Wan-Animate-2 review, pricing and limits
Alibaba's open-weight successor to Wan2.2-Animate, feeding a driving video straight into the diffusion transformer for identity-stable character animation.
- Best for
- Self-hosted dubbing and localization pipelines
- Not for
- Teams needing a managed, guaranteed-uptime API with no GPU ops
Should you use Wan-Animate-2?
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.
Best for
- Self-hosted dubbing and localization pipelines Apache 2.0 weights avoid per-generation API fees and give full control of the driving-video-to-DiT pipeline.
- Multi-character animated content Native single-to-multiple and multiple-to-multiple motion driving in one generation pass.
- Live avatar or streaming hosts The Wan-Animate-2-Lite variant reaches 24fps at 400x720, crossing the real-time threshold for interactive use.
Not for
- Teams needing a managed, guaranteed-uptime API with no GPU ops No official Alibaba hosted endpoint exists yet; access is self-hosted or via unofficial third-party resellers.
- Budget GPU setups under 16GB VRAM Even the FP8-quantized path needs about 16GB, and the full FP16 checkpoint needs about 28GB.
Wan-Animate-2 pricing
- Pricing model
- open-weights (self-hosted)
- Free tier
- None from Alibaba: the full model is open-weight (Apache 2.0) and free to self-host with no usage caps. Compute cost is the user's own GPU or cloud spend.
| Tier | Rate |
|---|---|
| Self-hosted (Apache 2.0) | Free to download and run on your own GPUs; no per-generation vendor fee. |
| Third-party hosted API (e.g. WaveSpeedAI) | Unofficial resellers charge roughly 10 credits/sec at standard resolution and 20 credits/sec at the higher tier; not an Alibaba-set price. |
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.
Self-hosting requirements
- Min vram gb
- 16
- Self-hosted
- Yes
- Recommended GPU
- RTX 4090 24GB (FP8) for local use; 8x A800 for Alibaba's official 720p reference config
What Wan-Animate-2 can do
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.
What goes in and what comes out
Reads
- image
- video
- text
Writes
- video
Capabilities
- Multi-character animationSupported
- Real-time streaming variantSupported
- Text driven camera controlSupported
- Video inputSupported
- Video outputSupported
- VisionSupported
- Audio inputNot supported
- Audio outputNot supported
- Code executionNot supported
- Function callingNot supported
- Structured outputNot supported
- Tool useNot supported
- Web browsingNot supported
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.
Known limitations
- No independently published numeric benchmark (VBench or similar); only blind user-study win rates are disclosed.
- Full 720p generation targets 8x A800 GPUs in Alibaba's own reference config; the unquantized FP16 checkpoint needs about 28GB VRAM.
- The base model is batch-only; real-time streaming requires switching to the separate Wan-Animate-2-Lite distilled checkpoint.
Who it is for
- ML and video engineers building character-animation or dubbing pipelines who want self-hosted control
- Studios and creators needing multi-character or camera-controlled animation without a per-second cloud fee
- ComfyUI and DiffSynth-Studio users already in the open video-generation tooling ecosystem
- Teams building live avatar or streaming products via the Lite variant
Wan-Animate-2 benchmarks
Standout result
win rate over the original Wan-Animate in Alibaba's blind pairwise user study
Source: Wan-Animate-2 arXiv paper (2608.06009), August 2026Strengths and limits
Strengths
- 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.
- Feeds the driving video directly into the DiT with no separate pose/expression extraction step, cutting identity drift versus prior pipelines.
Show 2 more
- Won over 70% of blind pairwise comparisons against the original Wan-Animate, and is reported to match closed platforms Dreamina and KLING MotionControl.
- Text-driven camera control (12 azimuth x 4 elevation grid) and native multi-character driving are rare in open character-animation models.
Limits
- 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.
Wan-Animate-2 alternatives
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| Wan-Animate-2 This page | Alibaba Cloud | -- | -- | image, video, text → video |
 Wan 3.0Pick Wan-Animate-2 for driving-video character animation with identity-preserving DiT; pick Wan 3.0 for general text/image-to-video generation up to 30 seconds with document input. | Alibaba Cloud | -- | -- | text, image, audio, video, pdf → video |
 Kling 3.0Pick Wan-Animate-2 for free, self-hosted character animation from a driving video; pick Kling 3.0 for a managed 4K/60fps API with native multi-language audio and no GPU ops. | Kuaishou Technology | -- | -- | text, image, video → video, audio |
 Gemini OmniSuperseded by Gemini Omni 1.2 Flash on 5 November 2026; this original Omni Flash from May 2026 is still what Google's AI Plus, Pro and Ultra plans use for video in the Gemini app and Google Flow. | -- | -- | text, image, audio, video → video, audio, image, text |
Wan-Animate-2 specs
- Architecture
- Diffusion Transformer (dual-branch DiT)
- Mixture of experts
- No
- Parameters
- 14B (dense DiT)
- VRAM requirements
- ~28GB FP16 unquantized; ~16GB with FP8 quantization; community GGUF builds fit single 24GB consumer GPUs; official reference config uses 8x A800 for 720p, 2x A800 for 480p
Weight formats
- safetensors
- GGUF (community quantized)
API endpoints
| Name | Endpoint | Auth |
|---|---|---|
| Hugging Face (Wan-AI/Wan2.2-Animate-2-14B) | https://huggingface.co/Wan-AI/Wan2.2-Animate-2-14B | None (download weights) |
| ModelScope demo | https://www.modelscope.cn/studios/Wan-AI/Wan2.2-Animate | ModelScope account |
| GitHub inference scripts | https://github.com/Wan-Video/Wan-Animate-2 | None (self-hosted) |
SDKs
- Python
Prompting patterns
Show 1 prompting pattern
Use named discrete camera angles
"right 60-degree view", "top angle"
Text-driven viewpoint control was trained over a fixed 12-azimuth by 4-elevation grid, so matching the trained vocabulary is more reliable than free-form cinematic language.
