Make a short science animation for a child
Start with the child's age, one thing they like, and a question they can answer by watching what happens. Use any details the user already supplied. If an optional detail is missing, pick a reversible assumption and keep writing. A usable first episode can be just a 15–20 second pilot. For a longer episode, sketch roughly 60–75 seconds and adjust to the attention and action needed.
Write the experiment into the story
- State the intended takeaway in one sentence, then identify what the audience can see that supports it. Give the characters a problem the experiment can solve. Show a first attempt, a clue, one changed variable, and a fair comparison. Leave time to guess before showing the answer.
- Verify the underlying explanation and try the real experiment where practical. Keep important conditions constant. Qualify the observed result: “this bridge held this block” is stronger teaching than an unsupported promise about all paper bridges.
- For each shot, specify the teaching beat, duration, first-frame and end state, one primary action, camera, and spoken line or silence. Decide what the child should see even with the sound off. Keep comparisons at the same viewpoint and scale.
- End with a safe, achievable activity the child can do with an adult. Use ordinary materials and avoid claims, demonstrations, or instructions unsuitable for the age.
Copy the brief and storyboard template into the project. For a concrete example, read the eight-shot paper bridge episode.
Lock the visuals before spending GPU time
- Approve one canonical design for each recurring character, then make front, side, and back views when they turn or appear from multiple angles. Write down visible identity features and allowed performance changes. A tracking light may move or tint as part of a character's reaction; the character's body design stays fixed.
- Generate each shot's first frame from the same references with GPT Image. Review the whole storyboard as a contact sheet. Fix identity drift, inconsistent scale, and key props before sending frames to H3. For a continuity-dependent sequence, derive start, middle, and end guides from the same layout.
- Test the shot most likely to make the lesson unclear: the key experiment, full crossing, or other contact-heavy motion. Ask for one action, an explicit screen direction, and a locked camera if the audience needs the full setup. Watch the result across the entire shot; a convincing thumbnail is insufficient.
- Queue the rest only after that shot passes or the storyboard has a workable alternate. Keep approved shots. Change one meaningful thing per retry, and rerun only the defective shot. Use the preceding shot's accepted tail frame when the next shot starts in the same scene.
For generation, queue recovery, Mandarin audio, and delivery, read Production in ComfyUI.
Review the final cut
Watch at normal speed with sound, with sound muted, and around hard joins frame by frame. Check that the result of each experiment remains visible long enough, motion does not jump back or repeat unintentionally, and subtitles match what is actually spoken. If a generated shot contradicts the lesson, shorten it, split it into simpler shots, or use a clear real demonstration. Export a shareable MP4 and retain the storyboard, approved takes, subtitles, and a record of generation versus export resolution. If possible, ask the child to explain the result back in their own words.