WanAnimate To Video (自定义)

WanAnimate is Alibaba's open motion-transfer model: you hand it a reference image of a character plus a pose/face driving video, and it re-renders the character performing that motion. The fiddly part has always been the preprocessing - the reference image has to be tiled 1+4n into latent frames, the pose video has to become a pose-conditioning latent, the face video gets its own channel, the character mask, the tail frames for continuation... on a stock workflow that's a half-dozen nodes wired by hand, and one wrong shape silently wrecks the output.

WanAnimateToVideoCustom collapses all of that into one node. You feed it the raw ingredients - reference image, pose video, face video, background, mask - and it encodes, arranges, and concatenates everything into the positive/negative/latent triple that goes straight into a KSampler. It's the node that makes the pack's "WanAnimate_ref_pics_loop+context" workflow readable at all.

The inputs that matter

You don't need to touch most of these to get a first result. Start with:

• positive / negative - your CLIP-conditioned prompts (required).
• vae - the Wan VAE (required).
• width / height (default 832×480, multiples of 16) and length (default 77, multiple of 4) - output geometry. 77 pixel frames is the classic WanAnimate chunk.
• reference_image - the character reference batch. Feed it the selected_images output from the pack's Reference Image Selector. The node handles the 1+4n tiling internally (first image ×1, each extra ×4).
• pose_video / face_video - the driving inputs. Both optional, both encoded into the right latent slots.
• ref_mode - 原模式 (original) is the default 1+4n batch encode; 兼容模式 (compatible) encodes each reference frame independently, which is what EverAnimate-trained LoRAs expect.
• mode - the continuation strategy for long videos: vanilla is stock behavior, legacy does a tail-frame fade mask, fix does black-frame detection with a transition zone (transition_width).
• continue_motion or prev_latent - how the next chunk starts from the last one. Feed the previous chunk's tail frames as continue_motion, or (better) feed its full concat_latent into prev_latent, which ignores continue_motion and replaces the neutral-gray frames directly.
• video_frame_offset - the chain counter; wire the previous chunk's video_frame_offset output into the next chunk's input.

Outputs

positive, negative, latent go to the KSampler. trim_latent and trim_image tell you how many leading frames are reference anchors you should cut when decoding. concat_latent is the fully-assembled latent (feed it to a VAE decode for output), and video_frame_offset plus latent_yaw_angles chain into the next chunk.

Reality check

This is a third-party wrapper around WanAnimate, and Wan 2.2 is the last open Wan - the interesting ecosystem now lives in derivatives like SCAIL-2, which this same pack also supports. The honest take: if your motion-transfer target is a different character or you need proportions preserved, SCAIL-2 is the better base; WanAnimate is where you land for a recognizable existing character and where the pack's reference-selector/yaw machinery really shines. It also expects a hefty supporting stack - the README's workflows pull in sdpose-ood, ComfyUI-WanAnimatePreprocess, VideoHelperSuite and friends, so installing this node alone won't run the full graph.

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. No model files ship with the pack - you need the WanAnimate model weights plus your pose/face detection stack separately.