Wan-Animate-2 - The Wan team's open-source next-generation character animation model

What is Wan-Animate-2

Wan-Animate-2 is the Wanteam's open-source next-generation character animation model. As a major architectural upgrade over Wan-Animate, it abandons reliance on explicit pose skeletons and auxiliary pose extraction networks, switching to an end-to-end dual-branch Diffusion Transformer (DiT) that captures motion priors directly from the reference video. The model supports text-driven viewpoint control and offers a real-time streaming variant capable of up to 24 fps.

Main Features of Wan-Animate-2

• Drive video-to-character animation: input a reference image reference.png and a driving template video template.mp4, letting the reference character replicate the motions and camera rhythm in the driving video.
• End-to-end motion transfer: The redesigned DiT directly consumes the driving video without going through an intermediate motion extractor, aiming to improve both motion fidelity and identity consistency.
• Identity/appearance preservation: Through reference image tokens and reference video tokens participating in attention computation, the reference appearance is injected into the generation process, reducing cases where "the motion looks right but the face falls apart."
• Text normalization control: Before inference, an LLM is first used to generate a fixed-format caption that only writes "character appearance description + background description," without describing motion or inferring emotion, which is then used as the prompt for Wan-Animate-2.
• Text-driven viewpoint control: Supports decoupling the output camera viewpoint from the driving video using text, making it convenient to generate "the same set of motions, different camera positions/viewpoints."
• Base high-quality mode: The default inference script uses the standard configuration; on the official Diffusers side it is num_inference_steps=40, leaning more toward quality first.
• Distillation low-step mode: The distillation weights example is 10 steps,guidance_scale=1.0, Euler solver/no CFG, aimed at faster output and low-latency validation.
• Real-time forward-looking feature: The paper proposes Wan-Animate-2-Lite, using teacher forcing, error buffer, Self-Forcing distillation, and chunk-wise backpropagation to push latency toward the real-time threshold, targeting streaming character animation.
• Local and online experience: Provides local Gradio scripts; use wan_animate_2_gradio.pyfor Base and wan_animate_2_gradio_distillation.pyfor distillation; the README also points to the ModelScope Studio online demo.
• Engineering ecosystem entry points: Weights are hosted on HuggingFace/ModelScope, with an example model ID of Wan-AI/Wan2.2-Animate-2-14B; Diffusers can already be installed from source for integration, while ComfyUI/DiffSynth-Studio are still Todo.
• Configurable parallel generation: By default, parameters are tuned for 8×A800, 720P, and 2×A800, 480P has also been tested; different hardware requires modifying the YAML parallel configuration.

How to UseWan-Animate-2

• Prepare hardware: Prioritize planning according to the official default 8×A800、720P ; the official team also tested 2×A800、480P, so do not assume it can run on a single consumer-grade GPU by default.
• Clone the repository：git clone --recursive https://github.com/Wan-Video/Wan-Animate-2.git。
• Set up the environment: Python 3.11; install torch 2.7.0 / torchvision 0.22.0 / torchaudio 2.7.0, CUDA 12.6 source; then install requirements.txt、flash-attn, and finally pip install -e .。
• Download weights: For HuggingFace use huggingface-cli download Wan-AI/Wan2.2-Animate-2-14B --local-dir ./ckpts/; for ModelScope use modelscope download --model Wan-AI/Wan2.2-Animate-2-14B --local_dir ./ckpts/。
• Generate the caption first: Use an LLM such as Qwen3.7-Plus, and following the official Chinese paradigm, write only "character appearance description + background description"; do not write actions, emotions, or evaluations.
• Run Base: enter infer, execute wan_animate_2_demo.py, pass --prompt、--refer-img-file、--refer-video-file、--config ./wan_animate_2.yaml。
• Run the distilled version: switch to wan_animate_2_distillation.yaml, and add --sample_guide_scale 1.0 --step 10; on the Diffusers side, the corresponding guidance_scale=1.0、flow_solver="euler"。
• Change the parallel configuration: when the number of GPUs/VRAM is inconsistent with the official setup, first change the YAML parallel configs, then troubleshoot OOM or abnormal speed.
• Quick experience: if you do not want to deploy, you can first try ModelScope Studio; locally, run wan_animate_2_gradio.py or wan_animate_2_gradio_distillation.py。

Follow on WeChat and reply "open source" to join theAI Open Source Project Communication Group

Core advantages of Wan-Animate-2

• Motion fidelity and identity consistency are the main selling points: removing the intermediate motion extractor theoretically reduces one layer of error accumulation.
• Closer to the production pipeline: the official setup makes the caption a fixed paradigm in advance, requiring descriptions of character appearance and background, and not describing actions, which facilitates stable reproduction.
• The low-latency route is clear: Distillation can already run 10-step inference; the Lite paper's framing points even more directly at real-time/streaming digital humans.
• Relatively high commercial friendliness: Apache-2.0 is friendlier for secondary development, but actual commercial use still requires handling 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
• Online experience Demo：https://www.modelscope.cn/studios/Wan-AI/Wan2.2-Animate

Comparison of Wan-Animate-2 with similar competing products

Application Scenarios of Wan-Animate-2

• Virtual streamer / digital human livestreaming: The Lite route clearly points to real-time thresholds and streaming character animation, suitable for low-latency livestream rooms where a real person drives a virtual avatar.
• Short videos / short drama character doubles: Use an actor's performance video as the driving video and apply it to character illustrations or reference images, suitable for fantasy characters, IP figures, and low-cost double shots; the core is end-to-end motion transfer and identity preservation.
• Dance, MV, gesture dance fan creations: Use a dance template to drive different character images, creating multi-character dance covers, same-frame costume changes, or stylized MVs; viewpoint control can also produce editing materials for 'same action, different camera angles'.
• E-commerce and brand content: Have brand mascots and virtual models display actions according to a unified template, for use in new product short films, event page videos, and social media materials; before commercial use, portrait, costume copyright, and platform compliance must be handled.
• Education / knowledge creator courseware: Drive an instructor image or cartoon instructor into fixed action templates, batch-generating opening, transition, and key point prompt shots, reducing the cost of repeated filming.
• Game NPC / cutscene previsualization: Planners first use live-action performance to quickly drive character prototypes, verify action rhythm, camera, and emotion, then decide whether to enter the formal motion capture and rendering pipeline.