AI video tools can generate a scene from a sentence, but character animation has always been a different problem: you want a specific person, mascot or avatar to perform a specific movement. That is exactly what Wan Animate 2 does. Released open source by Alibaba's Tongyi Lab on August 7, 2026, it takes one reference image and one driving video, and makes the character in the image perform the motion, expression and gestures of the video — without skeleton extraction, and with the camera angle you want.
What is Wan Animate 2?
Wan Animate 2 (model name Wan2.2-Animate-2-14B) is an end-to-end character animation framework. You give it two inputs:
- A reference image — a character, person, mascot, robot or even an animal. The output keeps this identity.
- A driving video — any clip with body motion, hand gestures and facial expressions. The output copies this performance.
The model then generates a new video where your character performs the driving video's motion, while the appearance stays faithful to the reference image. The background and camera perspective are no longer tied to the source footage — you describe both in a text prompt.
It is a different tool from Wan 3, the text/image-to-video generator. Wan 3 invents a scene from a prompt; Wan Animate 2 transfers a performance onto a character you choose. Used together, they cover both sides of the job: make the footage or the character with Wan 3, then animate with Wan Animate 2.
Key features
- No skeleton extraction. Older systems convert the driving video into bones or keypoints first, which drops fine details like micro-expressions and finger movement. Wan Animate 2 feeds the driving video directly into the model, preserving the subtle motion that makes a performance feel alive.
- Strong identity preservation. The reference character stays recognizably itself across the whole clip — including across big differences in body type, like a cartoon performing a human's dance.
- Text-driven camera control. The output camera is decoupled from the driving video. Ask for "top angle" or "slow orbit shot" and the model re-frames the same performance, without re-shooting anything.
- Prompt-controlled background. The scene behind the character is generated from your prompt, not copied from either input.
- Multi-character scenes. One generation can animate several characters at once, each keeping its own identity and motion.
- Long videos. Hosted services support driving videos up to 120 seconds, with segments chained seamlessly.
- Real-time streaming with Lite. The Wan Animate-2-Lite variant reaches about 24 fps at 400×720, opening the door to digital-human live streaming and interactive avatars.
How it works
The core idea is simple: the reference video is the best motion prior. Instead of compressing motion into skeletons or latent features and losing information along the way, a redesigned diffusion transformer (DiT) consumes the driving video directly.
Three architectural choices make this practical:
- Dual-branch DiT — a reference branch keeps the clean motion information at time t=0 and feeds key/value features to the generation branch, so motion does not decay during denoising.
- Time-aligned RoPE — reference and generated frames share one timeline, keeping frame-to-frame correspondence accurate even when input and output resolutions differ.
- Sparse reference attention — each generated frame only attends to its time-aligned reference tokens, cutting memory and compute without losing motion fidelity.
A viewpoint LoRA (trained on roughly 50K multi-view rendered samples) turns camera control into a text task, and the Lite variant uses a three-stage training recipe — teacher-forcing pretraining with an error buffer, Self-Forcing distillation and chunk-wise backpropagation — to stream generation chunk by chunk without accumulating errors.
The training data includes more than 100,000 paired synthetic videos covering full-body, half-body and close-up shots at up to 2.5K resolution.
How to use it
There are three ways to get started:
- Try the online demo — the ModelScope studio demo runs the same model for free.
- Use a hosted API — services such as WaveSpeedAI expose it as a REST API with 480p and 720p output, charged per generated second, ideal for products and batch workflows.
- Run it locally — the open-source repo, ComfyUI nodes, Hugging Face Diffusers pipeline and DiffSynth-Studio integration cover everything from inference to LoRA fine-tuning. Note that the official defaults target 8×A800 GPUs at 720P and 480P was tested on 2×A800, so a serious GPU setup (80GB-class) is recommended.
To get good results, prepare three things:
- A clear reference image of your character, full body preferred.
- A driving video with clean, visible motion — avoid heavy occlusion and motion blur.
- A prompt that describes appearance and background only. The recommended pattern is to caption the reference image with an LLM first (appearance + background, no action words), because the motion already lives in the driving video.
Practical tips: keep the reference image and driving video in the same framing (full-body to full-body), start with short clips at 480p, and describe the output background explicitly.
What it does best
Wan Animate 2 shines wherever you need a chosen character to follow a real performance:
- Digital humans and live streaming — the Lite variant brings streaming character animation to interactive scenarios like virtual hosts and live e-commerce.
- Dance and complex motion transfer — fast, non-rigid movements like dance, instrument playing and sports come out with far fewer artifacts than skeleton-based methods.
- Anime and IP characters — turn a static illustration into an animatable character, including multi-character scenes with independent identities.
- Marketing and e-commerce — bring brand mascots and product characters to life in ad creatives without a shoot.
- Education and content creation — turn avatars and course hosts into talking-head videos quickly.
- Game and metaverse prototyping — text-driven camera control makes it easy to render the same action from multiple angles for previsualization.
In the team's blind user studies, over 70% of users preferred Wan Animate 2 over the previous generation, and results were competitive with commercial tools such as Dreamina and Kling MotionControl. As with any creative tool, the fair way to judge it is to test it on your own characters and footage.
The bottom line
Wan Animate 2 is the most practical open character-animation option in 2026: Apache 2.0 licensed, free to use commercially, simple inputs, strong identity preservation and a real-time path for interactive applications. It is not a full scene generator — for inventing footage or character images from a prompt, that is Wan 3's job — but as the motion-transfer half of a video pipeline, it is hard to beat.
Ready to build the inputs? Generate your character images and driving footage with Wan 3, then animate them with Wan Animate 2.
