Skip to main content
AI video model
MiniMax
Hailuo

MiniMax H3

MiniMax H3 offers 768P and 2K output with text, first/last-frame, and image/video/audio reference scenes. Reference-video duration and every input image from image 6 onward affect pre-task authorization.

Scene
Text / first-last frames / references
Output
768P / 2K
Duration
4–15 seconds
Credits
32–406
30-second overview
  • Two output tiers: 768P and 2K
  • Either a first frame or a last frame can serve as image-mode input
  • Accepts up to 9 images, 3 videos, and 3 audio clips
  • Reference-video duration and images from image 6 onward affect preauthorization

Popular searches: How do I use MiniMax H3? · How much does MiniMax H3 cost? · What reference media does MiniMax H3 support?

Examples

Cinematic scene

Main Subject: Young American woman, early 20s, athletic build, one-piece swimsuit with a loose cover-up, hair in a tight braid, confident playful personality. Location: Small outdoor community pool, sunny midday. Diving board at the deep end, lounge chairs, a lifeguard tower in the background, clear blue water. No large crowds, no commercial branding. Visual Style: Ultra-realistic documentary realism, energetic candid feeling, bright natural sunlight. Camera Style: Early 2000s DV camcorder, handheld with water-splash blur occasionally on lens, exposure blown out in bright sun, soft grain. No stabilization. Timeline (30 sec): 00:00–00:03 → She climbs the diving board ladder, then walks to the edge. 00:03–00:06 → She turns to camera saying "okay don't judge my form," laughing nervously. 00:06–00:09 → She bounces once, then dives cleanly into the water. 00:09–00:12 → She surfaces, wiping water from her face, grinning. 00:12–00:15 → She swims to the pool edge, pulling herself up onto the deck. 00:15–00:18 → She wrings out her hair, then grabs a towel from a lounge chair. 00:18–00:21 → She dries off, glancing back at the diving board. 00:21–00:24 → She decides to go again, walking back toward the ladder. 00:24–00:27 → She climbs up, this time attempting a small jump with a tuck. 00:27–00:30 → She surfaces laughing, looking at camera saying "okay that one was way better!" Audio: Water splashing, distant pool chatter, wind, birds. Her dialogue as noted above. No music. Goal: A playful, confident pool moment energetic, fun, authentic.

Cinematic scene

电影级真实人像短片,总时长必须为15秒整,一镜到底,全程无切镜、无跳切、无转场、无突然换镜头。严格使用 {{Mixed 1}} 中的人物作为唯一主体,全程保持与参考图完全一致的人物身份、五官比例、脸型、眼睛、鼻子、嘴唇、眉毛、刘海、发型、肤色和整体气质,不重新设计人物,不改变长相,不改变年龄,不生成新的脸。人物始终保持参考图中的三分之二侧脸构图和朝向,全程绝不转头、绝不转回正面、绝不望向侧方大角度,仅允许自然呼吸、轻微眨眼、轻微眼神移动、极轻微嘴角变化以及发丝自然摆动,人物动作始终小于镜头运动。 真实手机后置主摄拍摄,RAW原片摄影质感,9:16竖屏,摄影师距人物约20-25厘米,全程超近距离怼脸拍摄,以三分之二侧脸特写为主,镜头始终围绕人物侧脸缓慢运动,不拉远,不改变人物姿态,不突然转头,不低头,不抬头,不改变参考图构图。 最高优先级:完整保留真实皮肤质感。超近距离清晰可见细微毛孔、鼻翼毛孔、脸颊真实纹理、额头轻微纹理、卧蚕纹理、眼下细纹、眼下轻微卡粉、鼻翼轻微卡粉、嘴角底妆粉感、极浅闭口、一两颗极淡小痘痘、轻微痘印、淡淡小痣、轻微泛红、T区轻微出油、鼻尖自然皮脂高光、脸颊真实水润反光、真实唇纹、下唇水光反射、真实绒毛、真实碎发和发丝边缘,所有细节自然、高清、干净,不磨皮、不美颜、不塑料,不降低人物颜值。 场景始终为夜晚高级咖啡馆窗边。背景严重虚化,仅保留暖黄色吊灯散景、奶油色墙面、木质桌面边缘、模糊咖啡杯轮廓、玻璃窗轻微反光以及窗外蓝色、紫色、橙色、红色霓虹和车灯散景。暖黄色吊灯作为主光,桌面形成柔和暖色补光,窗外蓝紫色霓虹形成轻微边缘光,玻璃窗形成极轻微反射光。曝光稳定,不过曝,不过暗,不频闪。 0-3秒:镜头从 {{Mixed 1}} 参考图当前构图开始,保持超近距离三分之二侧脸,焦点锁定眼睛和鼻梁。人物自然呼吸,轻轻眨一次眼,眼神缓慢看向镜头再轻微移开,嘴唇自然放松。镜头极慢推近并轻微横向滑动,展示脸颊毛孔、眼下纹理、鼻梁高光、鼻翼卡粉、真实碎发和发丝边缘。 3-6秒:镜头沿脸部缓慢滑向鼻尖,再滑向嘴唇附近,进行轻微弧线绕拍。人物仅有极轻微眼神变化,不转头,不改变姿态,一缕碎发随呼吸自然轻轻摆动。暖光照亮鼻梁、鼻尖、脸颊和嘴唇,桌面反射补光保持鼻翼、嘴角、眼下细节始终清晰可见。 6-9秒:镜头缓慢绕到侧脸另一角度,但始终保持参考图中的侧脸比例,不改变人物朝向,人物依然是三分之二侧脸对镜头,绝不转正、绝不转向侧方。人物偶尔轻轻眨眼,保持自然呼吸,重点表现卧蚕纹理、下眼睑细纹、睫毛、毛孔反光、鼻翼毛孔、嘴唇水光和真实皮肤纹理。 9-12秒:镜头轻微下移,突出嘴唇、下巴、鼻翼和脸颊,再连续缓慢回到眼睛,全程保持超近距离怼脸距离。人物保持参考图神态,仅有自然呼吸、极轻微眼神变化和轻微眨眼,嘴角自然放松,不做任何夸张表情,不微笑露齿,不转头。唇纹、鼻翼卡粉、毛孔、卧蚕纹理、眼下细纹、轻微痘印、底妆粉感始终保持高清可见。 12-15秒:镜头进行一次极轻微弧线收镜,缓慢推近,最后停留在眼睛、鼻梁、半边嘴唇构图。人物最后轻轻抬眼看向镜头,眼神自然、安静、有吸引力,不改变参考图整体神态和侧脸角度,不转头,不露齿微笑。最后一帧完整保留眼睛、卧蚕、睫毛、毛孔、鼻尖皮脂高光、脸颊纹理、真实唇纹、轻微卡粉、真实碎发和发丝边缘,全程无切镜、无跳切、无转场,全片时长满15秒。 整体风格:真实电影级手机摄影,夜晚咖啡馆窗边超近距离怼脸抓拍,高级杂志感,真实肤质,真实景深,克制自然,人物身份始终与 {{Mixed 1}} 参考图保持一致,不重新生成新的脸,不改变五官,不改变气质。 负面提示词:重新生成人物,人物变脸,改变五官比例,改变脸型,改变眼睛,改变鼻子,改变嘴唇,改变刘海,改变发型,改变肤色,改变年龄,改变身份,普通路人脸,AI脸,CG感,塑料皮,陶瓷皮肤,蜡像皮肤,过度磨皮,过度美颜,皮肤无毛孔,皮肤无纹理,滤镜感,浓妆,假睫毛,大幅转头,转头转正,转向侧方,露齿大笑,大幅低头,大幅抬头,大幅表情变化,张嘴过大,人物剧烈动作,镜头乱晃,突然拉远,突然切镜,跳切,转场,视频时长不足15秒,背景抢主体,咖啡馆全景,字幕,Logo,水印,可读文字,直播补光灯,环形灯,棚拍光,频闪,过曝,欠曝,低清晰度,低质量,AI运镜,眩晕感。 No BGM

