Wan 3.0 视频模型

一次生成三十秒——足够走完一个不断的运镜、说完一整段对白,或者让一个场面层层推进,而不是每五秒就重来一次。或者带上最多二十张图片、片段和音频作为参考,让它们在一条十五秒的镜头里始终保持一致。

可选
可选
50

Wan 3.0 的核心能力

  • 30 秒生成: 一次运行最长三十秒,不用把片段接起来。
  • 全模态参考: 每次生成二十个参考——图片、片段和人声。
  • 同步声音: 声音与画面在同一遍里生成。
  • 首帧与尾帧: 定下镜头从哪儿开场、落到哪儿,中间由模型补齐。
  • 角色一致性: 脸、衣服和道具在一条长镜头里都稳得住。
  • 精准视频编辑: 只改已有片子里的一个元素,画面其余部分纹丝不动。

30 秒生成

一次运行就能出到三十秒。不用拼接,也不用去凑两条几乎能对上的片子之间那个剪辑点。沿走廊推进的一镜到底还是一整段不断的运动;一句台词会在它开始的那个镜头里说完。你直接报出确切的长度,而不是从一个短短的菜单里挑——二到三十秒之间的任意整数秒——成本随之线性变化,所以同分辨率下 6 秒的片子确实是 12 秒的一半。没有理由往上取整。参考生成是唯一的例外:它的上限是十五秒,因为你的参考素材和生成出来的成片在上游共用同一个三十秒的额度。

全模态参考

一次最多挂二十份素材——十张图片、五段视频和五个音频文件。模型按你添加的顺序读取,所以就照你打算引用它们的顺序来排:先角色,再是要复刻的运镜,然后是一段人声。同一张脸、同一件产品、同一个房间,在这条镜头的每一个画面里都被稳住—— 同样是 参考生视频 的工作流,只不过它跑在一次生成的内部,而不是跨越多条独立片子。

INPUT

Girl A — outfit

Girl B — outfit

Girl A — face

Girl B — face

PROMPT

A cinematic fashion blockbuster featuring two female protagonists, with an extremely fast pace and pure jump cut editing logic. Shot with the Arri Alexa 65 and Panavision anamorphic lenses, it captures the texture of Kodak Vision3 500T film, boasting 8K resolution and photo-realistic quality, with no text. The facial features of the characters are fully locked onto Girl A @Image1 and Girl B @Image2. [0-3 seconds: rapid entrance and contrast] (hard cut) A stark white minimalist space is sharply shadowed by intense side lighting. Left screen/left side: Girl A wears a black deconstructed leather suit, her eyes cold and elegant under the cold light, swiftly turning to look directly at the camera. Right screen/right side: Girl B wears a flowing champagne gold silk dress, her eyes languid under the warm light, swiftly turning to look directly at the camera. Real film graininess. [3-6 seconds: extreme close-up of material] (hard cut) Extremely close-up macro shot. Girl A's reference outfit @Image3: Cold light casts on the sharp creases of black leather and metal accessories, presenting a cold and high-end luster. Girl B's reference outfit @Image4: Warm light outlines the lines of her collarbone, with silk fabric flowing like water waves, and the skin exhibits a hyper-realistic translucent and hydrated texture, with visible fine hairs. [6-9 seconds: Tension Interlace (High-Speed Photography)] (Hard Cut) The two characters walk towards each other in a pure white space and pass by each other. High-speed photography (1000fps slow motion). The toughness of black leather collides violently with the softness of champagne silk in an instant of interlace, with skirt hems and hair flying in the air. In the zero-point-one second of interlace, their eyes meet sharply. [9-12 seconds: Back-to-back pressure] (Hard cut) The two stand back-to-back. Strong rim lighting simultaneously outlines their facial contours and highlights the golden edges of their hair. They simultaneously turn their heads quickly and look directly at the camera. Their expressions exude a strong sense of pressure, professionalism, and the confident aura of top supermodels. [12-15 seconds: Relaxation and freeze] (Hard cut) The two people stand side by side, slightly tilting their heads to look at each other, revealing a contagious, relaxed, and confident smile. The camera slowly zooms out, revealing the perfect proportions of the two people in the grand light and shadow space, and the image gradually fades in a high-end glow. Camera movement: Quick handheld panning, rapid zooming, high-speed photography in slow motion, pure hard cuts, without fancy visual effects. Color: High-contrast black, white, and gray tones, embellished with cool metallic luster and warm champagne gold, with film-grade color grading.

