参考生视频 AI

你的主体,按你的方式来。一张参考图就够用,多几张更好——MaxH3 的参考生视频生成器会在你生成的每一帧里锁住脸和身形。

20
参考生视频 AI
模型
参考素材*
0/15
提示词*
0 / 7000
画面比例
视频时长
4s
生成数量
分辨率
公开可见
开启后,你生成的内容可能出现在公开画廊中。
版权保护
禁止他人下载或复制你生成的内容。

请先输入提示词。

加进参考里原本没有的东西

参考素材锁住的是不能变的东西——这只瓶子,它的形状、它的玻璃、它的光泽。其余的都随你提:一个从没出现在画面里的模特,一个原本不存在的房间,五个能剪到一起的场景。写清楚谁走进来、做了什么,从开场的推镜到最后一个镜头,产品始终是同一件产品。

参考素材

A perfume bottle on a glossy black surface in amber light and golden smoke

提示词

A Luxury perfume bottle on glossy black surface, warm amber lighting, golden smoke, dark elegant background, soft reflections, cinematic close-up, slow push-in, photorealistic 4K. Confident woman in a burnt-orange satin dress, long dark hair, luxurious dark interior with golden lights, subtle smoke, cinematic medium shot, photorealistic 4K. Close-up of the same woman spraying perfume on her neck. Visible fragrance mist, warm golden light, natural hair movement, shallow depth of field, cinematic beauty ad. Same woman walking confidently toward camera in the luxury interior. Smooth tracking shot, reflective floor, warm golden lighting, subtle smoke, elegant cinematic look. Same perfume bottle centered on glossy black surface with amber crystals and golden smoke. Dramatic spotlight, dark luxury background, slow push-in, photorealistic 4K. Consistency: Keep the same woman, face, hair, skin tone, outfit, and perfume bottle across all scenes. Warm amber lighting, realistic skin and movement. No extra people, distorted hands, duplicate objects, text, or logos. Style: Luxury fragrance commercial, cinematic, photorealistic, 4K.

输出视频

The same bottle through a five-scene fragrance ad — the product push-in, a woman in a burnt-orange dress, the spray on her neck, a walk to camera, and back to the bottle

用 MaxH3 的参考生视频 AI 把多张图片合成一条视频

交给它一张脸、一个地点和一套服装——参考生视频生成器会把三者融进一条连贯的片子。不是幻灯片,也不是三段剪在一起:就是一个镜头,你的主体、你的场景、你的服装都已经在里面。发生什么、按什么顺序,由你的提示词决定。MiniMax H3 最多能读九张参考图,外加参考视频和音频。

参考素材

A studio portrait of a woman with long brown hair, used as the face reference
A modern garden with a reflecting pool, lavender beds and a low glass house, used as the location reference
A flat-lay of cream pleated trousers, a black sleeveless top, a tan crossbody bag, sunglasses and gold hoops, used as the wardrobe reference

提示词

Close up shot of woman with sunglasses on top of her head, gold hood earrings, is walking in the garden, she is lost and asks where everyone is and what's going on.

输出视频

The woman from the portrait, in the flat-lay outfit, walking through the garden from the location reference and speaking to camera

同一个角色,一个镜头接一个镜头

一条片子自身连贯,只是基本功。更难的是第二条:换一条街、换一个时段、换一句台词——走进画面的还是同一个人。保留这组参考素材再跑一次,镜头就能剪到一起,而不是看起来像选了三次角。一组片子就是这样变成一支广告、一集剧、一整场营销活动的。

参考素材

A model in a dusty-rose suit walking through a white studio, used as the talent reference
A lip-oil bottle in a magenta splash, used as the product reference

