Wan SCAIL-2 Phase Sampler

SCAIL-2 is Z.ai's motion-transfer model on a Wan 2.1 base, and it's expensive in the way that matters: several minutes per 81-frame clip on a mid-range card, and past 81 frames you have to chain chunks by hand. The stock workflow's chain is sampler → decode → re-anchor → sampler again, with the arithmetic of when to step and how many frames to carry. This node is the pack's attempt to make that whole dance a single node: a complete two-phase sampler that takes pixel frames as its anchor input and internally does the VAE encoding, the anchor write-back, and the noise-mask management.

Think of it as a self-contained KSampler you can point at the same conditioning and latent your chunk node produces. When it's just a single chunk, it behaves like a normal sampler. When you connect the previous chunk's decoded tail frames, it switches into the two-phase SCAIL-2 continuation mode that the model was actually trained with - 81-frame chunks with a 76-frame step and a 5-frame anchor.

How the two phases work

The mechanism, from the source: it takes the tail previous_frame_count frames of previous_frames, upscales them to your target resolution, VAE-encodes them, and writes them into the front of the latent as an anchor. It builds a noise mask where the anchor region is set to phase1_noise and everything else is 1 (free). Phase 1 samples from your noise down to split_step. Then it force-overwrites the anchor region with the original encoded frames - this is what pins the chunk to the previous video - updates the mask to phase2_noise, and samples phase 2 from empty noise to the end. phase1_noise around 0.5 means the anchor is partially free during composition; phase2_noise around 0.1 means it's mostly frozen at the end.

The inputs that matter

• model1 - the Phase 1 model; model2 (optional) - a different model for Phase 2 (e.g. a distilled speed variant), falling back to model1 if not provided.
• positive / negative / vae / latent - your conditioning and the latent from the SCAIL conditioning node (WanSCAILToVideoMultiRef).
• steps / cfg / sampler_name / scheduler / denoise / seed / width / height - standard sampler knobs. Width and height step by 32 (SCAIL is coarser than WanAnimate's 16).
• split_step (default 2) - where the phases split. 0 = single phase. Two-phase only activates when split_step > 0, it's less than steps, and previous_frames is connected.
• phase1_noise / phase2_noise - the anchor freeze values described above (0 = frozen, 1 = free).
• previous_frame_count (default 5) - how many tail frames of previous_frames become the anchor. SCAIL-2 was trained at 5.
• previous_frames - the previous chunk's decoded frames. Without it, the node quietly degrades to single-phase.

The one output is latent - the sampled result, ready for VAE decode or to become the next chunk's previous_frames.

Reality check

This doesn't make SCAIL-2 cheap - nothing does. What it removes is the workflow cost: no manual sampler chaining, no off-by-one frame arithmetic, and the anchor behavior matches the model's training setup. Set previous_frame_count to 5, keep length at 81, and let the loop advance. And remember the community's standing advice if faces wander across chunks: drop the speed LoRAs and raise steps and CFG before you blame the sampler.

Install

Part of ComfyUI-CustomNodeKit. ComfyUI Manager → search "ComfyUI-CustomNodeKit", or:

cd ComfyUI/custom_nodes
git clone https://github.com/user2318/ComfyUI-CustomNodeKit.git
cd ComfyUI-CustomNodeKit
pip install -r requirements.txt

Restart ComfyUI. It drives your existing SCAIL-2 model - no extra weights from the pack.