The base Wan-Animate-2 checkpoint outputs 720p video at 24 frames per second, while the distilled Wan-Animate-2-Lite variant reaches real-time streaming speed at 400x720 for live avatars. Both ship as free, self-hostable Hugging Face downloads, giving teams an open alternative to closed, subscription-priced animation APIs.
Wan-Animate-2 is Alibaba Tongyi Lab's open-weight character animation model, released August 7, 2026, that beat the original Wan-Animate in over 70% of Alibaba's own blind pairwise comparisons. It feeds a driving video directly into a dual-branch Diffusion Transformer, skipping the pose-extraction step older animation models use, and adds text-driven camera control plus multi-character support.
Provider: Alibaba Cloud · Family: Wan
More about Alibaba Cloud on HokAI
Input modalities: image, video, text · Output: video
About Wan-Animate-2
Wan-Animate-2 is an open-weight character animation model from Alibaba's Tongyi Lab, released August 7, 2026 under the Apache 2.0 license, with arXiv paper 2608.06009 (HumanAIGC group). The base checkpoint, published on Hugging Face as Wan-AI/Wan2.2-Animate-2-14B, is a 14 billion parameter Diffusion Transformer succeeding Wan2.2-Animate-14B from September 19, 2025. Prior pipelines extracted skeletons or landmarks before generation, causing identity drift and lost hand and expression detail. Wan-Animate-2 instead feeds the raw driving video straight into the diffusion transformer. It runs a dual-branch DiT conditioned on a reference image and driving video together, using Time-Align RoPE for alignment and Sparse-Ref Attention to pull only informative reference features. Alibaba has not published a VBench-style score; the only evaluation disclosed is a blind pairwise user study on quality, motion, identity, and expression. Alibaba reports Wan-Animate-2 beating the original Wan-Animate in the majority of these comparisons, and matching or exceeding closed platforms Dreamina and KLING MotionControl on the same axes. There is no token-based context window since this is a video model. Output tops out at 720p at 24 frames per second in Alibaba's reference setup, which also scales down to a lighter resolution on fewer GPUs. It supports single-to-multiple and multiple-to-multiple character driving in one pass, plus text-driven camera control via a discrete 12-azimuth by 4-elevation grid. Inputs are a reference image, a driving video, and a text prompt; output is an animated video carrying the reference character with the driving video's motion and expression, plus an optional text-redirected camera angle. There is no audio input or output, tool-calling, or code execution. Wan-Animate-2-Lite trades quality for speed via teacher-forcing/self-forcing distillation, reaching 24fps at 400x720 for real-time use the base model can't serve. There is no official per-token or per-second Alibaba price; the model is free to self-host under its open license, so GPU time is the only real cost. The unquantized checkpoint needs roughly 28GB of VRAM, and a quantized path brings that down to about 16GB, fitting a single high-end consumer GPU. Third-party clouds that host the model charge per-second credits that roughly double at higher output resolution, a rate the reseller sets, not Alibaba. Alibaba ships inference scripts, base weights, and a distilled checkpoint on Hugging Face and ModelScope, installable with Python 3.11, PyTorch 2.7, CUDA 12.6, and flash-attention. It has native ComfyUI support (a WanAnimate2Cache node halves generation time) and DiffSynth-Studio support with layer-by-layer execution for low-VRAM machines. Third-party clouds like WaveSpeedAI and Atlas Cloud also expose it as a hosted API. Alibaba has not published a formal system card or red-team disclosure; its hosted terms of service prohibit generating deepfakes of real, identifiable people without consent, with violations subject to account termination. Because the weights are open, there is no vendor-side content filter once self-hosted, and policy enforcement falls entirely to the deployer. Teams wanting full control and no per-generation fee are the clearest fit, especially on ComfyUI or DiffSynth-Studio. Live-avatar builders should use Wan-Animate-2-Lite specifically. Teams without dedicated GPU capacity are a worse fit; closed platforms like Kling MotionControl offer a managed API instead. Alibaba has not disclosed a training cutoff, dataset composition, or compute budget. Wan2.2-Animate-14B shipped September 19, 2025 as the family's first unified animation model; Wan-Animate-2 replaced it August 7, 2026 with the direct-driving-video DiT redesign, camera control, multi-character support, and the real-time Lite variant. Alibaba has not announced a successor or folded animation into the separately released Wan 3.0 flagship.