提示词

Ultra-realistic 8K luxury beauty commercial in vertical 9:16, shot on an ARRI Alexa 65 with Cooke S4/i prime lenses, cinematic HDR, premium editorial color grading in warm rose and blush tones, shallow depth of field, realistic optics, natural skin texture, luxury cosmetic campaign quality, smooth gimbal movement, photorealistic physics, no CGI look, no cartoon appearance. Use the exact same female model throughout every shot: an exceptionally beautiful, elegant Caucasian woman, approximately 175 cm tall, slender runway proportions, fair neutral skin with subtle natural freckles, luminous dewy complexion, expressive green-hazel eyes, thick defined brows, long natural lashes, soft naturally pink glossy lips, light brown hair styled in a sleek high ponytail, minimal luxury makeup with champagne highlighter and satin finish, wearing a tailored dusty-rose oversized blazer with matching wide-leg trousers, cropped bandeau top, nude pointed heels, confident editorial expressions. Use the exact same product throughout every shot: luxury transparent glass bottle labeled "MOIST HEIST LIP OIL", metallic silver cylindrical cap, vibrant glossy magenta lip oil inside, perfectly centered logo, premium reflections, identical proportions in every scene. 00:00–00:01 — Fade in from pure black. 00:01–00:02 — Hero product reveal. The lip oil bottle floats in mid-air while an explosive crown splash of glossy magenta liquid erupts around it. Crystal-clear droplets suspend naturally in slow motion with realistic refraction, sparkling highlights, and premium product lighting. Slow cinematic push-in. 00:02–00:03 — The model confidently walks toward the camera through a minimalist white luxury studio with elegant pink architectural panels. Smooth backward gimbal tracking, natural fabric movement, confident eye contact, polished editorial fashion campaign feel. 00:03–00:04 — Medium close-up. The model extends the MOIST HEIST LIP OIL directly toward the lens. Perform a smooth rack focus from her face to the bottle so the product becomes perfectly sharp while her face remains softly blurred behind it. 00:04–00:05 — Clean overhead shot. Elegant hands slowly unscrew the metallic cap. The applicator lifts smoothly from the bottle, creating luxurious glossy lip oil strings with realistic viscosity and fluid dynamics before separating naturally. 00:05–00:06 — Extreme beauty close-up. The plush doe-foot applicator gently glides across the lower lip, depositing a rich glossy magenta lip oil. Hyper-realistic pores, peach fuzz, soft beauty lighting, cinematic macro focus. 00:06–00:07 — Ultra-macro lip shot. Capture every natural lip line while the formula melts smoothly into the lips, creating an intense glass-like shine with bright specular reflections and a hydrated luxury finish. 00:07–00:08 — Fantasy product shot. The bottle floats inside swirling viscous magenta liquid while elegant glossy waves wrap around the glass. Suspended droplets sparkle like gemstones. Slow 180-degree cinematic orbit. 00:08–00:09 — Continue the hero shot. The bottle rotates gracefully as larger floating droplets orbit around it. Tiny beads of moisture shimmer on the metallic cap while the logo remains perfectly readable. Rich reflections, luxury cosmetic advertising quality. 00:09–00:10 — Extreme macro beauty finale. Fully hydrated glossy lips fill the frame with a flawless glass-like finish, multiple sparkling highlights, subtle confident smile, premium editorial beauty lighting.

输出视频

One model and one bottle through six shots of a beauty campaign — the product reveal, the walk to camera, the applicator, the macro finish

每个模型能接受哪些参考素材

MaxH3 上每个能读参考素材的模型,以及上传器实际执行的限制——几张图、几段视频、每段能多长、多少参考音频,以及最终输出什么。

模型参考图片参考视频参考音频输出时长分辨率
MiniMax H393 · 2–15s each · 15s total3 · 15s total4–15s768P · 2K
Seedance 2.53010 · 2–30s each · 30s total10 · 30s total4–30s480p · 720p · 1080p
Gemini Omni 1.1 Flash103—3–10s360p · 720p · 1080p · 4k
Wan 3.0 Prime105 · 1–15s each · 15s total5 · 15s total2–15s480p · 720p · 1080p
Wan 3.0105 · 1–15s each · 15s total5 · 15s total2–15s480p · 720p · 1080p
Seedance 2.093 · 2–15s each · 15s total3 · 15s total4–15s480p · 720p · 1080p · 4k
Seedance 2.0 Fast93 · 2–15s each · 15s total3 · 15s total4–15s480p · 720p
Seedance 2.0 Mini93 · 2–15s each · 15s total3 · 15s total4–15s480p · 720p
HappyHorse 1.19——3–15s720p · 1080p
HappyHorse 1.09——3–15s720p · 1080p
MiniMax H3 Max93 · 2–15s each · 15s total3 · 15s total5–15s480P · 768P
Kling V3 Omni71—3–15s720p · 1080p · 4k
Grok Imagine7——6–15s480p · 720p
Kling O171—3–10s720p · 1080p
Wan 2.755—2–10s720p · 1080p
Veo 3.13——8s720p · 1080p · 4k
Veo 3.1 Fast3——8s720p · 1080p · 4k
Veo 3.1 Lite3——8s720p · 1080p · 4k

