🎬 视频创作快速上手
视频生成的胜负手不在模型，在顺序：先把首帧定死，再让它动起来。这一篇讲清公式、四种生成方式和模型分工。
1. 第一段视频，五步走完
- 打开创作台，切到「视频」。
- 选生成方式（下一节细说）：没素材就文生视频；已经有一张满意的图，就图生视频。
- 选模型、时长（5–30 秒）、分辨率（480P / 720P / 1080P）和画幅。
- 写提示词：重点写「怎么动」，价格会先亮出来，确认再生成。
- 视频是后台生成的——关掉页面、锁屏都不断，完成后自动存入我的作品。
想要结果最稳：先在图片区生成一张定妆首帧，再用图生视频让它动——人物、服装、场景全被首帧钉住，视频只负责表演。
2. 视频提示词公式
主体 + 连续动作 + 场景变化 + 镜头运动 + 景别 + 光线 / 氛围 + 风格 + 节奏
例：清晨的山间公路上，一位骑手俯身加速，摩托车掠过弯道，路边的松树向后飞掠。镜头低角度跟拍、轻微手持晃动，中景，朝阳逆光勾出金色轮廓，写实电影感，节奏紧凑。
- 只写看得见的动作：抬头、转身、走向镜头、挥手——越具体越稳；「很有感觉」这类抽象词模型接不住。
- 一个短镜头只安排一个主要动作：动作塞太多，画面就会跳变、融化。
- 镜头运动要喊出来：固定、推进、拉远、横移、环绕、跟拍、手持——不写，模型默认慢吞吞地飘移；想要快节奏，就把「快速跟拍、急推急拉」写进提示词，并明说节奏。
- 图生视频别重复描写外观：首帧里已经有的人物长相、服装、场景，提示词里只写「接下来怎么动」。
- 声音：每条视频带声音开关——需要对白与音效就开着，想后期自己配音就关掉出无声片。
3. 四种生成方式怎么选
| 方式 | 适合什么 | 提示词重点 |
|---|
| 文生视频 | 从零探索画面与动作 | 主体、场景、动作、镜头、风格全写清 |
| 图生视频 | 首帧（人物 / 产品 / 场景）已经定好 | 只写动作、镜头运动和环境变化 |
| 参考生视频 | 要同时锁住多个人物 / 物体 | 给每张参考图派活：谁是人物、谁是场景；在分镜词里用 @人物 / @场景 锚点指认，别用图号 |
| 视频编辑（Wan 2.7） | 在已有的一段视频上换内容、换风格 | 写清保留什么、改什么；按源视频时长计费 |
还有一个隐藏好手：承接上一镜——画布与导演台里可以截取上一镜的尾帧当下一镜的首帧，两镜之间严丝合缝，长故事就是这样一镜一镜接出来的。
4. 视频模型阵容
| 模型 | 定位 | 选用建议 |
|---|
| Seedance 2.5 Pro Spicy（默认） | 旗舰，动作与质感的第一选择 | 重要镜头、人物表演——默认就是它 |
| Wan 2.6 Spicy | Spicy 线的性价比之选 | 批量出片、试动作时先用它探路 |
| Seedance 2.5 | 标准旗舰，多参考图（最多 9 张） | 多人物、多素材的常规题材 |
| Wan 3.0 · 通义万相 | 最高 1080P，支持参考视频 | 要高清成片、要照着一段参考视频重绘时 |
| Kling 3.0 · 可灵 | 运镜与人物表演细腻，720P / 1080P | 讲究镜头语言的段落 |
| Wan 2.7 · 视频编辑 | 直接改源视频本身 | 换风格、换元素的 vid2vid 活儿 |
| Grok Imagine 1.5 | 便宜量大，15 点 / 秒起 | 草稿、氛围镜头、大批量尝试 |
5. 让画面稳下来的四个习惯
- 首帧要干净：主体完整、少遮挡、背景别太杂——首帧的质量就是成片的上限。
- 动作由小到大：人物镜头先跑通轻微动作，再逐步加大幅度；一上来就大开大合最容易崩。
- 抖了就加约束词：「镜头平稳、运动连贯、主体结构稳定」，写进提示词立竿见影。
- 长故事拆单镜头：单次调用最长 30 秒；更长的故事拆成单个短镜头逐个生成，再去「一键成片」或导演台合成——这正是下一篇的内容。
🎬 Video Creation — Quick Start
Video generation is won on sequence, not on model choice: lock the first frame, then set it in motion. This guide covers the formula, the four generation modes and what each model is for.
