Wan 2.2: AI Video Generation by Tongyi Lab

Alibaba Group, through its Tongyi Lab research center, has introduced a completely new generation of its video generation model called Wan 2.2. This advancement in artificial intelligence is offered as an open-source solution available worldwide, including in the Czech Republic.

Technical Innovations and MoE Architecture

Wan 2.2 introduces a groundbreaking architecture called Mixture-of-Experts, which has not previously been used in video generation models. This architecture divides the denoising process across timesteps using specialized expert models. Specifically, the model uses two expert systems: a high-noise expert for the initial stages, focusing on the overall layout, and a low-noise expert for the later stages, which fine-tunes video details.

Each expert model contains approximately 14 billion parameters, totaling 27 billion parameters, while only 14 billion are active at each step. This solution significantly increases the model's overall capacity while maintaining the same computational cost as previous versions.

Training Data and Performance

Compared to the previous Wan 2.1 model, Wan 2.2 was trained on a significantly larger dataset containing 65.6% more images and 83.2% more videos. This expansion significantly improves the model's ability to generalize across many dimensions, including motion, semantics, and aesthetics.

On the new Wan-Bench 2.0 benchmark, Wan 2.2 outperforms leading commercial models across most key evaluation dimensions. The model achieves the best performance among all open-source and commercial video generation models.

Cinematic Control and Creative Capabilities

Wan 2.2 offers professional cinematic storytelling through deep control of film language. The model provides fine-grained control over lighting, colors, and composition for versatile styles with nuanced details. Users can create complex motion sequences with improved fluidity and control.

The model can process complex scenes with multiple objects and accurately follow detailed instructions. Examples include generating dynamic hip-hop dance scenes, combat sequences, parkour action, or acrobatic performances on an airplane in midair.

Three Main Model Variants

Wan AI has released three main variants of the Wan 2.2 model:

Wan 2.2-T2V-A14B is a text-to-video model supporting the generation of 5-second videos in both 480P and 720P resolutions. It is built with a Mixture-of-Experts architecture and achieves outstanding video generation quality.

Wan 2.2-I2V-A14B is designed specifically for image-to-video generation and supports both 480P and 720P resolutions. Thanks to its MoE architecture, it achieves more stable video synthesis with fewer unrealistic camera movements and offers enhanced support for a variety of stylized scenes.

Wan 2.2-TI2V-5B is a hybrid model built with the advanced Wan 2.2-VAE, which achieves a compression ratio of 16×16×4. This model supports both text-to-video and image-to-video generation at 720P resolution and 24 frames per second and can run on individual consumer GPUs such as the NVIDIA 4090.

Availability in the Czech Republic

As an open-source solution, Wan 2.2 is available worldwide, including in the Czech Republic, without any regional restrictions. Czech users can download and run the model through GitHub, Hugging Face, and ModelScope. The model is also available online through Hugging Face Space, where the TI2V-5B model can be tested.

Running it locally requires hardware capable of supporting PyTorch 2.4.0 or later, with sufficient computing resources. Without specific optimization, the TI2V-5B model can generate a 5-second video at 720P resolution in less than 9 minutes on a single consumer GPU.

Pricing Policy and Subscriptions

Free basic version: The basic open-source version of the Wan 2.2 model is completely free for both personal and commercial use. Users can download the code and models from GitHub, Hugging Face, and other platforms at no cost.

Wan 2.2 Plus: For advanced features, tools, and templates, there is a paid version called "Wan 2.2 Plus," available through the official WanVideo website.

Hardware costs: The only potential costs are those associated with computing hardware or the use of cloud GPU services, which are not set by Wan AI.

WanBox - a Comprehensive Creative Platform

Wan AI also offers the WanBox creative platform, which enables users to perform various creative tasks, including image generation, video generation, and video editing. The platform offers flexible timeline-based video editing for cutting clips and generating additional content.