OUTPUT VIDEO

0:15

同步声音

声音是和画面一起生成的,而不是事后叠上去的,所以对白落在嘴上,脚步声落在脚步上。用写「该看到什么」的那份心思去写「该听到什么」——模型也在读说明里的这一半,而且不额外收费。想要一段自己来配乐的无声素材时,把它关掉即可。

首帧与尾帧

把你想开场的那张图和想落到的那张图交给 Wan 3.0,它会把两者之间的运动搭出来。 这就是把 图生视频 的两头都钉住、而不是只钉一头的情形。当两个端点都已确定、只剩中间那条路要走时特别有用——一件必须停在特定位置的产品、一次接向下一个镜头的匹配剪辑、一个必须严格落在品牌手册指定位置的 logo 组合。

INPUT

First frame

Last frame

PROMPT

【One-Take, Slow Aesthetic Camera, No Cuts, 15-Second Long Take — Rock-Color Painting Mineral Pigment Texture Meets 3D Rendering, Azure Blue × Gilded Gold Color Clash】 Global Settings: Chinese rock-color painting aesthetic with visible mineral pigment grain texture and gold leaf accents. Strong contrasting palette of azure blue and gilded gold. The scene is West Lake at dawn in thin mist: the lake surface rendered in matte azure-blue washes, golden morning light piercing through cloud layers to form Tyndall light beams. In the distance, the stone pagodas of Three Pools Mirroring the Moon and Leifeng Pagoda emerge faintly through the fog. In the foreground, lotus leaves and lotus blossoms. The central figure is Xizi, the Water Goddess of West Lake: rendered in rock-color painting texture, with an oval face, willow-leaf eyebrows, phoenix eyes, a vermillion floral mark at the center of her brow, skin like mutton-fat jade glowing with a soft luminescence, and a pale golden halo floating behind her head. Her hair is swept up in a high bun adorned with a single gilded gold lotus hairpin. Her skirt and flowing shawl are composed entirely of living lake water, gradient azure blue, with a shimmering, rippling surface. Audio & Sound Design: Background features gentle guzheng and bamboo flute. Ambient sound includes subtle water ripples and early morning birdsong. Starting at the 8-second mark, an ethereal, tender female voice slowly recites: "If I compare West Lake to the fair Xizi—" The second line follows during the glance-back segment: "Light or heavy makeup, she is always beautiful." The delivery is slow, with a slight reverb, leaving a lingering resonance. [0-3s: A Water Drop] Extreme close-up, macro shot. At the edge of a lotus leaf at dawn, a water droplet gathers, plump and full, reflecting golden morning light, then slowly falls. The camera follows the droplet with a tilt down as it drops into the lake, triggering a ring of gold-and-blue ripples that expand outward in concentric layers. [3-7s: Water Takes Shape] At the center of the ripples, the lake water slowly rises and gathers into the form of the goddess Xizi: the water transforms into her azure-blue skirt and translucent shawl, with fine streams of water flowing and swirling around her body. She rises slowly from the center of the lake, her eyes gently closed, then slowly opening. The gilded gold lotus hairpin glows with a faint light. The camera slowly orbits her half-lowered eyes and the skirt woven from flowing water. [7-11s: One Step, One Lotus] She walks barefoot across the lake surface, each step sending out a ring of golden ripples from beneath her feet, lotus blossoms blooming one after another within the ripples. The camera pulls back to a medium-wide shot: the thin mist parts, and in the distance the three stone pagodas of Three Pools Mirroring the Moon light up with warm glow, Leifeng Pagoda emerges in the morning light, and willow branches along the shore brush the water. At the 8-second mark, the female voiceover begins: "If I compare West Lake to the fair Xizi—" Her shawl trails a long streak of gold-and-blue water light behind her. [11-15s: The Glance Back · Always Beautiful] She stops and looks back over her shoulder toward the camera. The camera slowly pushes in to a half-body close-up: morning light falls across her face, her water-woven shawl drifts gently at her side, a few luminous water droplets fall from her hair tips, dissolving into fine particles of light mid-air. The voiceover continues with the second line: "Light or heavy makeup, she is always beautiful." She smiles softly. Behind her, Leifeng Pagoda stands in the morning glow, with Three Pools Mirroring the Moon hazy in the distance. In the final second, the frame freezes completely on her glance-back close-up, locked into a perfect cinematic poster composition.

