Wan 2.2 animation: how Animate and Replace modes work, what they need and what they cost
Wan 2.2 Animate is Alibaba's open-weights model with two modes: animation puts a still character in the motion of a driving video, and replacement swaps a person in that video for your character. Alibaba's hosted API takes 2 to 30 second videos (Alibaba Cloud). If you need control, run it locally; otherwise pay per second.
Key facts
| Release and license | Open weights released 2025-09-19 under Apache 2.0[1] |
|---|---|
| Two modes | Animation (character copies the video's motion) and replacement (character takes the video person's place)[2] |
| Alibaba API input limits | Video 2 to 30 s, up to 200 MB; image up to 5 MB, 200 to 4096 px per side[5] |
| Alibaba API price (Singapore) | $0.12 to $0.26 per output second, as of 2026-10-26[7] |
| fal price | $0.04 (480p), $0.06 (580p), $0.08 (720p) per billed second[9] |
| People in the driving video | Mask extraction supports single-person videos only[3] |
Every limit and price below comes from the model maker's own documentation or the host's pricing page, linked in the sources list. Each price is dated.
What is Wan 2.2 Animate?
Wan 2.2 Animate (Wan-Animate) is an open-weights model from Alibaba's Wan team, one of the open Wan video models, that copies a performer's body motion and facial expressions onto a character image. The weights and inference code were released on 2025-09-19 under Apache 2.0, and the paper was posted to arXiv on 2025-09-17.
It takes two inputs: a character image and a driving video. The paper says body motion is driven by skeleton signals, and expressions come from facial features in the video. An optional Relighting LoRA helps a replaced character match the scene's light.
Animate vs replace: which mode do you need?
Wan 2.2 Animate offers animation mode, which makes your character move like the video, and replacement mode, which puts your character into the video in place of the person.
| Animation (Move) | Replacement (Mix) | |
|---|---|---|
| Result | New video of your character in the motion of the source | Original video with the person swapped out |
| Background | Not taken from the source video | Kept from the source video |
| Extra preprocessing | Pose retargeting (retarget_flag, optional use_flux) |
Mask extraction (replace_flag) |
| Retargeting default | Strongly recommended | Off, to keep contact with the scene |
| Relighting LoRA | Not used | use_relighting_lora at inference |
| Cost on Alibaba API (Singapore, per second) | $0.12 standard, $0.18 pro | $0.18 standard, $0.26 pro |
Choose animation when you want a new scene around your character. Choose replacement when the source footage, set and camera work are already what you want. Prices are from Alibaba Cloud's pricing page, checked 2026-10-26.
What does the reference video need?
Use a driving video with one clearly visible person. Alibaba's API accepts MP4, AVI or MOV files of 2 to 30 seconds, up to 200 MB. Each side must be 200 to 2048 px, with an aspect ratio between 1:3 and 3:1. Alibaba recommends higher resolution and frame rate for better results.
The character image can be JPG, JPEG, PNG, BMP or WEBP, up to 5 MB, 200 to 4096 px per side. In ComfyUI, set the generated video's width and height to multiples of 16.
Two rules come from the preprocessing guide:
- The official mask extraction handles single-person videos only. For multi-person footage, use your own masking tool.
- Basic pose retargeting needs the character image and the first driving frame to show a front-facing, stretched pose.
How do you run Wan 2.2 Animate locally?
You can run Wan 2.2 Animate locally with the official scripts or with a ComfyUI template. Both preprocess the video first, then generate.
With the official repo:
- Run
preprocess_data.pywith--resolution_area 1280 720. - For animation, add
--retarget_flagand optionally--use_flux. For replacement, add--replace_flagplus the mask settings--iterations 3 --k 7 --w_len 1 --h_len 1. - Run
generate.py --task animate-14Bon the output folder. For replacement, add--replace_flag --use_relighting_lora. - On one GPU, the model card lists
--offload_model True --convert_model_dtype.
With ComfyUI, load the Animate template, install ComfyUI-KJNodes and comfyui_controlnet_aux, and upload the image and video. DWPose turns the video into pose and face control videos. Each Video Extend group adds 77 frames, about 4.8 seconds. ComfyUI's page lists a 4-step acceleration LoRA, and says first-time users should start at a small resolution in case VRAM is short.
For VRAM numbers, see Wan 2.2 in ComfyUI for the VRAM numbers published for the other Wan 2.2 models.
What goes wrong, and what the docs say to do
The official guide documents several input and preprocessing problems. This table maps each documented limit to a fix.
| Symptom | Documented cause | Fix |
|---|---|---|
| Body shape or limbs look deformed | Large proportion mismatch between character and performer | Turn on retargeting in animation mode. In replacement mode, pick a performer with similar proportions |
| Pose retargeting looks off | FLUX-enhanced retargeting is not guaranteed to work | Check the intermediate pose video before generating |
| Replacement fails or swaps the wrong person | Mask extraction handles one person only | Crop or trim to a single performer |
| Halo or background changes around the character | Mask size trade-off | Tighter mask keeps more background but limits the area; looser mask gives freedom but can shift the background |
| Choppy motion | Low processing fps |
Raise fps; lower values are faster |
| Odd results with a custom style | LoRAs trained on Wan 2.2 may misbehave with Wan-Animate | Do not stack Wan 2.2 LoRAs |
How much does Wan 2.2 Animate cost?
