Wan 2.1 and 2.2
Model, video model, by Alibaba Wan
Assessment
The default open video model for a home box: an image and a sentence become a few seconds of believable motion in a few minutes. Use it for stills brought to life and short clips; reach for a lip-sync model when a mouth has to match words.
2026-10-09
Strengths
- Open-weights text-to-video and image-to-video that runs on consumer cards
- The 1.3B model fits a single modest card; the 14B spans two 12 GB cards
- Good motion coherence for its size, especially from a still image
Limitations
- Minutes per clip: about four for ten seconds on two 12 GB cards
- Faces and text in frame still drift; it is a mood-and-motion tool, not a lip-sync tool
Wan is Alibaba's family of open video generation models.[1] Wan 2.1 shipped text-to-video and image-to-video at 1.3B and 14B parameters; Wan 2.2 refined motion and added a mixture-of-experts variant. The 1.3B model is notable for fitting a single consumer card, and the 14B for producing clips that hold together over several seconds when split across two 12 GB cards.
What it is for
Bringing a still to life: a product shot that turns, a portrait that breathes, a landscape with moving weather. Short text-to-video clips for illustration. Anything where the motion matters more than exact faces or readable text.
What it costs
On two 12 GB cards in ComfyUI, about four minutes for ten seconds of video at a modest resolution, with the LLM off the cards for the duration plus an idle timeout.[2] On a 4 GB card the 1.3B model was never worth running.
Things to know
- Start from an image whenever you can; image-to-video is both faster to get right and more controllable than text alone.
- Keep prompts about motion and camera, not about content the image already shows.
- Frame count, not resolution, is the main cost lever.
History
- 2026-10-09: article written and published.
- 2026-10-09: entry created.
Comments
Public comments on each entry are coming. Nothing is collected here yet.