Vocabulary flashcard videos: one Wan 3.0 scene per word, your captions

For a vocabulary flashcard video, make one two to four second Wan 3.0 scene per word and burn the word and its translation yourself with video captions cues. Alibaba's Wan 3.0 README (read 2026-10-05) lists text rendering in 12 languages and a 2 second lower bound is in Sume's docs, since wan-3.0 accepts 2 to 30 seconds. But a flashcard is the case where text must be right: a wrong accent or stroke teaches the wrong thing. So the model draws the picture, and your captions carry the language.

The card layout

Each card has the same four parts. Keep them fixed across the set, so a learner knows where to look.

• Scene (0 to 2 s): the thing or action, shown without text.
• Word (1 to 3 s): the target-language word, large, as a caption.
• Translation (2 to 4 s): the learner's language, smaller.
• Optional audio: record or synthesize the word separately; it is not part of the Wan clip.

Generate the scenes cheaply

A set of twenty words is twenty short jobs. Each is one wan-3.0 request with a duration of 3 seconds and a plain prompt. Start at 480p to check the idea, as the catalog lists 480p, 720p and 1080p. Keep every prompt in the same shape, so the set looks consistent: subject, action, lighting, camera. Use the same style reference image in each request, up to the 10 images the catalog allows.

Caption styles and scripts

The captions docs describe styles that resolve by script: Latin text defaults to slam, Korean to black-outline, and the page warns that Korean text on a Latin style burns as tofu. They also say language only hints speech-to-text and never selects the style or font. For a script the page does not document, run a one-card test and look at the result before you render the set. Pass style and font explicitly for Hangul sets, and keep the design overrides the same for every card.

Review and join

Have a speaker of the target language read all cards before you publish. A native check catches the errors that no tool will. Then join the cards with Timeline 1.0: the render accepts 1 to 200 video slots, and above 12 segments the default strategy chunks the render. Add a short fade between cards and keep each slot's duration equal to the card length, so the rhythm is steady. A final card with your name or channel closes the set.

Keep the set small at first. Ten words in a clip is a good unit for a learner and for your review: you can re-do one scene without touching the others, and a mistake costs one short generation. When the format works, scale it to a series, with a fixed order of beats and a fixed set of reference images, so every episode looks like part of the same show. Avoid putting faces of real people in the scenes unless you have their agreement, and keep each prompt about objects and actions.