What is Wan? Choosing tasks across the video model family and cloud services
Wan is Alibaba’s visual generation model family with multiple versions and task branches. Learn how text-to-video and image-to-video relate to open weights, cloud APIs and audio-capable outputs.
Meet this model
Wan is Alibaba’s visual generation family, commonly creating video from text, images or reference clips. Public weights such as Wan2.2 and later cloud models do not all share one specification. Identify the version and task branch before checking supported inputs and outputs.
Text-to-video and image-to-video start with different material
Use text-to-video when starting from a written shot description. If product appearance is fixed, image-to-video can use its photo as the starting reference. The picture alone does not define the entire sequence; describe movement and check for changed parts after rotation.
The cloud catalog separates task-specific models, some supporting reference clips or audio conditions. Changing only a model name may not switch the task correctly. Check required fields and whether the service can read the image URL, rather than merely whether it opens locally.
Wan2.2 expert architecture matters for local inference
Wan2.2’s A14B branch divides denoising between high-noise and low-noise experts, emphasizing layout early and details later. This is a version-specific design, not a universal Wan architecture or a way to infer local GPU memory requirements.
The public repository provides inference implementations, but local requirements depend on checkpoint size and precision. Open weights can suit research that modifies inference; a hosted API may suit software submitting generation jobs. No weights were downloaded and no device speed was tested here.
Verify audio versus silent output for each model
Alibaba Cloud’s text-to-video guide distinguishes silent and audio-capable models, listing audiovisual synchronization for later Wan2.6 and Wan2.7 versions. Do not attribute those outputs to every local Wan2.2 checkpoint. Audio input also does not necessarily mean its original sound is preserved.
For a narrated product clip, decide whether narration is generated or added in editing. Play the whole output and compare speech with the product action. Inspect actual duration and resolution rather than generalize one model’s specifications to the family.
Check repository licenses separately from cloud availability
The Wan2.2 repository uses Apache 2.0 and provides model download and usage instructions. Check the specific model card and accompanying material. That license does not establish open weights for every newer cloud model or remove cloud account quotas and regional conditions.
Cloud generation generally involves submission followed by result retrieval. A task ID establishes acceptance, not completion; retrieve the finished video. Review character continuity across shots. This article explains official repository and documentation differences without generating clips or measuring business outcomes.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.