AI video for lessons: Seedance 2.5 demo clips for teachers

Yes, Seedance 2.5 can make short lesson videos, and ByteDance names teaching as a use. Its launch post says the model can turn the historical context, characters and storylines behind a lesson into more vivid visuals, and help teachers produce instructional videos by turning scientific principles, historical events and experimental procedures into dynamic demonstrations. On Sume you call it as seedance-2.5 through the Video Router: 4 to 30 seconds per clip at 480p, 720p or 1080p, with image, video and audio references.

The model makes the picture. You still own the lesson: the facts, the order, and a check that the clip shows what you mean. This page covers what a teacher can ask for, where the limits sit, and how to turn clips into one lesson video.

What does ByteDance say Seedance 2.5 can do for education?

The ByteDance Seed launch post (dated 2026-07-31) has an education passage. It says the model has begun to enter real learning settings, with an example from the Doubao Learning app's classroom scenario, and that it lowers the barrier to producing educational material while allowing flexible customization of content.

It also states its own limit: the Seed team says there is still room for improvement in the physical plausibility of complex motions and in scenes where several subjects interact. For a lesson, that matters most for experiments and anything with a correct physical outcome.

What can you request on Sume?

The Video Router doc lists seedance-2.5 at 4 to 30 seconds and 480p, 720p or 1080p. The Video generation doc adds that audio and video references are honored by the Seedance 2.x models, and that a request is either first/last frame image-to-video or reference-to-video; if you send both, frame_images wins.

A practical lesson clip is one idea per request: one diagram-like scene, one motion, 8 to 20 seconds. A 30-second request is better kept for a narrative scene such as a historical street. Start at 480p while you check the idea, then rerun the winner at a higher resolution. Sume does not accept a seed, so a rerun is a new take, not a repeat.

import os, requests

r = requests.post(
    "https://api.sume.com/v1/videos",
    headers={
        "Authorization": f"Bearer {os.environ['SUME_API_KEY']}",
        "Idempotency-Key": "lesson-water-cycle-001",
    },
    json={
        "model": "seedance-2.5",
        "prompt": "Classroom explainer style: rain falls on hills, runs into a river, "
                  "the sea evaporates into clouds. One continuous slow camera move.",
        "duration": 12,
        "resolution": "480p",
        "aspect_ratio": "16:9",
    },
)
print(r.status_code, r.json())

How do you turn clips into one lesson video?

Timeline 1.0 takes one audio spine plus ordered video[] slots and returns one MP4. Your narration is the spine; each Seedance clip is a slot with a start, a duration and an optional fade or wipe. POST /v1/timeline-1.0/plan is an unbilled preflight that returns the duration, segment count and an estimated cost before you render.

Slots must be this workspace's media.sume.com files, so check that the finished job lists an artifact URL on that host (Jobs and results) before you build the slot list. Timeline's public rate is $0.10 per rounded-up output minute, and the doc says to confirm it live in the catalog.

For on-screen vocabulary, Video captions can burn authored cues with text, start and end, which skips speech-to-text. That is the safer route for key terms, because a speech-to-text pass on a silent clip fails with caption_no_speech.

What should a teacher check before sharing?

Treat each clip as an illustration. Look at the first and last second, and any frame with hands, text or a moving object, and confirm it matches the lesson. If a clip is wrong, change the prompt and rerun, rather than trimming around the error.

Label generated visuals as generated when you present them. A single-subject, single-motion clip is the pattern the vendor's own limits note points to.