Wan-Animate-2 - A new generation of character animation models open sourced by the Wanxiang team. - AiBoss

What is Wan-Animate-2?

Wan-Animate-2 is WanxiangThe team's open-source next-generation character animation model, a major architectural upgrade to Wan-Animate, abandons the reliance on explicit pose skeletons and auxiliary pose extraction networks, replacing them with an end-to-end two-branch Diffusion Transformer (DiT) that directly captures motion priors from reference videos. The model supports text-driven viewpoint control and provides real-time streaming variants at up to 24 fps.

Main functions of Wan-Animate-2

Drive video to character animationInput reference image reference.png With driver template video template.mp4This allows the reference character to replicate the actions and camera rhythm in the video.
End-to-end motion transferThe redesigned DiT direct-consumption driven video eliminates the need for an intermediate motion extractor, aiming to simultaneously improve motion fidelity and identity consistency.
Identity/Appearance PreservationBy using reference image tokens and reference video tokens in attention calculations, reference appearances are injected into the generation process, reducing the situation where "the action is similar but the face is distorted".
Text normalization controlBefore reasoning, use LLM to generate a fixed paradigm caption, only writing "character appearance description + background description", without describing actions or speculating on emotions, and then use it as a prompt for Wan-Animate-2.
Text-driven perspective controlIt supports decoupling the output camera view from the driving video using text, making it easy to generate "the same set of actions, different camera positions/views".
Base High Quality ModeThe default inference script uses the standard configuration; the official example uses Diffusers as the default. num_inference_steps=40It prioritizes quality.
Distillation Low Steps ModeExample of distillation weighting: 10 stepsguidance_scale=1.0Euler solver/no CFG, aimed at faster output and low-latency verification.
Real-time forward-looking capabilitiesThe paper proposes Wan-Animate-2-Lite, which uses teacher forcing, error buffer, self-forcing distillation, and chunk-wise backpropagation to reduce latency to a real-time threshold, making it suitable for streaming character animation.
Local and online experiencesProvides local Gradio scripts, for Base use. wan_animate_2_gradio.pydistillation wan_animate_2_gradio_distillation.pyThe README also points to the ModelScope Studio online demo.
Engineering Ecological EntranceWeights are applied via HuggingFace/ModelScope, with the model ID example as follows. Wan-AI/Wan2.2-Animate-2-14BDiffusers can now be installed and integrated from source code, while ComfyUI/DiffSynth-Studio is still in the Todo list.
Configurable parallel generationThe default settings are 8×A800, 720P, but 2×A800, 480P have also been tested; different hardware requires modification of the YAML parallel configuration.

How to useWan-Animate-2

Prepare hardwarePrioritize using the official default settings. 8×A800、720P Planning; Official tests conducted separately 2×A800、480PConsumer-grade single cards should not be assumed to be able to run the game.
Cloning repository：git clone --recursive https://github.com/Wan-Video/Wan-Animate-2.git。
Building the environmentPython 3.11; install torch 2.7.0 / torchvision 0.22.0 / torchaudio 2.7.0, CUDA 12.6 repository; then install... requirements.txt、flash-attn,at last pip install -e .。
Download weightHuggingFace huggingface-cli download Wan-AI/Wan2.2-Animate-2-14B --local-dir ./ckpts/ModelScope modelscope download --model Wan-AI/Wan2.2-Animate-2-14B --local_dir ./ckpts/。
First, the captionUse an LLM like Qwen3.7-Plus, and follow the official Chinese template to only write "character appearance description + background description", do not write actions, emotions or evaluations.
Run Base:Enter infer,implement wan_animate_2_demo.py,pass --prompt、--refer-img-file、--refer-video-file、--config ./wan_animate_2.yaml。
Running distillation version: Change wan_animate_2_distillation.yamland add --sample_guide_scale 1.0 --step 10Diffusers side corresponding guidance_scale=1.0、flow_solver="euler"。
Modify parallel configurationIf the number of GPUs/video memory is inconsistent with the official specifications, first modify the YAML parallel configurations, then check for OOM or speed anomalies.
Quick ExperienceIf you don't want to deploy, you can try ModelScope Studio first; otherwise, run it locally. wan_animate_2_gradio.py or wan_animate_2_gradio_distillation.py。

Wan-Animate-2's core advantages

Authenticity of movement and consistency with identity are the main selling points.Removing the intermediate motion extractor theoretically reduces one layer of error accumulation.
Closer to the production chainThe official approach sets the caption as a fixed format, requiring a description of the character's appearance and background, but not of actions, to facilitate stable reproduction.
Low-latency route is clearDistillation can already run 10 steps of inference; the Lite paper even points directly to real-time/streaming digital humans.
Highly commercially friendlyApache-2.0 is more friendly to secondary development, but in actual commercial use, it is still necessary to deal with material copyright, portrait rights and platform rules.

