# Wan Animate: Character Swap and Motion Transfer, Explained

> Wan Animate (Wan 2.2 Animate) explained from its official sources: animation and replacement modes, inputs, license, hardware, limits and consent rules.

By Anna Krejcova | Published: 2026-10-21T08:00:00Z

Canonical: https://www.fanscreator.ai/blog/wan-animate-character-swap

- Wan Animate (Wan2.2-Animate-14B) is an open-weight Alibaba Tongyi Lab model released on 19 September 2025 under Apache 2.0. It takes one video and one character image.
- Animation mode moves a still character with the performer's motion and expressions. Replacement mode swaps the performer for your character and keeps the scene's lighting and color tone.
- Running it locally needs a preprocessing pipeline and serious GPU memory; the repository documents its other 14B models on a GPU with at least 80GB VRAM. Mask extraction works on single-person videos only.
- Only swap in your own face, your own AI persona or a person whose consent you can prove. Fansly bans photorealistic face-swapped content, and OnlyFans requires the verified creator.

**Wan Animate** (released as Wan2.2-Animate-14B) is an open-weight model from Alibaba's Tongyi Lab that does two jobs with one set of weights. It can animate a still picture of a character with the movements and expressions of a person in a reference video. Or it can replace the person in that video with your character and keep the scene, its lighting and its color tone.

Every Wan Animate detail below comes from the Wan2.2 GitHub repository, the Hugging Face model card, the paper and the project page, all read on 5 October 2026.

## What Wan Animate is

Wan Animate is a character animation and replacement model built on Wan 2.2. The team introduced it on 19 September 2025 as "an unified model for character animation and replacement with holistic movement and expression replication". The weights and inference code came out the same day, according to the [Wan2.2 GitHub repository](https://github.com/Wan-Video/Wan2.2).

The [paper](https://arxiv.org/abs/2509.14055) is credited to the HumanAIGC Team at Tongyi Lab, Alibaba. It describes the model as built on the Wan image-to-video model and post-trained with extra control signals. It also appears as Wan 2.2 Animate and Wan-Animate; all refer to the same model.

| Fact | What the official sources say |
| :--- | :--- |
| **Name** | Wan2.2-Animate-14B |
| **Released** | 19 September 2025, weights and inference code |
| **Built on** | Wan image-to-video (Wan-I2V) |
| **Inputs** | One video and one character image |
| **Modes** | Animation and replacement |
| **License** | Apache 2.0 |
| **Integrations** | ComfyUI and Diffusers (Diffusers added 13 November 2025) |
| **Hosted demos** | wan.video, ModelScope Studio and a Hugging Face Space |

## The two modes: animation and replacement

Wan Animate takes the same two inputs in both modes. What changes is which part of the result comes from the video.

### Animation mode: motion transfer onto a still character

In animation mode, the model "generates a video of the character image that mimics the human motion in the input video", in the words of the repository. The paper says it replicates both the expressions and the movements of the person in the reference video.

The result is a new video of your character, doing what the performer did. This is the "motion control" use case: the reference video acts as the motion source, and the picture decides who appears on screen.

### Replacement mode: character swap that keeps the scene

In replacement mode, the model puts the animated character into the reference video in place of the original person. According to the [project page](https://humanaigc.github.io/wan-animate/), it keeps the performer's expressions and movements and replicates the original lighting and color tone so the new character fits the environment.

This is what most people mean by "character swap AI" or "replace person in video AI". The scene stays. The person changes.

| Mode | Input video provides | Character image provides | Background in the result |
| :--- | :--- | :--- | :--- |
| **Animation** | Motion and expressions | Who appears and what they look like | Comes from the picture |
| **Replacement** | Motion, expressions and the scene | Who appears and what they look like | The original video's scene |

## How Wan Animate works

The model separates body motion from facial expression and controls each one differently. The paper describes three parts:

- **Body motion** comes from skeleton signals extracted from the reference video and aligned to the frame.
- **Facial expression** comes from face images cropped from the reference video, encoded so the expression is kept but the identity is not.
- **Lighting** in replacement mode is handled by an extra "Relighting LoRA", which applies the scene's lighting and color tone while keeping the character's appearance.

That split explains why the model needs a preprocessing step. Before generation, the video is turned into a face video and a pose video. Replacement mode also produces a background video and a mask video that marks where the person was. The [preprocessing guide](https://github.com/Wan-Video/Wan2.2/blob/main/wan/modules/animate/preprocess/UserGuider.md) lists the extra models this needs. Pose detection is always needed. Mask extraction is needed for replacement mode. An image editing model is optional and only used for the enhanced pose retargeting in animation mode.

## Model size, license and hardware

**Size.** It is a model named Animate-14B (listed at 17B parameters, BF16, on the [Hugging Face model card](https://huggingface.co/Wan-AI/Wan2.2-Animate-14B)).

**License.** The weights are released under Apache 2.0. Separately, the model card's usage note says the team claims no rights over generated content, and that use "must not involve sharing any content that violates applicable laws, causes harm to individuals or groups", among other limits.

