WanAnimate2 To Video (自定义)

What it is, and why it exists

ComfyUI ships WanAnimate2ToVideo as a core node, and it works: it takes the Wan 2.2 Animate model (character animation and replacement, built on the Wan 2.1 I2V base) and injects the pose branch as key/value rather than piling everything into one concat. Two things are missing. It won't take a batch of reference images - the six-angles-of-your-face trick first-gen WanAnimate users rely on - and the extension path has a reputation: the classic report is that chunk two replays the first chunk with a slow zoom (1nwrlgx, "WanAnimate Comfy native does not extend").

This pack's node is a fusion rather than a wrapper. You get the second-gen mechanism (the pose branch, plus pose_strength, the pose time window, reference_image_strength, a pose-specific prompt and CLIP vision) and the first-gen node's habits on top: reference images encoded one at a time, the fix/legacy/vanilla mask modes, and a concat_latent output a context-window node can chew on. Wan 2.2 is the last open Wan and SCAIL-2 now owns the motion-transfer default, so you're here for the pose path specifically.

How it actually works

The node VAE-encodes each reference image separately - one image, one latent frame - and concatenates them on the time axis. That per-image encoding is deliberate: it's what keeps EverAnimate-style LoRAs happy, and the old "1+4n batch encode" was dropped because it behaved badly here. Then it builds the image zone: a full-length grey placeholder, filled at the head with your continuation frames if you supplied any. The two blocks become the concat latent the model consumes, and a single-channel concat mask marks the reference/continuation region as known (0) and the rest as to-generate (1).

The pose branch is where the interesting fix lives. Your pose_video gets upscaled to the output resolution, VAE-encoded, then zero-padded at the front by trim_latent - 1 frames, because the second-gen model asserts pose_latents.shape[2] == f_gen - 1. Without that padding, multi-reference runs fail the check.

The inputs you'll actually touch

• reference_image - the raw batch (e.g. selected_images from the pack's Reference Image Selector). Do not pre-splice the 1+4n pattern yourself; this node encodes per image on its own.
• pose_video - your skeleton render as an IMAGE sequence. Sliced from video_frame_offset onward, padded by repeating the last frame if it runs short.
• length (default 81, step 4) and width/height (multiples of 16). Shave length first if you're tight on VRAM.
• pose_strength (1.0 = trained behaviour) and the pose time window, pose_start_percent / pose_end_percent. Motion is largely established early, so pose_end_percent = 0.7 loosens fine detail while keeping the choreography - and it's faster, because the pose branch is skipped outside the window.
• reference_image_strength - below 1 loosens identity adherence, above 1 tightens it against drift. Reach for this before you blame the LoRA.
• For chunked runs: continue_motion, continue_motion_max_frames (5 is sane) and video_frame_offset from the previous chunk's output of the same name.

mode defaults to vanilla; fix and legacy (plus tail_frame_count, the tail strengths, mid_frame, neutral_mix_min/max) handle black frames carried over from a previous chunk. Leave them alone until you see black frames in your continuation input.

Outputs

positive and negative carry the injected concat latents into your KSampler, and latent is the empty latent shaped to the total frame count. concat_latent is the full strip - decode that one. trim_latent / trim_image tell downstream nodes how many head frames are reference/continuation rather than generated output, so use them instead of eyeballing a crop. video_frame_offset comes back incremented by length: wire it into the next chunk, or every chunk re-predicts from frame one.

Installing it

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

Restart ComfyUI, or search "ComfyUI-CustomNodeKit" in ComfyUI Manager. That's the whole install for this node - the pack's install.py also pip-installs groundingdino-py and transformers for its unrelated detection nodes on first import. If that part fails, the pack still loads.

No models ship with it: you still need the Wan 2.2 Animate UNet, the Wan VAE, CLIP vision and a text encoder.

Where people get burned

pose_start_percent > pose_end_percent isn't a sampler crash, it's an immediate ValueError from the node. It reads like a framework bug the first time.

If a chunk comes out looking like the previous one with a slow zoom, check the wiring first. video_frame_offset has to actually reach this node, and pose_video has to be long enough to slice from that offset - the node drops everything before it, then pads by repeating the final frame, so a stale offset gives you a plausible-looking run with the wrong skeleton throughout. That's the honest answer to most "the extend nodes don't extend" complaints against the native workflow.

One upgrade gotcha: v1.8.0 removed WanSCAILSparseAttention outright - O(S²) attention masks that exploded memory on long sequences for no measurable benefit. Old workflows will just show a missing node; delete the orphan.