1. Your first video in five steps
- Open the Studio and switch to Video.
- Pick a mode (details below): nothing prepared → text-to-video; a still you're happy with → image-to-video.
- Choose a model, a duration (5–30s), a resolution (480P / 720P / 1080P) and an aspect ratio.
- Write the prompt around how it moves. The price shows before you confirm.
- Videos render in the background — close the page or lock your phone, it keeps going and lands in My Works when done.
For the steadiest results: generate a portrait still first, then animate it with image-to-video. Face, outfit and setting are pinned by the first frame; the video only has to perform.
2. The video prompt formula
Subject + Continuous action + Scene change + Camera move + Shot size + Light / Mood + Style + Pacing
Example: On a mountain road at dawn, a rider leans in and accelerates through a bend, pine trees streaking past. Low-angle tracking shot with a slight handheld shake, medium framing, golden backlight rimming the silhouette, realistic cinematic look, tight pacing.
- Only visible actions. Looks up, turns around, walks to camera, waves — the more concrete, the steadier. Abstract words like “full of feeling” give the model nothing to hold.
- One main action per short shot. Cram in too much and the picture jumps and melts.
- Call the camera move. Static, push-in, pull-out, pan, orbit, tracking, handheld — unwritten, the model drifts slowly by default. For fast pacing, say so: fast tracking, snap zooms, and name the rhythm.
- In image-to-video, don't re-describe the look. Face, outfit and setting are already in the first frame; the prompt only says what happens next.
- Sound: every video has a sound toggle — on for dialogue and effects, off if you'd rather score it yourself later.
3. Choosing between the four modes
| Mode | Best for | Prompt focus |
|---|
| Text-to-video | Exploring picture and motion from zero | Subject, setting, action, camera and style — everything |
| Image-to-video | A first frame (person / product / scene) already decided | Action, camera moves and environmental change only |
| Reference-to-video | Locking several people or objects at once | Give every reference a job — who is the character, what is the scene; in shot prompts, point with @Character / @Scene anchors, not image numbers |
| Video Edit (Wan 2.7) | Changing content or style of an existing clip | What to keep, what to change; billed by source length |
One more hidden gem: chaining — in the Canvas and Director Studio you can take the previous shot's last frame as the next shot's first frame, so cuts join seamlessly and a long story is built one shot at a time.
4. The video model lineup
| Model | Position | Reach for it when |
|---|
| Seedance 2.5 Pro Spicy (default) | Flagship — first choice for motion and texture | Hero shots and character performance; it's the default for a reason |
| Wan 2.6 Spicy | The value pick on the Spicy line | Batch runs and testing moves before you commit |
| Seedance 2.5 | Standard flagship, up to 9 reference images | Multi-character, multi-asset everyday subjects |
| Wan 3.0 | Up to 1080P, supports reference video | HD finals, or redrawing along a reference clip |
| Kling 3.0 | Refined camera work and performance, 720P / 1080P | Passages where camera language matters |
| Wan 2.7 Video Edit | Edits the source clip itself | Restyles and element swaps (vid2vid) |
| Grok Imagine 1.5 | Cheap and cheerful, from 15 pts/sec | Drafts, mood shots, bulk experiments |
5. Four habits for stable footage
- Clean first frame. Complete subject, little occlusion, uncluttered background — the first frame caps the quality of everything after it.
- Grow the motion. Get small movements working on character shots first, then scale up. Big moves from a cold start are where things break.
- Add constraint words when it shakes: “steady camera, continuous motion, stable subject structure” — it works immediately.
- Split long stories into single shots. One run caps at 30 seconds; generate shot by shot, then assemble in One-Click Film or the Director Studio — which is exactly the next guide.