Wan-Animate-2 project address

Project official website：https://humanaigc.github.io/wan-animate-2
Github repository：https://github.com/Wan-Video/Wan-Animate-2
ModelScope：https://modelscope.cn/models/Wan-AI/Wan2.2-Animate-2-14B
arXiv technical paper：https://arxiv.org/pdf/2608.06009
Experience the demo online：https://www.modelscope.cn/studios/Wan-AI/Wan2.2-Animate

Comparison of Wan-Animate-2 with similar competing products

Comparison Dimensions	Wan-Animate-2	MagicAnimate	MusePose
Core mission	The reference image and driver video generate character animations, emphasizing end-to-end motion transfer and identity preservation.	Animations are generated by combining a single image with motion video, prioritizing timing consistency and fidelity to the reference image.	The video is generated by combining reference images with pose sequences, focusing on dance/full-body movements.
technical route	Redesign the DiT direct consumption-driven video, removing the intermediate motion extractor; add text-driven viewpoint control.	Early diffusion schemes rely on the DensePose driver by default; SD1.5, MSE VAE, and project checkpoints are required.	diffusion-based + pose-guided; based on/optimized Moore-AnimateAnyone routes, and provides pose align.
Action signal	Using driving video directly eliminates the need for an intermediate layer of attitude/keypoint representation.	Depends on the motion video/DensePose class conditions.	The explicit dwpose sequence first aligns the dance video to the reference image before reasoning.
Identity/Details	The goal is high-fidelity motion and strong identity preservation; for realistic effects, it is still recommended to test the face, hands, and flickering with the same material.	The official website admits that faces and hands may be distorted, and the default configuration may result in a style drift from anime to realism.	The official Limitations state that facial areas and complex clothing details remain unstable, and complex backgrounds may exhibit noise or flickering.
Speed/Real-time	Distillation example: 10 steps, no CFG, Euler; the paper also mentions Lite real-time streaming.	It belongs to the early high-quality diffusion route and does not focus on real-time updates as its main selling point.	It doesn't primarily focus on real-time performance; its advantage lies in using pose align to improve availability, rather than a low-latency architecture.
Resolution/Hardware	The default is 8×A800, 720P; the official test showed 2×A800, 480P; different hardware requires modification of YAML for parallel processing.	The official installation requirements are Python 3.8, CUDA 11.3, and ffmpeg; no promises have been made regarding next-generation video memory.	The official specifications state that the 512×512×48 model has approximately 16GB of VRAM, and the 768×768×48 model has approximately 28GB of VRAM; the training machine example has 8×80GB.
Eco-friendly/Easy to use	HuggingFace/ModelScope weights, Gradients, and Diffusers source code integration; ComfyUI is still in the Todo list.	It has early access points such as HuggingFace, Replicate, and Colab, but the architecture is outdated.	Comfyui-MusePose is now supported; training codes were released on March 4, 2025.

Application scenarios of Wan-Animate-2

Virtual Anchor / Digital Human Live StreamingThe Lite approach clearly focuses on real-time thresholds and streaming character animations, making it suitable for creating low-latency live streaming rooms with "real people driving virtual avatars".
Short video/short drama character stunt doubleThe idea is to use the actor's performance video as driving video and overlay it onto the character illustration or reference image. This is suitable for fantasy characters, IP characters, and low-cost stunt double shots. The core is end-to-end motion transfer and identity preservation.
Dance, music video, and hand gesture dance fan creationsUse a dance template to drive different character images, create multi-character dance covers, costume changes in the same frame, or stylized music videos; the perspective control can also create "same action, different camera angles" editing footage.
E-commerce and brand content: Allow brand mascots and virtual models to display actions according to a unified template for use in new product launch videos, event page videos, and social media materials; before commercial use, handle portrait rights, clothing copyrights, and platform compliance.
Educational / Knowledge Blogger CoursewareThe system drives the lecturer's image or cartoon lecturer into fixed action templates, generating opening, transition, and knowledge point prompt shots in batches, reducing the cost of repeated shooting.
Game NPCs / Cutscene PreviewThe plan is to first use live-action performances to quickly drive the character prototype, verify the movement rhythm, camera angles, and emotions, and then decide whether to proceed to the formal motion capture and rendering pipeline.