What Wan 2.2 text-to-video is
Wan 2.2 launched with three open models. Wan2.2-T2V-A14B is the text-to-video model: a mixture-of-experts diffusion transformer with 27 billion parameters, 14 billion active at each step, where a high-noise expert sets the overall layout and a low-noise expert refines the details. Wan2.2-I2V-A14B animates an existing image, and Wan2.2-TI2V-5B is a smaller hybrid that does both and runs on a consumer card at 720p and 24 frames per second.
Compared with Wan 2.1, the 2.2 models were trained on 66% more images and 83% more videos, with detailed labels for lighting, composition, contrast and colour tone. That is why a prompt can name the light and the framing with film terms such as "soft lighting", "warm colors" or "low angle shot", all taken from Alibaba's own prompt vocabulary.
How to write a video prompt
A video prompt describes a short scene, not a still. A structure that works is subject, action, setting, camera, style: "A red fox walks through fresh snow in a pine forest, turns its head towards the camera, low angle tracking shot, soft morning light, cinematic".
- Subject and action: one subject, one clear action with a verb. Five seconds is enough for a gesture, a walk, a wave, not for a story.
- Setting: the place and the time of day set the light.
- Camera: static, pan, tilt, tracking, push-in, drone shot. Name it, or the model picks.
- Style: cinematic, documentary, anime, claymation, 35 mm film. One style word is enough.
Write the action in plain sentences; keywords can set the light and the framing, as in Alibaba's own example prompts, but a list of tags does not describe a motion. Use the negative prompt for the classic defects (blurry, flicker, distorted hands, extra fingers, watermark, subtitles). It has an effect only when the guidance is above 1; the 4-step Lightning LoRA workflows run at 1 and ignore it.
Settings and generation time
The official defaults are 81 frames at 16 fps, 480p or 720p, 40 steps and a guidance scale of 3 to 4; ComfyUI workflows use 20 steps at a guidance of 3.5, or 4 steps at 1 with the Lightning LoRA. With the official settings a 480p clip takes about five and a half minutes on one H100 data-centre GPU, and 720p about three times longer for the same number of frames. The seed works as in image generation: fix it to compare two prompts, change it to get another take of the same prompt. The aspect ratio is chosen before generation, usually 16:9 or 9:16; the model cannot change it afterwards.
Because each attempt is slow, test the prompt at 480p and only then re-run at 720p. The seed of a 480p take gives a different clip at 720p, since the noise has another shape. Most bad results come from prompts with too many actions or contradictory camera instructions, not from the settings.
What text-to-video cannot do yet
Five seconds, 81 frames, is the length the model is set up for; longer generations drift. Long text inside the video is unreliable. Hands and fast, complex interactions between several people still fail often. Physics is plausible rather than exact: water, cloth and hair move well, but a ball can bounce wrong. There is no audio. For a precise look, image-to-video is easier: generate the first frame with Stable Diffusion, SDXL or Flux, then animate it with Wan 2.2.
Running Wan 2.2 T2V locally
The weights are on Hugging Face and run in ComfyUI, Diffusers and the official repository. With the official script, which Alibaba says needs an 80 GB card, the 14B T2V model peaks at about 41 GB of VRAM at 480p and 60 GB at 720p. In ComfyUI it runs on a 24 GB card with fp8 checkpoints and offloading, at several minutes per clip. The 5B model produces a 720p, five-second clip on an RTX 4090 in under ten minutes. In ComfyUI it fits on an 8 GB card; below that, use a cloud GPU or the free demo on this site, which starts from an image used as the first frame.