WanSCAILToVideoMultiRef

SCAIL-2's whole pitch is that it dropped the stick-figure pose representation and takes raw inputs instead: a reference image, a driving video, and support masks. That's more flexible - cross-identity replacement, animal driving, multi-character - but it also means a lot of conditioning engineering before the sampler ever runs. WanSCAILToVideoMultiRef is that engineering in one node: the conditioning hub for this pack's SCAIL/SCAIL-2 support. You throw in reference images, a pose video, colored per-identity masks, and it returns the positive/negative/latent triple plus a frame-offset for chaining chunks.

It's the SCAIL sibling of WanAnimateToVideoCustom, and it does the multi-reference thing properly: every reference image is independently encoded into its own reference_latents (SCAIL-2 trains with reference images concatenated into the latent, so more references means more identity anchors), and the pose video is downscaled to half resolution before encoding.

The inputs that matter

• positive / negative / vae - conditioning and the VAE (required).
• width / height (default 512×896, step 32) and length (default 81) - output geometry. 81 is the SCAIL native context; keep it there for single chunks.
• reference_image - a batch of images (N, H, W, 3). Each is encoded independently and added as a separate reference latent.
• ref_encoding_mode - 1+4n批量 (first ×1, rest ×4, one batch encode - closest to training, the default), 逐帧编码 (each frame independently encoded, higher per-frame fidelity), or 混合编码 (first N-1 batch + last frame per-frame - a balance that eases chunk-to-chunk jumps).
• pose_video - the driving video for pose conditioning.
• pose_video_mask - SCAIL-2 only: a SAM3 colored per-identity mask video at the same resolution as the pose video. This is what makes multi-person work.
• replacement_mode - False (default) is Animation Mode, where the pose mask has a black background; True is Replacement Mode with a white background. The mask background color must match this setting or you get garbage.
• reference_image_mask - SCAIL-2 only: colored masks matching the reference images, same count.
• previous_frames + previous_frame_count (default 5) - the anchoring pair for chunk continuation: feed the previous chunk's decoded output; only the last previous_frame_count frames are used as the extension anchor. 5 is what SCAIL-2 was trained with (81-frame chunks, 76-frame step).
• video_frame_offset - the chunk counter; wire the previous chunk's offset output into this input.
• pose_strength (default 1), pose_start / pose_end - how hard and when the pose conditioning applies during denoising.

Outputs

positive / negative (the assembled conditioning, including the 28-channel mask fields), latent (the empty generation-size latent), and video_frame_offset (your offset plus length, for the next chunk). From here the flow is: this node → WanSCAILContextWindows (wraps the model) → WanSCAIL2PhaseSampler or a KSampler → VAE decode → next chunk's previous_frames.

Where people get burned

The mask/background contract is the #1 trap: Animation Mode wants black-background pose masks and produces white reference-mask backgrounds; Replacement Mode flips both. The pack's Create SCAIL-2 Colored Mask node handles this pairing automatically, so use it rather than hand-building masks. Also, multi-reference isn't free - each extra reference adds latent frames and VRAM, and the compute is already the wall with this model. And remember the community's verdict on quality: outputs are genuinely impressive, faces can wander, and 20 seconds of video costs tens of minutes of generation.

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. You supply the SCAIL-2 weights and the SAM3 tracking yourself.