAIL-capable Wan model (Wan 2.1 14B SCAIL; the fp8 scaled version from kijai's WanVideo_comfy_fp8_scaled repo is what most people run on 16GB cards), plus a pose/driving video and a character reference image.

Common issues

The default negative_prompt is a long Chinese quality-negative list - works fine, but you'll probably want your own. If identity drifts between segments, check clip_vision is wired and pose_strength isn't cranked so high the pose overrides the character. And if the chain stops early, your pose video ran out of frames before the relays finished.