Wan 2.2 Animate: ComfyUI Setup, Files and Animate 2 (2026)

Wan 2.2 Animate is an open-weight model from Alibaba's Wan team that takes a character image and a driving video, then either animates the character with the video's movement and expressions or replaces the person in the video with the character. It is licensed under Apache 2.0 and runs in ComfyUI from the built-in Wan2.2 Animate template, which needs five model files and two custom node packs.

Read on 13 October 2026 from the Wan team's Wan2.2 repository, its preprocessing guide and the model card, Comfy's Wan2.2 Animate and Wan Animate 2 tutorials and changelog, the Wan-Animate-2 repository, and the Hugging Face pages for each file.

Quick answer

What is Wan 2.2 Animate?

Wan is Alibaba's family of open video models, published on GitHub as Wan-Video and on Hugging Face as Wan-AI. The repository's news log dates the Wan 2.2 release to 28 July 2025 and adds Wan2.2-Animate-14B on 19 September 2025, calling it a "unified model for character animation and replacement with holistic movement and expression replication". Animate is not a text-to-video model. It needs a performance to copy.

The mode names differ by source: the Wan team says animation and replacement, ComfyUI says Move and Mix. Comfy describes Mix as preserving the performer's expressions and movements "while replicating the original lighting and color tone".

Wan 2.2 in ComfyUI: which model does what

A search for Wan 2.2 in ComfyUI usually means one of five models. Each has an official template, and none needs a wrapper to run:

Sources: the Wan2.2 model list and Comfy's Wan 2.2 and S2V guides. For prompted clips from the 5B or 14B models, start with our best ComfyUI models guide.

ComfyUI-WanVideoWrapper is a separate community node pack by kijai, under Apache 2.0. Its own README answers whether you need it: "Unless it's a model/feature not available yet on native, you shouldn't." Animate has a native template, so start there.

Wan 2.2 Animate, Wan Animate 2, or Wan 2.5 and 3.0?

On 13 October 2026 the Wan-AI organization on Hugging Face listed 27 models, and none is a Wan 2.5, 2.6, 2.7 or 3.0 checkpoint. In ComfyUI those versions are Partner Nodes: the pricing page bills wan2.5, wan2.6, wan2.7 and wan3.0 models in credits per second. So a search for Wan 2.5 in ComfyUI leads to a paid API node, not a download. Alibaba's Wan 3.0 announcement of 13 August 2026 lists API prices from $0.05 to $0.20 a second by resolution. We track the weights question in Wan 3.0 and open weights.

Two alternatives do the same job. SCAIL-2, built on Wan 2.1, handles animation and replacement with native ComfyUI nodes. The Kling 2.6 Motion Control Partner Node is paid and supports continuous actions of up to 30 seconds (Comfy's guide). Kling's own prices are in our Kling AI guide and in Kling vs Runway.

What you need to run it

Software and custom nodes

Comfy's guide expects an up-to-date ComfyUI and names two custom node packs: ComfyUI-KJNodes (GPL-3.0), which supplies the Points Editor, and comfyui_controlnet_aux (Apache 2.0), whose DWPose Estimator turns the driving video into pose and face control videos. With ComfyUI-Manager installed, the Install missing nodes button adds both. If Manager is not showing, follow our ComfyUI Manager install guide.

Model files

That is about 27 GB. File names follow Comfy's guide and sizes come from each file's Hugging Face page (the FP8 model, Comfy's repackaged files). A full-precision 34.5 GB model is optional. The LightX2V file is a LoRA for 4-step sampling.

GPU memory

Neither the Wan team nor Comfy publishes a minimum for this workflow. What they do say:

• The Wan team's single-GPU commands for the 14B text, image and speech models carry the note "at least 80GB VRAM", and the 5B command says at least 24GB. The Animate command has no stated figure.
• The FP8 Animate model alone is an 18.4 GB file. ComfyUI's startup reference says async weight offloading is on by default on Nvidia and AMD cards, and Comfy says the 5B model "should fit well on 8GB vram" with native offloading. A card smaller than the model can still run it, more slowly.
• Comfy's advice for a first run is "a small size for video generation, in case you don't have enough VRAM".