Hosted Wan 2.2 Animate costs $0.04 to $0.26 per second of video on the two providers (fal and Alibaba) whose prices were checked on 2026-10-26. Running locally costs only your own hardware and time.
| Provider and tier | Rate per second | 10-second clip |
|---|---|---|
| fal Animate, 480p | $0.04 | $0.40 at 16 fps |
| fal Animate, 720p | $0.08 | $0.80 at 16 fps, $1.50 at 30 fps |
| Alibaba Singapore, Move standard | $0.12 | $1.20 |
| Alibaba Singapore, Move pro | $0.18 | $1.80 |
| Alibaba Singapore, Mix standard | $0.18 | $1.80 |
| Alibaba Singapore, Mix pro | $0.26 | $2.60 |
Alibaba bills the output duration of successful videos only, and failed calls are free. Its Beijing region lists lower rates, such as $0.06 for Move standard. fal bills by frames divided by 16, so a 10-second clip at 30 fps is 300 frames, or 18.75 billed seconds, which is $1.50 at 720p. The 10-second rows are my arithmetic from those rates, and fal's page leaves the output frame rate unstated, so the table shows 16 and 30 fps.
To compare other models, use the AI video cost calculator. If one take in three is usable, multiply by three. A keeper on Alibaba's Mix pro would then cost about $7.80 for 10 seconds.
Is Wan 2.2 Animate free?
The weights are free under Apache 2.0, and Alibaba's Singapore region lists a free quota of 50 seconds valid for 90 days. Beyond that you pay per second or supply your own GPU. A local GPU costs whatever hardware you already own.
Likeness and rights
Replacement mode puts a character into footage of a real person, and animation mode can animate a photo of one. Use only your own characters, licensed characters, and videos of people who gave written consent. Alibaba's API moderates inputs and outputs and can reject a job as suspected IP infringement. Run face photos through the character photo checker before you spend credits. For a comparable hosted motion-transfer tool, see Kling Motion Control, and for the hosted-only Wan versions see the Wan 2.5 guide.
Checklist before you generate
- Driving video of 2 to 30 seconds for Alibaba's API, with one person if you use the official mask extraction.
- Character image sharp, with a front-facing, stretched pose if you use basic retargeting.
- Similar body proportions, or retargeting on in animation mode.
- Generated video width and height in multiples of 16 for ComfyUI.
- Preview the pose video first, then spend on the full render.
- Rights confirmed for the character and the performer.
Questions
Is Wan 2.2 Animate free?
The weights are free to download under Apache 2.0, so running it on your own GPU costs no license fee. Hosted versions charge per second of output. Alibaba's Singapore region lists a free quota of 50 seconds valid for 90 days, as of 2026-10-26.
What is the difference between animate and replace in Wan 2.2?
Animate only borrows motion and expressions from the video, so the background comes from you. Replace keeps the original video, including its background, and swaps the person for your character, using a mask and an optional relighting LoRA.
What are Move and Mix in ComfyUI and Alibaba's API?
They are the same two modes under other names. Move is animation and Mix is replacement. ComfyUI's template switches between them by connecting or disconnecting the background video and character mask inputs.
Can I use anyone's face or a copyrighted character?
Use only characters you own or have written permission to use, and only videos of people who agreed. Alibaba's API runs content moderation on inputs and outputs and can return an IP infringement error.
Sources
- Wan2.2-Animate-14B model card
- Wan2.2 repository (Animate section)
- Wan-Animate preprocessing guide
- Wan-Animate: Unified Character Animation and Replacement with Holistic Replication
- Wan2.2 animate move API reference
- Wan2.2 animate mix API reference
- Model pricing
- Wan2.2 Animate in ComfyUI
- Wan 2.2 14B Animate Move
Related guides
- ComfyUI Wan 2.2: how to run it locally, VRAM needs and GGUFRun Wan 2.2 in ComfyUI: official 5B needs 8 GB VRAM, 14B files add up to 28.6 GB in fp8. File sizes, GGUF quant table and setup steps, checked 2026-10-26.
- Wan 2.5 guide: features, access, cost and open-source statusWan 2.5 is hosted-only. Wan 2.2 is the newest open-weights Wan. See the version table, per-second API prices and what a 10-second clip costs.
- Kling Motion Control: how to use it, the input rules and what a clip costsKling Motion Control copies motion from a 3-30 second video onto a still image. Input rules, a per-clip cost table and a pre-flight checklist.
- Best photo for image to video: check yours before you spend creditsCheck a photo for image-to-video before you pay: format, size, aspect range, blur and face size, with limits from Runway's docs. Runs in your browser.
