Wan 2.2 Animate: How to Animate a Character Photo for Free (Step-by-Step Guide)

Introduction

Before Wan 2.2 Animate, animating a character photo meant writing prompts about motion and hoping. You would type "a woman waving slowly while turning her head" into an image-to-video model and get back a half-right result — the wave happens, the head turn is a suggestion, and the face belongs to someone else by frame 40.

Wan 2.2 Animate changes the input, not the prompt. Instead of describing motion in words, you hand the model two files: a reference image of your character and a driving video of a real performance. The model transfers the body motion, facial expression, and lip movement from the video onto your character, frame for frame. You do not describe the motion, because you supply it.

This guide covers how the model actually works, the difference between its two operating modes, how to prepare inputs so the output does not fall apart, and how to run it free — in a browser tool or locally in ComfyUI.

TL;DR

• Wan 2.2 Animate (Wan2.2-Animate-14B) is Alibaba's open character animation and replacement model, released September 19, 2025 under Apache 2.0 and built on the Wan-I2V video foundation.
• It has two modes: Move (animation) transfers motion from the driving video onto your character image; Mix (replacement) swaps your character into the existing footage while keeping the scene, lighting, and camera intact.
• Inputs are a character reference image plus a driving video. No motion prompting — the performance is the prompt.
• Output goes back to free if you run it locally (ComfyUI or the official CLI, Apache 2.0 weights) or through a browser tool with free credits if you do not own a capable GPU.
• Four preparation rules decide the result: 720p or higher source, locked-off camera, neutral background, matched framing between reference and driving video.
• Expect it to fail on camera pans, zooms, and multi-person footage. The model cannot separate camera motion from subject motion, and it locks onto the largest face in frame.
• The practical sweet spot is a 4–5 second clip. Longer sequences are built by extending, not by generating in one pass.

What Wan 2.2 Animate Is

Wan 2.2 Animate is the motion-transfer member of Alibaba's Wan 2.2 family. Where the base Wan 2.2 text-to-video and image-to-video models generate a scene from a prompt, Animate takes a character image and a performer's video and returns the character performing that performance.

Technically, it is a unified character animation and replacement framework built on the Wan-I2V foundation with a Diffusion Transformer (DiT) backbone. It was released as Wan2.2-Animate-14B on September 19, 2025 under the Apache 2.0 license, and the weights are downloadable, which is why the local workflows exist at all.

The job it does is the same job Runway Act-One, Hedra's character tools, and Microsoft's research models do — except open-weight, so the free path is real rather than a free trial.

Two Modes, Two Different Jobs

The modes are named differently depending on where you run the model — ComfyUI calls them Move and Mix, while the official docs call them animation and replacement. Same pipeline, different output:

Mode	ComfyUI name	What you get	Use it for
Animation	Move	Your character image performs the motion, expression, and lip movement from the driving video	Talking characters, dance clips, performance transfer
Replacement	Mix	Your character replaces the actor inside the original footage, with the scene's lighting and colour preserved	Stunt or double replacement, swapping a person in existing footage

The underlying pipeline is shared: skeleton extraction, implicit facial feature extraction, and your character image as visual reference. The difference is what stays and what gets replaced. In Move mode you disconnect the background and character mask inputs; in Mix mode you keep everything connected so the original footage drives the scene.

If you want output today, start here: Launch Wan 2.2 Now →

What You Need Before You Start

Two files and a decision.

1. A character reference image

• Single person, clearly visible, unoccluded
• Neutral-ish background is easier for the model to separate
• Same framing as the driving video (face shot drives face shot, half-body drives half-body, full-body drives full-body) — mismatched framing is the single most common cause of uncanny crops
• 720p or higher; lower-resolution sources amplify artefacts when the model downsamples internally

2. A driving video

• Locked-off camera. Handheld jitter transfers to your character and there is nothing you can do about it afterwards
• One person in frame. Wan Animate locks onto the largest face; a second person creates chaos
• No camera pans or zooms. The model cannot separate camera motion from subject motion
• 4–5 seconds is the practical sweet spot for a first run
• Green screen or a plain wall if you plan to composite later

3. A route to run it

Route	Cost	Requirements	Best for
Browser tool	Free credits to start	Nothing — works from the upload page	Testing an idea, no GPU, quick client previews
Local ComfyUI	Free (weights are Apache 2.0)	A capable GPU and some patience with node installs	Volume work, full parameter control, no per-clip cost
Official CLI	Free	Python environment, model downloads	Scripted/batch pipelines

Running It Free in the Browser

If your goal is to see whether your footage works before investing an evening in a local install, the browser route is the faster test:

1. Open the Wan 2.2 Animate generator and start a generation.
2. Upload your character reference image.
3. Upload the driving video — 4–5 seconds, locked-off camera, one person.
4. Choose your mode: animate (Move) to transfer the performance, replace (Mix) to swap your character into the footage.
5. Generate and watch the first two seconds closely. If the framing is wrong there, it will be wrong at frame 100 too.

