Skip to content

Wan 2.2 Lightning

View markdown Try it in the explorer

Wan 2.2 Lightning is Alibaba’s fast video generation model from the Wan family. It generates video in around 22 seconds — making it practical for interactive applications and high-volume pipelines.

Wan 2.2 Lightning uses 14B active parameters in a Diffusion Transformer (DiT) architecture with three key components:

  • T5 text encoder — encodes multilingual text input with cross-attention in each transformer block
  • Spatiotemporal 3D VAE — compresses video frames simultaneously across space and time, dramatically reducing compute requirements
  • DiT backbone — processes the compressed latent space with shared MLP modules across transformer blocks

Wan 2.2 Lightning generates video in just 4 diffusion steps without requiring classifier-free guidance (CFG), enabling fast generation while maintaining strong visual quality.

  • Social media content — fast turnaround for short-form video at 720p
  • Rapid prototyping — quickly test video concepts before committing to longer, higher-quality generation
  • High-volume pipelines — the ~22s generation time makes batch processing practical
  • Image animation — bring product shots, illustrations, or photos to life with the img2vid mode

For higher resolution (1080p), longer duration (up to 15s), or features like audio-driven lip sync and video continuation, consider Wan 2.7 instead.

Job type Description
inference.wan2-2.lightning.txt2vid.v0 Generate a video from a text prompt
inference.wan2-2.lightning.img2vid.v0 Generate a video from an input image and prompt
  • prompt (required) — text description of the video to generate, up to 2,500 characters
  • resolution — output resolution: 720p (default, 1280x720) or 480p (832x480)
  • seed — integer for reproducible results
  • image (img2vid only) — input image filename to animate

Wan 2.2 Lightning responds well to specific, action-oriented prompts. Include details about movement, camera angle, and visual style:

  • Be specific about motion: “A cat walking slowly through a garden” works better than “a cat in a garden”
  • Include visual style cues: “cinematic lighting”, “slow motion”, “4k” help guide quality
  • Describe camera movement: “tracking shot”, “pan left”, “aerial view” improve spatial coherence
  • Keep it concise: the model performs best with focused, clear prompts rather than long descriptions

Text-to-video:

{
"type": "inference.wan2-2.lightning.txt2vid.v0",
"config": {
"prompt": "Two anthropomorphic cats boxing on a spotlighted stage, cinematic lighting, dynamic camera angles",
"resolution": "720p"
}
}

Image-to-video (animate a still image):

{
"type": "inference.wan2-2.lightning.img2vid.v0",
"config": {
"prompt": "The person slowly turns their head and smiles, natural movement",
"image": "portrait.jpg",
"resolution": "720p"
}
}