怎么用 Omni Reference 做角色一致的视频

1
锁住不能变的东西

每个要保持不变的东西配一张干净的图——脸、服装、产品、场景。每张图只放一个主体,正面、光线均匀:之后每个镜头都要拿这张图来比对,参考图杂乱或昏暗,后面的相似度就会打折扣。

2
把这组素材放进 Omni Reference

把这组素材拖进提示词框。Omni Reference 不只收图片:MiniMax H3 最多读九张图来守住一张脸或一件产品,三段参考视频来交代运动,还有三个音频文件——可以在同一次生成里混用,每类各有上限。Wan 3.0、Seedance 2.5 和 Veo 3.1 也能读参考素材;每个模型接受什么,选择器里都标着。

3
剩下的交给提示词,然后复用这组素材

参考素材给不了的,都由提示词来补:动作、运镜、镜头顺序。最后加一句,点明哪些必须保持一致,然后运行;下一个镜头继续用同一组素材跑——几条片子能剪到一起,靠的就是这个。

准备好让同一个角色贯穿每个镜头了吗?

常见问题

当镜头里的某样东西已经存在,而且必须原样再出现时——一张脸、一件产品、一个场景。光靠提示词,每次运行都会重新发明一遍;参考生视频则从你交给它的素材出发生成。主体还是那个主体,场景、运镜和动作由你的提示词决定。

取决于模型,而且差别很大:MiniMax H3 最多九张图,Wan 3.0 十张,Seedance 2.5 三十张。有几个模型还能在图片之外读参考视频和参考音频——H3 各收三个。上面的表格列出了每个模型的上限,上传时选择器也会按上限来卡。

能,前提是模型声明支持。MiniMax H3 上的 Omni Reference 指的就是这个——图片、视频和音频放进同一次生成,每类各有上限:九张图片,三段每段 2 到 15 秒的视频(合计 15 秒),三个音频文件。图片用来锁定外观;视频则是交代动作或运镜最自然的方式。

每张图一个主体,正面、光线均匀,尺寸要让模型读得清——在 H3 上每条边在 256 到 5760 像素之间,宽高比在 0.4 到 2.5 之间。多样胜过重复:同一件夹克拍三个角度,比同一张照片传三遍教得更多。把脸、服装和场景拆成各自独立的参考,让每张图只担一件事。

保留这组参考素材再跑一次。单条片子内的一致性是模型的事;跨条的一致性则来自每次都喂同样的参考素材,只改提示词。正因如此,一组片子才能剪成一支广告或一集剧,而不是看起来像选了三次角。

MaxH3 上有十五个视频模型能读参考素材,包括 MiniMax H3、Wan 3.0、Seedance 2.5、Veo 3.1 和 Kling——并不是每个模型都支持,所以选择器只列出能用的那些。切换模型时参考素材会保留,你可以把同一组素材交给第二个模型跑,不用重新上传任何东西。

新账号会拿到免费积分,够在轻量一些的模型上跑一条短片,开始使用也无需绑卡。积分按秒计价,随模型和分辨率而变;运行前生成器会显示消耗。下载不带 MaxH3 水印的文件需要订阅套餐或购买积分包。

产品片——同一个瓶子在每个镜头里都得一模一样;广告系列——同一位模特贯穿一整季的帖子;角色内容——同一张脸撑起一整集;还有样机合成——把真实产品放进一个从没拍过的场景里。凡是画面里的东西必须两次都一眼认得出是同一个的,都适合。

主体只定一次,处处都是它

上传一张参考图,剩下的拍摄交给 MaxH3。