That last step matters more than it sounds. Almost every failure I see at the end was visible in the first second of output — mismatched framing, a changing face, a copied camera wobble.

For related free tools on the same site, image-to-video with Wan models lives at the free image-to-video generator, and the Wan 2.2 model page is Wan 2.2 free generator if you want to compare plain generation against motion transfer.

Running It Locally in ComfyUI

The local route is more work but free at scale, and it gives you control over the parameters that decide output quality.

Install path

1. Install the Wan 2.2 Animate nodes via ComfyUI Manager. The two common options are the WanVideoWrapper custom node pack (the most actively maintained port, which includes Animate) and native Wan 2.2 nodes once your ComfyUI build supports them. Restart after installing; the nodes to look for are the Wan video loader, sampler, and animate conditioning nodes.
2. Download the model weights. The official repository ships the Animate-14B weights, plus the text encoder, VAE, CLIP-vision, LoRAs, the SeC model, and the pose/detection ONNX files — each goes into a different folder, which is the part people underestimate.
3. For the native workflow you also need a pose estimation node pack for skeleton preprocessing (DWPose), and it is worth adding a step-acceleration LoRA if you plan to iterate a lot.
4. Load a workflow: either the complete community workflow with custom nodes, or the core-node workflow if you prefer to preprocess the image yourself first.

Two constraints to remember

• Output width and height must be multiples of 16. Non-conforming sizes are a common source of errors that look like model failures.
• Video Extend adds 77 frames per extension (about 4.8 seconds at 16 fps). You build longer sequences by chaining extends and carrying the frame offset forward, not by asking for a longer clip in one pass.

Switching modes in ComfyUI

Ready to try it yourself? Try Wan 2.2 Free →

To use replacement (Mix) mode, keep every connection intact so the background video and character mask feed the sampler. To use animation (Move) mode, disconnect the background video and character mask inputs from the output subgraph — that is the entire difference between the two behaviours.

A Workflow That Produces Usable Output

Step 1 — Match the framing before anything else. Face shot with face shot, half-body with half-body, full-body with full-body. If your driving video is a waist-up shot, do not use a full-body reference image.

Step 2 — Record the performance the way you want it delivered. Slow, deliberate movement transfers better than fast motion, because fast motion is where facial artefacts appear. Keep the camera still — tripod, phone propped up, whatever gets you a stable frame.

Step 3 — Generate a short test first. Four seconds is enough to judge framing, face fidelity, and whether the background behaves. Fix problems at four seconds, not at forty.

Step 4 — Extend, do not regenerate. Once the short clip is right, extend frame by frame with Video Extend and the correct video_frame_offset between segments.

Step 5 — Review at full speed, then frame by frame. Playback hides small artefacts that a client sees on a second viewing. Check the mouth region and the hands specifically — those are the two areas that go wrong first. For a closer look at how it stacks up against other models, see Kling 2.6 Motion Control vs Wan 2.2 Animate.

Where Wan 2.2 Animate Breaks

Plan around these rather than fighting them:

• Camera pans and zooms. The model cannot separate camera motion from subject motion, so a panning driving video produces a scene that fights itself.
• Multiple people in frame. It locks onto the largest face. Group footage is not the use case.
• Extreme framing mismatches. A tight face reference driven by a full-body performance gives you an uncanny crop, not a result.
• Long single-pass clips. Animate is built for short segments plus extensions. Asking for a 30-second one-shot is not how the pipeline works.
• Low-resolution sources. Artefacts amplify when the model downsamples internally, so the source quality sets the ceiling.
• Hands and fast gestures. Motion transfer on fine hand articulation is still the weak point across every model in this category.
• Heavy compositing expectations. Mixed lighting between your reference image and the driving footage does not get fully reconciled — it copies the scene, it does not relight your character to match a different set.