Without a suitable card, the template opens on Comfy Cloud, or you can rent a GPU: see our ComfyUI cloud GPU comparison.

Step by step: Wan 2.2 Animate in ComfyUI

The prompt is optional: Comfy's step reads "Update the prompt if you want". The template ships in Mix mode. To switch to Move, disconnect the background_video and character_mask inputs of the Video Sampling and output subgraph node.

How long can a Wan 2.2 Animate video be?

There is no fixed maximum. The Wan team's Animate config sets 77 frames per segment, and Comfy says iterative generation keeps motion consistent in extended videos. In ComfyUI the length is the number of Video Extend groups you chain. Comfy counts each 77-frame extension as about 4.8 seconds and says an input under 5 seconds may not need one. Every extra segment adds render time, so check the joins before you commit to a long run. The team's preprocessing guide adds that a lower frame rate improves efficiency but "may cause stuttering or choppiness".

Wan Animate 2: what changed

Wan Animate 2 (Wan2.2-Animate-2-14B) was released on 7 August 2026 under Apache 2.0. It "directly consumes driving videos", so there is no pose or skeleton preprocessing, and it adds "text-driven viewpoint control to decouple the output camera perspective from the driving video". Comfy's tutorial adds that it generates a new background from your prompt instead of copying the driving video's.

• Two weight sets: the base model runs 40 steps in the team's example. The distilled model runs 10 steps with a guidance scale of 1.0.
• ComfyUI: native since v0.31.0 on 7 August 2026, with WanAnimate2ToVideo and WanAnimate2Cache nodes. The template loads wan_animate_2_int8_convrot.safetensors, a 16.7 GB file, with the same LoRA, text encoder and CLIP Vision file as above.
• Hardware: the team's scripts are "tuned for 8× A800 GPUs" at 720P and were tested at 480P on two, so the quantized ComfyUI template is the practical route on one card.
• Prompt: the team says to caption the reference image first, describing the character's appearance and the background but no actions, and use that caption as the prompt.

Comfy's Animate 2 template covers animation only. To swap a person inside existing footage, stay with Wan 2.2 Animate's Mix mode or use SCAIL-2.

Choosing inputs, and a note on LoRAs

• Match body proportions. In replacement mode the team warns that a significant mismatch between the two characters can cause "artifacts or deformations".
• One performer for Mix. Multi-person footage can break pose tracking and the mask.
• Leave Wan 2.2 LoRAs out. The repository says: "we do not recommend using LoRA models trained on Wan2.2, since weight changes during training may lead to unexpected behavior." The LightX2V LoRA in the template is part of Comfy's setup.

That answers most questions about Wan 2.2 LoRA training for this model: train character LoRAs for the stills, not for Animate. For the other Wan 2.2 models, the repository points to DiffSynth-Studio for LoRA training and musubi-tuner lists Wan 2.1 and 2.2 among its architectures. The method is in our LoRA training guide.

Troubleshooting

• Nodes are missing or red: update ComfyUI, install both node packs, restart, and check the startup log for import failures.
• CLIP Vision model not found: the folder ComfyUI reads is models/clip_vision (folder_paths.py). The file tree in Comfy's Animate guide spells it clip_visions.
• Out of memory: lower the size in multiples of 16, shorten the driving video, remove Video Extend groups and keep the FP8 model.
• Jumps between segments: each Video Extend group must take its inputs from the group before it, not from the first one.

License, likeness and disclosure

The Wan2.2 license section says the models are "licensed under the Apache 2.0 License" and that the team claims no rights over generated content, while you are "fully accountable" for your use. Apache 2.0 allows commercial use.

Label realistic AI video when you post it. Each platform sets its own rule, and Instagram's is in our Instagram AI label rules guide. The AI influencers library covers the persona and consistency steps that come before motion, and our free AI video generator comparison sets self-hosted Wan beside the hosted free tiers.

Wan 2.2 Animate: FAQ