Skip to main content
by artokun · June 16, 2026 · wan · video · ComfyUI · model highlight Most “AI video” tools let you make a clip from scratch and then leave you stranded the moment you want to put your character into that performance. WAN Animate 2.2 is the open model that nails the part everyone actually wants: hand it one reference image and a driving video, and it either drops your character into the existing footage — same lighting, same camera, same motion — or animates your still to copy the performer’s pose and expression frame for frame. And you can run the whole thing locally in ComfyUI, no API, no per-second render fee. Below: what it’s genuinely best at (with the paper’s receipts), how it works under the hood, where it sits relative to WAN VACE and the rest of the WAN stack, the VRAM you need, and the fastest way to get it running — a one-command install with comfyui-mcp and the sidebar Panel, instead of hand-cloning six custom-node repos and chasing dtype errors.
TL;DR — one-command setup. Install comfyui-mcp, apply the wan-animate pack (apply_manifest --path packs/wan-animate/manifest.yaml, or run the generated installer), install SageAttention + Triton with the bundled .bat, then drive the graph from your own Claude session via the Panel. Jump to Install.

​
What is WAN Animate 2.2?

WAN Animate (from Alibaba’s WAN team, released September 19, 2025 as Wan2.2-Animate-14B) is a unified character animation and replacement model built on the Wan-I2V video foundation with a Diffusion Transformer (DiT) backbone. One architecture, two jobs:
  • Animation mode — animate a static character image so it copies the body motion, facial expression, and lip movement of a driving video. Upload a photo
    • a performance, get your character performing it.
  • Replacement mode — swap the character already in a video for your reference character, while preserving the original lighting, color tone, and camera so the swap blends into the scene.
Critically, it’s Apache 2.0 licensed — fully permissive, commercial use included. That’s rare in the video-model space and is the main reason it caught on so fast versus the closed Runway/ByteDance options. (The weights and inference code are on Hugging Face.)

​
Why it’s the v2v “character replace” king

This is the headline, and the paper backs it up. In the Wan-Animate paper, the authors position it as “the most comprehensive and highest-performing open-source model to date” for character animation, with:
  • Quantitative wins over open baselines (reported SSIM 0.834 / LPIPS 0.205 / FVD 94.65 on portrait data).
  • Human-preference wins vs. closed-source Runway Act-Two and ByteDance DreamActor-M1 in a 20-participant study.
  • A clean read on the open competition: Animate Anyone “exhibits significantly lower generation quality” (UNet-based foundation), VACE shows “instability in character animation tasks,” and DreamActor-M1 “tends to have slightly lower quality in local details.”
A fair caveat: those benchmarks and the human study are run by the model’s own authors, so treat the “beats everyone” framing as strongest for character animation/replacement specifically, not as a universal video-quality crown. Independent reviewers consistently scope the win the same way — and so do we.

​
WAN Animate vs. the rest of your video stack

There’s no single “best” — pick by the job: Where rivals win: VACE for breadth of edit types, T2V for from-scratch creation, LTX for raw speed. WAN Animate earns the spotlight for the one thing it’s built to dominate — driving a character with another character’s performance.

​
System & VRAM requirements

The pack targets the 14B fp8-scaled build. Plan for 24 GB+ for a smooth 720p experience on a single card: Treat 24 GB as the target floor for fp8 at 720p. On a 24 GB RTX 4090 you’re in good shape; for 16 GB and below, drop to a GGUF quant and lower resolution. (Numbers vary with clip length, resolution, and offload settings — flag this as approximate.)

​
Install WAN Animate 2.2 in ComfyUI

The manual route is real work: clone the WanVideoWrapper, KJNodes, VideoHelperSuite, segment-anything-2, SecNodes, and WanAnimatePreprocess custom-node repos, then download the diffusion model, encoder, VAE, CLIP-vision, two LoRAs, the SeC model, and the ViTPose/YOLO ONNX detection files — each into a different folder.

​
The fast way — comfyui-mcp + the Panel

That folder-juggling is exactly what the comfyui-mcp wan-animate pack removes. One declarative manifest installs every custom node and pulls every model to the correct folder — and the same manifest drives both an MCP-native install and the generated double-click scripts:
Required: SageAttention + Triton. The WanVideoWrapper sampler needs them. Run the bundled install_triton_and_sageattention_auto.bat from your portable Python folder (...\ComfyUI_windows_portable\python_embeded\) before your first generate. It detects your Torch/CUDA, installs triton-windows and the matching woct0rdho SageAttention wheel, and drops the Python include/libs zip so Triton can compile. Restart ComfyUI afterward, then load workflow.json.
Because the pack ships with the plugin, your own Claude session can drive the live graph through the Panel — load the reference image, point at the driving video, pick the SeC vs. auto-mask path, set widgets, and iterate conversationally, with full Ctrl+Z undo and no extra API keys. Every model URL in the pack is CI-validated for reachability and size, so a link never quietly rots (here’s why that matters).