OUTPUT VIDEO

0:15

角色一致性

脸在一条长镜头里始终保住自己的五官,不会在镜头之间慢慢漂走。衣服、道具和场景陈设也都留在原处。参考就是你手里那根杠杆:把必须保持不变的脸、产品或房间交给模型,它会把它们贯穿整次生成。

INPUT

Product image

PROMPT

The lip gloss product image @Image1, generate a model lip color test video, 15 seconds, portrait orientation, highlighting the lip color effect

OUTPUT VIDEO

0:15

精准视频编辑

用大白话描述改动,Wan 3.0 就只重写已有片子里的那一部分——去掉一个物体、重新给场景打光、改变某个人正在做的事——而其余一切都保持拍摄时的样子。你也可以选定一个时间区间只重生成那一段,或者把已有素材往后延长。它覆盖的正是时间线类工具完全做不到的那部分 AI 视频编辑:不是剪你拍到的东西,而是改画面里的东西。

INPUT

Source clip

PROMPT

In Video 1, the man on the left is removed, and the surrounding background is used to fill the space naturally, while the rest of the characters, actions, and images remain unchanged.

OUTPUT VIDEO

0:05

开始使用

怎么用 Wan 3.0 创作

1
选好 Wan 3.0 和你的起点

一句提示词、一张首帧,或者一组参考。 如果从一帧开始,请到 图生视频 上传。

2
把所有必须保持一致的东西挂上去

角色图片、要复刻的一段运镜、要对上的一段人声。十张图、五段片子、五个音频文件。

3
定下长度和分辨率,然后反复试

还在找镜头的阶段就留在 480p。一次只改一处,再跑一遍,最后用 1080p 重跑。

准备好做第一条视频了吗?

常见问题

Wan 3.0 是阿里巴巴的视频生成模型,2026 年发布。它把文生、图生和参考驱动的生成收在同一个模型里,在出画面的同时出声音,并且能按一句大白话指令重写已有素材的一部分。

凡是需要跑得久、还得撑得住的都行——产品演示、对白场景、UGC 风格广告、短篇叙事、音乐演出片段。三十秒的上限加上参考控制,让它更适合完整的场面,而不是那种你打算之后再剪到一起的零散镜头。

在一次运行里最多挂二十份素材——十张图片、五段视频、五个音频文件——并在提示词里逐一直接点名。某张图定住角色,某段片子提供运镜,某个音频文件提供人声。

用干净、光线充足、主体占满画面的图片,并且在不同次运行之间把对这个主体的描述写得一模一样。要明确点名每一份素材,别指望模型自己去猜哪个是哪个。在同一次运行里少换场景,也有助于参考稳住。

能。描述改动,它就会重写那一部分,而画面其余仍是拍摄时的样子——去掉一个物体、重新打光、改变某人正在做的事。它也能延长已有素材,输入和输出加起来上限为三十秒。

免费——赠送的积分可以覆盖 Wan 3.0 的运行。三十秒的 1080p 片子在这里属于比较重的渲染,所以试提示词的阶段先从 480p 起步。

用 Wan 3.0 做视频

一次生成三十秒,配乐已经在里面了——外加最多二十个参考,让这条镜头从第一帧到最后一帧都保持一致。