WanVideoFree

Wan Text to Video

Wan text-to-video turns a scene description into a short generated clip. It works best when the prompt describes an observable action and camera plan, rather than a long collection of visual adjectives.

Generate a Wan video from text

Choose a draft or delivery model

  • Wan 2.2 5B, 8 credits: lowest-cost five-second 720p visual draft.
  • Wan 2.5 480p, 25 credits: native-audio five-second generation for iteration.
  • Wan 2.5 1080p, 75 credits: final HD pass after motion works.
  • Wan 2.7 720p, 50 credits: newer native-audio model with 2–15 second provider capability; first site tier is locked to five seconds.

Prompt formula

Write six parts in this order: subject, action, camera, setting, light, exclusions.

A ceramic robot folds blue paper into a crane, one continuous slow clockwise camera orbit, clean wooden workbench, warm morning window light, physically plausible hands and paper, no captions, no logo.

If audio matters, name it plainly: soft paper folds, quiet mechanical ticking, no dialogue.

A low-waste workflow

  1. Test motion with Wan 2.2 or a 480p tier.
  2. Change one failed instruction at a time.
  3. Keep successful camera and timing language.
  4. Move to native audio or 1080p only when the shot works.
  5. Download the result and record the model/settings used for reproducibility.

Server-owned cost controls

The model catalog is shared by the UI and the generation route. The server looks up credits, strips display-only tier suffixes, locks the matching duration and resolution, then charges atomically before provider submission. Unknown or canary-unverified models are rejected.

Last checked: August 12, 2026. WanVideoFree is an independent tool, not the official Wan site.

Frequently asked questions

Which Wan model should I use first?

Use Wan 2.2 for the cheapest motion test. Use Wan 2.5 when native audio matters. Use Wan 2.7 for newer controls, longer provider durations, continuation, or editing workflows.

How long should a prompt be?

Long enough to specify subject, action, camera, environment, and constraints, but short enough that instructions do not compete. One clear paragraph is usually easier to debug than a page of adjectives.

What happens when a job times out?

After the server timeout boundary, a still-pending task is settled as failed and eligible credits are returned once. The browser does not own that refund decision.

Are prompts sent to analytics?

Product event properties are sanitized to remove prompt, email, token, media, and URL fields. The inference provider still receives the prompt because it is required to perform the requested generation.

Where this is used

Two use-case walkthroughs build on this page: cinematic shots for depth-of-field and camera language, and short-form social video for 9:16 hook testing at the 8-credit tier.