Reading note
This course is for readers who can already express requirements in natural language and want greater control over video results. Chapter 1 introduces professional generation platforms and tutorials; Chapter 2 covers refinement after generating an initial result and ways to let AI operate tools directly; Chapter 3 provides existing workflows beginners can follow.
Resources checked: October 31, 2026. Research verification and the three chapters are complete; connection and generation instructions have not been tested in local software, paid calls or output acceptance checks. Chapter 3 collects creators' published cases and tutorials, distinguishing their demonstrations from actual reproduction by this site. Research choices and workflow file records are in the [research verification notes](VIDEO_CREATION_RESEARCH.md).
1. Distinguish models, platforms and workflows
| Input and task | Main controls | A first exercise |
|---|---|---|
| Text-to-video | Subject, action, scene and cinematography | One subject, one action and one short shot |
| Image-to-video | Initial appearance and composition, followed by motion | Add slow movement to a product or scene image |
| First and last frames | Starting point, endpoint and the transition | Two similar states of the same object |
| Multiple references | Characters, props, actions, camera movement or sound | Assign one purpose to each reference |
| Combining clips | Order, pacing, transitions and audiovisual continuity | Assemble three checked shots into a short video |
2. Three generation platforms
Based on the independent evaluations, creative controls and course needs available at verification time, this chapter focuses on three routes. Their order is a learning sequence. Text-to-video, image-to-video and editing have different leaderboards; they are not merged into an undefined overall ranking.
| Platform and focus model | Reasons and suitable tasks | Official entry points and tutorials |
|---|---|---|
| Jimeng / Dreamina: Seedance 2.5 | First overall choice for this course, suitable for reference-based multishot content and subsequent instruction editing. Ranked first in the checked Artificial Analysis image-to-video list and second in video editing | Jimeng, model overview, official prompt guide |
| Tongyi Wan: Wan 3.0 | Focus on generating shots from text and editing afterward. Ranked first in both checked Artificial Analysis text-to-video and editing lists | Alibaba Cloud model overview and entry point |
| Hailuo: MiniMax H3 | Focus on image-to-video and multimodal references. Ranked first in the checked Arena image-to-video list and third in Artificial Analysis text-to-video | Hailuo, official MiniMax video tutorial |
3. Where to start learning each platform
| Platform | First-choice tutorial | Supplement |
|---|---|---|
| Jimeng / Dreamina: Seedance 2.5 | Official Seedance 2.5 prompt guide (Chinese) | Official Xiaoyunque Seedance 2.5 manual (Xiaoyunque interface) |
| Tongyi Wan: Wan 3.0 | Official Wan 3.0 video prompt guide (Chinese) | Official Wan 3.0 video generation guide (API) |
| Hailuo: MiniMax H3 | Official MiniMax H3 tutorial and workflow center | Official MiniMax H3 video generation guide (API) |
4. Learn the main video controls
Control actions before increasing complexity. For a short shot, specify who does what, how the action starts and ends, and whether the camera moves. If the image already contains the subject, repeated appearance adjectives can obscure the action that matters. For multiple shots, explain each shot's role separately. See the Adobe prompt tutorial.
Give references separate roles. A character image supplies identity, a product image appearance, a video action or camera movement, and audio dialogue or rhythm. When references conflict, reduce them first and inspect the result. References are conditioning inputs; outputs still need inspection by frame or segment.
First and last frames constrain the ends; the transition still needs design. If the first frame is a closed box and the last an open box, describe the opening action. Large differences in scene, angle or objects leave more changes for the model to invent. Start with two images having similar framing and composition. Flow first/last-frame documentation