**Hardware.** The repository does not publish a memory figure for Animate-14B on its own. It gives single-GPU and multi-GPU commands, and the multi-GPU example runs on 8 GPUs. For the other 14B models in the same repository, the single-GPU commands are documented as running on "a GPU with at least 80GB VRAM". Treat that as the closest official guide. A consumer graphics card is not what the official setup was written for. Community ComfyUI setups may need less memory, but the official sources publish no figure for them.

If you run it yourself, you will also install the preprocessing models, manage the intermediate videos and check them before generation. Running it locally gives full control. It also means setup and render time are your job.

## Limits the official sources state

The preprocessing guide is direct about where the open release falls short. These are the points that matter for creators:

- **One person per video.** The mask extraction for replacement mode is "designed for single-person videos ONLY" and may fail or give wrong results with several people in frame.
- **Body proportions matter in replacement mode.** Pose retargeting is off by default in this mode. The guide warns that a significant mismatch in body proportions between the character and the performer may cause artifacts or deformations.
- **Retargeting is not guaranteed.** In animation mode, retargeting is recommended when the two bodies differ. But the guide says the simplified version is "NOT guaranteed to be perfect" and advises checking the preprocessing results first.
- **Frame rate is a trade-off.** A lower frame rate speeds up generation but may cause stuttering or choppiness.
- **No Wan 2.2 LoRAs.** The repository does not recommend using LoRA models trained on Wan 2.2 with Wan Animate, because they may lead to unexpected behavior.
- **Research framing.** The project page states the project is "intended solely for academic research and effect demonstration." The weights themselves are released under Apache 2.0, with the model card's usage limits on top. Read both before you build on the model.

For general motion problems such as warping hands or smeared frames, see [how to fix AI video artifacts and motion blur](/blog/fix-ai-video-artifacts-motion-blur).

## How this maps to character swap and video face swap

Wan Animate is one model. Creators usually think in tasks: "change the face", "change the person" or "make this picture move". Here is how the two modes line up with the swap features in FansCreator. This is a comparison of what each one does. It is not a statement about which model FansCreator runs behind its swap features.

| Task | Wan Animate | FansCreator |
| :--- | :--- | :--- |
| **Replace the whole person, keep the scene** | Replacement mode | Character swap on video |
| **Replace only the face, keep the body** | No face-only mode | [Video face swap](/ai-video-face-swap) |
| **Make a still picture move** | Animation mode, with motion from a reference video | Image-to-video from a picture of your persona |

### Character swap: the whole person

Character swap replaces the whole person in a video with your persona and keeps the scene. It covers the same job as Wan Animate's replacement mode, without the local setup. Use it when the hair or the build of the person in the clip is clearly different from your persona. In that case a new face alone would sit on a body your subscribers do not recognize.

### Video face swap: only the face

Video face swap places a face from a picture onto the person in a video and keeps the body, the movement, the scene and the sound. Wan Animate has no mode for this. In FansCreator, video face swap works on clips up to 30 seconds with one person in them. It keeps the original audio, takes a minute at most and adds no watermark on any plan. It exports at 720p on the free plan and up to 1080p on paid plans. The source can be an uploaded video or a clip generated in FansCreator. GIFs are not supported.

The rule of thumb from [how to keep one face across AI video clips](/blog/keep-one-face-across-ai-video-clips) holds: face swap when the body already matches, character swap when it does not. For still pictures, use [AI face swap](/ai-face-swap).

### Animation mode and image-to-video

The closest FansCreator workflow to animation mode is image-to-video. It starts from a picture of your persona and adds motion, so the face and body in the clip are the ones you approved. The [Wan NSFW page](/wan-nsfw) explains how FansCreator runs Wan 2.2 video in the browser, and [text-to-video vs image-to-video](/blog/text-to-video-vs-image-to-video) covers when to start from a picture.

## Consent and platform rules

Any tool that swaps a person into a video raises consent first. The rule is short: only use your own face, your own AI persona, or the face or body of a person whose consent you hold and can prove. The FansCreator [Terms of Use](/terms) prohibit anything else.

The source video needs the same care. A character swap changes who appears, but the motion and the scene still come from the original recording. Use video you generated yourself, or a recording of a performer who agreed to have their motion used in a swap.

