Create a Wan Video

This is the working order, not the theory. Most first attempts fail at step 1 — starting from text when the shot was already clear in a photograph, or the reverse — and the rest fail at step 3 by iterating at final size.

Step 1: decide whether you have a first frame

Ask one question: does the shot already exist as a picture?

• Yes → use image to video. You supply the frame and describe only the movement. This is the shorter path by a wide margin, because composition, lighting and framing are already settled.
• No → use text to video, and expect to spend longer on the prompt, because the model is inventing the scene as well as the motion.

If the shot needs a style carried over from clips you already have, reference to video is the third entry point. If it needs to continue something you already rendered, that is extend a video.

Step 2: write one sentence with three parts

Name the subject, the action, and how the camera moves. "Slow push in on a woman in a red coat walking through a rainy street, shallow depth of field" contains all three. A subject with no direction tends to drift.

Keep it literal. Named lighting — "overcast afternoon light" — does more than mood words like "cinematic". Anything you want suppressed goes in the negative prompt, not the positive one; more positive clauses dilute the instruction rather than sharpen it. The Wan 2.2 Prompts page goes further.

Step 3: iterate cheap, then render once

This is where budgets quietly disappear. Prompt iteration is the part where most runs get thrown away, and every one of them is paid for at whatever output band you chose.

• Set the lowest output band you can still judge, and settle the prompt and motion there.
• Fix the seed so that changing the wording is the only variable.
• Re-run only the attempts you actually like at the size you will publish — see Wan 2.2 Resolution for why the band is part of the route here.

Judging a small render as a proxy for a large one has limits: artefacts invisible at a low band can be obvious once the same shot is rendered bigger. Use the small render for composition and motion, not fine detail.

Step 4: check before you build on it

Treat the first render as a test. Look for motion that reads as intended, artefacts around faces and hands, and whether the shot matches what you had in mind. If it keeps failing the same way, change the input or the route rather than adding clauses — a prompt cannot fix a source image that is too soft or a movement so large the model has to invent most of the frame.

What it costs

The credit cost of your chosen settings appears beside the form before you submit, so you can price a test before running it. Credit prices are on pricing.

Questions people ask

How long does a clip take? It depends on the route, resolution and duration. The form shows what you are calling before you commit.

Should I write the prompt in English? English is the safest choice if you are unsure; prompt interpretation is broadest there.

Can I use the result commercially? Check the licence for the model route you used; terms differ between versions.

Where do I start? Open the Wan workspace and choose the route from step 1.