Known quirks
Show 4 known quirks
Full model needs 8x A800 for 720p
The base 14B checkpoint's official reference config targets 8x A800 GPUs for 720p, and even 480p testing used 2x A800; the unquantized FP16 weights alone need about 28GB of VRAM.
Workaround: Use the FP8-quantized or community GGUF builds (e.g. QuantStack/Wan2.2-Animate-14B-GGUF) to fit a single 24GB consumer GPU like an RTX 4090, accepting a quality or speed tradeoff.
Base checkpoint can't stream in real time
The full 14B base checkpoint is a batch diffusion model; interactive or live-avatar use cases will stall well below real-time frame rates.
Workaround: Switch to the Wan-Animate-2-Lite distilled variant, which uses teacher-forcing pretraining plus self-forcing distillation to hit 24fps at 400x720 for streaming scenarios.
No independently verified numeric benchmark
Alibaba's release material reports only blind pairwise user-study win rates against Wan-Animate v1, Dreamina, and KLING MotionControl; no VBench-style public leaderboard score has been published as of August 2026.
Workaround: Treat the reported 70%+ win-rate figure as directional, not comparable across papers, and run a side-by-side test on your own footage before committing a production pipeline.
Camera control is a fixed 48-view grid, not free camera
Text-driven viewpoint control is trained over a discrete space of 12 azimuth angles by 4 elevation angles, not a continuous camera path.
Workaround: Phrase camera prompts using the trained vocabulary (e.g. "right 60-degree view", "top angle") rather than cinematic free-form language, and chain multiple generations for continuous camera moves.
Wan-Animate-2 facts
- Trains on your data
- No
- Licence
- Apache License 2.0 · Open source
- Safety posture
- Permissive
- Released
- 07 Aug 2026
- Current version
- Wan2.2-Animate-2-14B
- Family
- Wan
- Variant
- Multimodal
- Deployment
- Self-hosted · Hugging Face · ModelScope · ComfyUI · DiffSynth-Studio
- Data retention
- No Alibaba-side retention applies to the self-hosted open-weight distribution; a generation run through a third-party hosted API instead falls under that reseller's own retention terms.
Content restrictions
- Wan-Animate-2's own hosted terms of service prohibit generating deepfakes or manipulated content of identifiable real people without consent.
- No model-level content filter is enforced once the open weights are self-hosted; policy enforcement is entirely up to the deployer.
Safety controls
- Rlhf
- No
- Red teaming
- not publicly disclosed
- System prompt
- No
- Constitutional AI
- No
- Content filter API
- No
- Moderation endpoint
- No
Data governance
- Trains on inputs
- No
Version history
Wan2.2-Animate-2-14B · 07 Aug 2026
Direct driving-video DiT redesign with Time-Align RoPE, Sparse-Ref Attention, multi-character support, text-driven camera control, and the real-time Wan-Animate-2-Lite variant.
Wan2.2-Animate-14B · 19 Sep 2025
Predecessor. First unified character animation and replacement model in the family; relied on intermediate motion/pose extraction.
What changed
Questions about Wan-Animate-2
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.
Where does your data go when you self-host Wan-Animate-2?
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, nothing leaves your own infrastructure and no Alibaba-side collection applies. Running the model through a third-party hosted API instead puts you under that reseller's own retention terms, not Alibaba's.
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 want a vendor-managed API instead of running their own infrastructure, are better served by a closed commercial platform like Kling MotionControl.
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.
Read the full background
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. It sits alongside other 2026 video-generation models such as Seedance 2.5 and LTX-2.5, part of the wider directory of AI video and animation models.
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, Runway or Pika offer a managed API instead, and a team turning stills into short clips without any GPU ops at all may prefer our guide on turning a photo into a video with free AI tools.
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.
Sources and guides
Official sources
- Wan-Animate-2 model page Vendor
- Licence text Licence
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Alibaba Cloud's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Hokai guides that mention Wan-Animate-2
Lists that include Wan-Animate-2
The provider
More AI Models on Hokai
- Kling 4.0Kling 4.0 is Kuaishou's AI video model for long single-take clips with multi-keyframe control, Omni Reference and stereo audio, announced in September 2026.
- Genie 4Genie 4 is Google DeepMind's interactive world model, announced in October 2026, with native audio, spoken dialogue and scenes steered by text prompts.
LTX-2.5Lightricks' open-weights video and audio model that renders synchronized 4K HDR clips from text, image or video input, self-hostable and fine-tunable.Gemini Omni 1.1 FlashGemini Omni 1.1 Flash (Google, GA 2026) generates and edits AI video through chat, adding native synchronized audio and multi-pass scene extension.
HappyHorse 1.0Built by Alibaba's ATH Business Unit, HappyHorse 1.0 generates synchronized video and audio together from a text, image or reference-image prompt.- Veo 3.1Google DeepMind's Veo 3.1 generates 4, 6 or 8-second video clips with native synchronized dialogue and audio, rendering up to 4K.