Each platform then adds its own conditions. Policies change, so check the linked pages before you post. All were read on 5 October 2026.

| Platform | What it says about AI and swapped content |
| :--- | :--- |
| **OnlyFans** | AI content is allowed only if "it clearly features the verified Creator" and is labelled with a signifier such as #ai or #AIGenerated. Accounts or content that are "wholly AI-generated" cannot be shared, and content that "features heavily altered images" is also listed. |
| **Fanvue** | Fully AI-generated content is allowed. Deepfaked or face-swapped content is allowed, but "you MUST verify the real model behind the AI". |
| **MYM** | AI content must be labelled "Virtual Image", #AI or #AIgenerated. Deepfakes need written permission from the person being deepfaked. |
| **Fansly** | Photorealistic AI content is not allowed, including "face-swapped content", and labelling it as AI does not make it permissible. |

**OnlyFans.** The [Terms of Use](https://onlyfans.com/terms) allow AI-generated, altered or enhanced content only when it features the verified creator and a user can tell it is AI. The [Help Centre](https://onlyfans.com/help/196/205/211) repeats that all content must belong to a verified creator, and its [content guide](https://onlyfans.com/help/196/205/206) rules out accounts or content that are wholly AI-generated, as well as content that "features heavily altered images". A fully AI persona swapped into a video is not a fit for OnlyFans. A persona built from scratch belongs on Fanvue or MYM, with their AI labels; [how to create an AI influencer](/how-to-create-an-ai-influencer) covers building one.

**Fanvue.** Its help article [Is AI content allowed on Fanvue?](https://help.fanvue.com/en/articles/9538738-is-ai-content-allowed-on-fanvue) allows face-swapped content if the real model is verified. Its [KYC article for AI creators](https://help.fanvue.com/en/articles/9539091-passing-kyc-and-creating-multiple-accounts-as-an-ai-creator) goes further: "All real people need to be verified, even if you're only using their body and not their face". That matters for face swap, where the body in a recorded clip belongs to a real person.

**MYM.** The [MYM AI content article](https://support.mym.fans/hc/en-us/articles/14997634314524-AI-Generated-Content-on-MYM-What-You-Need-to-Know) requires the AI label and written permission for any deepfake. MYM warns that publishing a sexual deepfake without meeting its conditions risks imprisonment and fines.

**Fansly.** The [Fansly AI content article](https://help.fansly.com/en/articles/12315578-ai-generated-content-on-fansly) bans photorealistic AI content outright. Only non-photorealistic virtual personas can apply.

The platform-by-platform detail for swapped video is in [is video face swap allowed on OnlyFans, Fanvue and Fansly](/blog/video-face-swap-onlyfans-fanvue-fansly-rules).

## Wan Animate in short

Wan Animate is an open-weight, Apache 2.0 model that takes a video and a character image. Animation mode moves your character the way the performer moved. Replacement mode puts your character into the performer's scene and matches the lighting. Running it locally means a 14B model, a preprocessing pipeline and, going by the repository's other 14B models, a GPU with at least 80GB of memory.

Hosted tools cover the same tasks: character swap replaces the whole person and [AI video face swap](/ai-video-face-swap) replaces only the face. Whichever route you take, the face and the body have to be yours to use, and the platform has to allow the result.

## Frequently asked questions

### Is Wan Animate free and open source?

The weights and inference code for Wan2.2-Animate-14B are published on GitHub and Hugging Face under the Apache 2.0 license. The model card says the team claims no rights over generated content, and asks that use not involve content that breaks the law or harms people. Running it locally is free apart from your own hardware.

### What is the difference between Wan Animate's animation and replacement modes?

Both modes take a video and a character image. Animation mode creates a new video of the character copying the performer's motion and expressions. Replacement mode puts the character into the original video in place of the performer and matches the scene's lighting and color tone.

### What GPU do I need to run Wan Animate locally?

The official repository does not publish a separate memory figure for Animate-14B. Its single-GPU commands for the other 14B Wan 2.2 models are documented as running on a GPU with at least 80GB VRAM, and its multi-GPU example uses 8 GPUs. You also need the preprocessing models, including pose detection.

### Is character swap the same as face swap?

No. A face swap replaces only the face and keeps the body, the movement and the scene. A character swap replaces the whole person and keeps the scene. Use face swap when the body in the clip already matches your persona, and character swap when it does not.

### Can I post a character-swapped video on OnlyFans or Fansly?

On OnlyFans, AI content must clearly feature the verified creator and carry a label such as #ai, so a fully AI persona does not qualify. Fansly does not allow photorealistic AI content, including face-swapped content, even when it is labelled. Fanvue allows face swaps if the real model is verified.

## Sources

- https://github.com/Wan-Video/Wan2.2
- https://github.com/Wan-Video/Wan2.2/blob/main/wan/modules/animate/preprocess/UserGuider.md
- https://huggingface.co/Wan-AI/Wan2.2-Animate-14B
- https://arxiv.org/abs/2509.14055
- https://humanaigc.github.io/wan-animate/
- https://onlyfans.com/terms
- https://onlyfans.com/help/196/205/211
- https://onlyfans.com/help/196/205/206
- https://help.fanvue.com/en/articles/9538738-is-ai-content-allowed-on-fanvue
- https://help.fanvue.com/en/articles/9539091-passing-kyc-and-creating-multiple-accounts-as-an-ai-creator
- https://support.mym.fans/hc/en-us/articles/14997634314524-AI-Generated-Content-on-MYM-What-You-Need-to-Know
- https://help.fansly.com/en/articles/12315578-ai-generated-content-on-fansly

Keywords: wan animate, wan 2.2 animate, character swap ai, motion control ai, wan animate 2.2, replace person in video ai