Use Cases That Actually Fit

Want to see the difference on your own footage? Start creating with Wan 2.2 →

• Talking-head content without filming talent — animate a still portrait with your own recorded dialogue performance
• Character consistency across a series — keep one reference image, drive it with different performances
• Product demos with a presenter — animate a presenter still over a scripted performance, useful when nobody wants to be on camera
• Localisation — re-drive an existing performance with a different character reference instead of re-shooting
• Replace a performer in existing footage — Mix mode, keeping the original scene's lighting and colour
• Backup shots for a shoot — if the take is unusable, use the performance with a character image

If your footage is a locked-off, single-person, four-second performance, Wan 2.2 Animate will do the job free. If it is a handheld group scene with a camera pan, no amount of prompt engineering will rescue it.

Try Wan 2.2 Animate Free

You do not need a GPU, a ComfyUI install, or a credit card to test motion transfer on your own footage.

• Upload a character photo plus a driving video — the performance itself is the prompt
• Both modes available — animate your character, or replace a performer while keeping the original scene
• No node installation — skip the model downloads, the folder layout, and the ONNX files
• Free credits to start, so you can validate your framing rules before committing hours to a local setup
• Works with the entire Wan toolset on the same site for text-to-video and image-to-video follow-ups

Test with your shortest, simplest clip first. A four-second locked-off performance from a frame-matched reference will tell you in one generation whether this workflow fits your project.

Related guides

• Kling 2.6 Motion Control vs Wan 2.2 Animate: AI Motion Generation Comparison
• Kling Motion Control vs Wan Animate: Which Motion Transfer Tool Wins in 2026?
• Wan 2.2 vs Wan 3.0: Complete Comparison Guide (Open Weights vs API-Only)

FAQ

What is Wan 2.2 Animate?

It is Alibaba's open-source character animation and replacement model, released as Wan2.2-Animate-14B on September 19, 2025 under Apache 2.0. It takes a character reference image and a driving video and transfers the performer's body motion, facial expression, and lip movement onto your character — or replaces the performer in the original footage.

Is Wan 2.2 Animate free to use?

Skip the setup and test it in the browser: Experience Wan 2.2 Free →

The weights are Apache 2.0, so running it locally in ComfyUI or through the official CLI is free. Browser-based tools typically offer free starting credits, which is the practical option if you do not have a suitable GPU.

What is the difference between Move mode and Mix mode?

Move (animation) transfers motion onto your character image — the character is animated by the performance. Mix (replacement) keeps the original footage and swaps your character in, preserving the scene's lighting and colour. In ComfyUI, Move mode is enabled by disconnecting the background video and character mask inputs.

How long can the output video be?

The model works in segments. Each Video Extend adds 77 frames — roughly 4.8 seconds — and you chain extends with the correct frame offset to build longer clips. Short segments of 4–5 seconds are the practical unit of work.

Why does my output look uncanny or badly cropped?

Almost always a framing mismatch between the reference image and the driving video, or a resolution that is too low. Match the shot type — face to face, half-body to half-body — and use 720p or higher sources. Also check that output dimensions are multiples of 16.

Can Wan 2.2 Animate handle a moving camera?

No. The model cannot separate camera motion from subject motion, so handheld jitter, pans, and zooms in the driving video transfer as artefacts. Shoot locked-off footage.

Does it work if there are several people in the video?

Not reliably. Wan Animate locks onto the largest face in frame, so multi-person footage produces unpredictable results. Single-subject footage is the intended input.

Do I need a powerful GPU to run it?

Running the 14B model locally needs a capable GPU and a proper ComfyUI setup with the correct model folders and pose-detection files. If that is more than you want to manage, the browser route runs the same workflow without an install and is the faster way to check whether your footage is suitable.

References

• Wan2.2 Animate — ComfyUI native workflow — official-style workflow documentation for Move and Mix modes, including mode switching
• WAN Animate 2.2 in ComfyUI — character replace guide — setup details, node packs, and what each model file is for
• Wan2.2 Animate explained: modes, specs, API pricing & setup — release details, licensing, and the FPS/pose-alignment requirements
• Wan 2.2 Animate: open-source character animation workflow — preparation rules for driving footage and known failure cases
• Wan-AI on GitHub — the official repository for weights, the CLI, and the Animate example commands