Pricing
Alibaba has not published an official per-token or per-second price; the Apache-licensed weights download and run for free, leaving only your own compute as a real expense. Reseller clouds like WaveSpeedAI list output at roughly 10 credits per second on their standard tier and about 20 credits per second on their higher tier, rates the reseller sets independently of Alibaba.
Key Features
- Direct driving-video DiT ingestion: Feeds the driving video straight into the diffusion transformer with no separate pose or expression extractor, cutting identity drift and preserving micro-expressions and hand movement.
- Dual-branch DiT with Time-Align RoPE and Sparse-Ref Attention: Conditions generation on a reference image and reference video simultaneously, with temporally aligned positional encoding and attention that selectively pulls informative reference features.
- Text-driven camera viewpoint control: A discrete 12-azimuth by 4-elevation viewpoint grid lets a text prompt place the output camera differently from the driving clip's own camera.
- Multi-character animation: Supports single-to-multiple and multiple-to-multiple character motion driving in a single generation pass.
- Wan-Animate-2-Lite real-time variant: A distilled checkpoint trained with teacher-forcing pretraining and self-forcing distillation reaches real-time frame rates for streaming avatars and live hosts.
Pros
- Open weights under the Apache 2.0 license, commercially usable, with no per-generation vendor fee.
- Beat the original Wan-Animate in the majority of Alibaba's own blind comparisons, and is reported to match closed tools like Dreamina and KLING MotionControl.
- Native ComfyUI and DiffSynth-Studio integration lowers the barrier for teams already in the open video-generation tooling ecosystem.
Cons
- Full-resolution generation is GPU heavy: the unquantized base checkpoint needs about 28GB of VRAM.
- No independently published numeric benchmark; only blind user-study win-rate results have been disclosed.
- Real-time use requires the separate Lite or distilled checkpoint; the base model is batch-only.
Benchmarks
- user study win rate vs predecessor pct: 70
Frequently Asked Questions
How much does Wan-Animate-2 cost to use?
There is no official Alibaba per-token or per-second price: the Apache-licensed weights are free to download and self-host, and the real cost is GPU time. Third-party resellers such as WaveSpeedAI charge roughly 10 credits per second of output at lower resolution and about 20 credits per second at higher resolution, but that rate is set by the reseller, not Alibaba.
How does Wan-Animate-2 compare to Kling MotionControl and Dreamina?
Alibaba has not published a numeric VBench-style score against these competitors. Alibaba's blind pairwise study put Wan-Animate-2 ahead of the original Wan-Animate on over 70% of comparisons, and reported it performing on par with or better than the closed rivals Dreamina and KLING MotionControl on visual quality and identity preservation, though that claim has no independent verification yet.
Is Wan-Animate-2 open source?
Yes. Wan-Animate-2 is released under the Apache License with model weights, inference scripts, and a distilled checkpoint published on Hugging Face and ModelScope, and commercial use is permitted without royalty. Alibaba claims no rights over content generated with the model.
Does Wan-Animate-2 train on user data?
Alibaba has not published a data retention or training policy specific to Wan-Animate-2. Because the primary distribution is open weights for self-hosted use, no Alibaba-side data collection applies unless a generation runs through a third-party hosted API, whose own retention terms govern instead.
Who should use Wan-Animate-2, and who should avoid it?
Teams building self-hosted dubbing, localization, or multi-character animation pipelines with existing ComfyUI or DiffSynth-Studio infrastructure get the most value, especially with the real-time Wan-Animate-2-Lite variant for live avatars. Teams without a dedicated GPU, or that need a managed API with guaranteed uptime, are better served by a closed commercial platform like Kling MotionControl.