​
How WAN Animate works

The pack wires up the full kijai pipeline, but it helps to know what each piece does:
  • Pose / detection (ViTPose + YOLO). A YOLOv10 detector finds the person, then ViTPose-L (or -H) WholeBody extracts 2D whole-body skeleton keypoints per frame. Those skeleton signals are spatially aligned to the character and used to drive body motion. (Runs via ComfyUI-WanAnimatePreprocess.)
  • Face / expression. Rather than abstract landmarks, the model feeds the raw face crop directly into dedicated “Face Blocks” in the DiT — that’s why expression and lip-sync come through so cleanly.
  • Segmentation masking (SeC). Replacement mode needs a precise mask of the character to remove. The pack defaults to SeC (Segment Concept, SeC-4B), a concept-driven video segmenter that uses a vision-language model for robust tracking through scene changes — it reports an 11.8-point gain over SAM 2.1 on the SeCVOS benchmark (SeC paper). A SAM2 auto-mask path is bundled but muted by default.
  • Relight LoRA. An auxiliary LoRA (trained on IC-Light-synthesized lighting pairs) adjusts the new character’s lighting and color tone to match the destination scene — the difference between a “pasted-in” swap and a believable one.
  • lightx2v speed LoRA. A step-distill LoRA (lightx2v_I2V_14B_480p_cfg_step_distill) that collapses sampling to a handful of steps for dramatically faster generation.
The diffusion model itself is the 14B fp8-scaled WAN Animate checkpoint (Wan2_2-Animate-14B_fp8_scaled_e4m3fn_KJ_v2), paired with the umT5-XXL text encoder, the Wan 2.1 VAE, and CLIP-vision — all pulled to the right folders by the pack.

​
Where WAN VACE fits

You’ll hear WAN VACE in the same breath as Animate, so here’s the clean split. VACE (“All-in-One Video Creation and Editing,” ali-vilab/VACE) is the generalist — one framework for reference-to-video, video-to-video editing, masked editing, and free composition of those tasks (Move-Anything, Swap-Anything, Animate-Anything, Expand-Anything, etc.). It’s the Swiss-army knife. WAN Animate is the specialist. It does one family of jobs — drive/replace a character — and does it better than VACE does that specific job. The Wan-Animate paper explicitly calls out VACE’s instability on character tasks. So:
  • Reach for Animate when the job is “put this character into this performance” or “make this still copy this motion.” Best identity, expression, and scene-blending.
  • Reach for VACE when you need broad, mix-and-match video editing (inpaint, outpaint, extend, generic reference control) and character fidelity isn’t the whole point.
If you’re already running our WAN stack, note that the WAN 2.2 T2V pack also ships VACE modules for reference/pose/depth conditioning on text-to-video — see the WAN 2.2 post for that side.

​
What to make with it

  • Character / outfit / object replace — drop your hero, your VTuber, or your client’s mascot into existing footage with the original lighting preserved.
  • Motion transfer to a still — animate a single portrait or concept-art character to copy a real performance (dance, dialogue, gestures).
  • Virtual influencers & avatars — consistent identity driven by an actor.
  • Lip-sync dialogue — the raw-face-crop conditioning makes mouth movement track the driving clip.
  • Previz & stand-in swaps — block a scene with one actor, swap the character in post.

​
Troubleshooting

  • self and mat2 must have the same dtype (WanVideoWrapper). The known fix: delete and re-clone kijai/ComfyUI-WanVideoWrapper, then reinstall its requirements with the portable Python — python_embeded\python.exe -m pip install -r ComfyUI-WanVideoWrapper\requirements.txt.
  • RTX 5000-series: onnxruntime “QuickGelu” / CUDA-provider error on the ViTPose detection pass. Pin the version: pip install onnxruntime==1.20.1.
  • Torch not compiled with CUDA enabled. Your torch is a CPU build. Reinstall from the matching CUDA index, e.g. pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128 (use the cuXXX matching your driver).
  • Triton/SageAttention errors at sampling. Re-run install_triton_and_sageattention_auto.bat from python_embeded\ and restart.
  • Masking is wrong / character not isolated. The shipped workflow defaults to SeC (manual/prompted); the SAM2 auto path is muted. Pick the matching mask / background_image getter per the in-graph notes.
  • Missing nodes after install. Restart ComfyUI, or use ComfyUI-Manager → Install missing custom nodes to pull any node Python deps.

​
FAQ

What does WAN Animate 2.2 actually do? Two things from one model: animate a static character image to copy a driving video’s motion/expression, or replace the character in a video with your reference character while keeping the original lighting and camera. Do I need a reference video? Yes — Animate is video-to-video. You bring a reference image (the character) and a driving video (the performance/footage). For text-to-video from scratch, use the WAN 2.2 T2V stack instead. Is it free / commercial-use OK? Yes. The weights are Apache 2.0 — the most permissive license in open video, commercial use included. How much VRAM do I need? ~24 GB for the fp8 build at 720p; GGUF quants bring it to roughly 8–12 GB at 480p with more system RAM and slower runs. WAN Animate vs. WAN VACE — which one? Animate for character drive/replace (better identity and blending); VACE for general all-in-one video editing. The Animate paper notes VACE is unstable on character-specific tasks. Why does it need SageAttention and Triton? The WanVideoWrapper sampler depends on them. The pack ships an auto-installer that matches your Torch/CUDA — run it once before your first generation. Is it better than Runway Act-Two or DreamActor-M1? In the authors’ human study, yes — users preferred WAN Animate. That study is vendor-run, so read it as “competitive with / preferred over the closed tools for character animation,” not a universal benchmark.

​
Get it running in one command

  1. Install comfyui-mcp and the Panel — the panel auto-starts a background agent on your Claude subscription (no API keys; sign in with claude once).
  2. Apply the wan-animate pack — nodes + models land in the right folders, validated.
  3. Run install_triton_and_sageattention_auto.bat, restart, and load workflow.json.
  4. Open the Panel and let the panel’s agent load the image, point at the driving video, and iterate.
That’s the whole point of the project: expert ComfyUI setups that install in one step and drive themselves from your own agent session. Next in the series: WAN Transparent Expressions — looping, alpha-channel character sprites from WAN 2.2 I2V.