Cinematic scene

Format: 16:9 Concept: "Getting Ready to Meet My Friends" Create a single continuous self-filmed vlog recorded on a handheld MiniDV camcorder. The camera is always held by the girl, creating a natural first-person selfie perspective with subtle hand movement, realistic autofocus adjustments, slight exposure changes, and authentic walking motion. The footage should feel completely spontaneous with no cuts, transitions, or cinematic camera moves. Visual Style: Authentic early-2000s MiniDV camcorder aesthetic with soft detail, light tape grain, natural skin tones, gentle highlight bloom, realistic indoor lighting, and minimal color processing. Character: HAEUN, a Korean female idol in her 20s with long straight black hair, fresh natural makeup, and a bright, friendly personality. She starts in comfortable loungewear before changing into a stylish casual outfit—an oversized knit sweater, pleated skirt, white sneakers, and a small shoulder bag. Location: A cozy modern apartment followed by quiet city streets leading to a trendy outdoor café with a warm, lively atmosphere. The recording begins with HAEUN picking up the camera from her desk. She smiles and says, "I'm finally catching up with my friends today." She walks through her apartment while deciding what to wear, holds up two outfits with a playful expression, laughs, and quickly changes into her favorite one. Standing by the mirror, she fixes her hair, puts on simple jewelry, sprays a little perfume, grabs her bag, phone, and keys. "Okay... I think I'm ready." She leaves the apartment and walks through peaceful streets, chatting casually about how busy everyone's been lately and how long it's been since they all met. As she approaches the café, she spots her friends sitting outside. She smiles excitedly and waves toward them. One of them waves back and stands up to greet her. She laughs, "There they are!" Still holding the camera, she walks over as her friends welcome her with smiles and cheerful greetings. The vlog ends naturally with everyone laughing together as she lowers the camera, as if she's putting it away to enjoy the rest of the day. Generation Requirements: One uninterrupted recording with consistent facial identity, realistic body movement, natural hand interactions, smooth walking, accurate lighting transitions, believable environmental audio, and no AI artifacts, jump cuts, anatomy errors, duplicated objects, or unrealistic motion. The final result should feel like a genuine personal vlog filmed on a handheld MiniDV camcorder.

Fantasy scene

Fantasy scene

Cinematic scene

Create a 15-second hyper-realistic cinematic lifestyle vlog featuring a confident 23-year-old female content creator enjoying her peaceful sunrise walk along a scenic lakeside park surrounded by lush trees, blooming flowers, and soft morning mist. She wears a cream athletic T-shirt, sage-green running shorts, white sneakers, and a minimalist smartwatch. The atmosphere feels calm yet energetic, with premium influencer aesthetics inspired by high-end Instagram Reels and TikTok lifestyle videos. Use ultra-smooth gimbal tracking, cinematic handheld movement, dynamic whip transitions, speed ramps, shallow depth of field, natural lens flares, soft golden sunrise lighting, crisp ambient nature audio, and realistic facial expressions with accurate lip sync. No text overlays, logos, or watermarks. Scene 1 (0–2 sec) Close-up sequence of her putting on wireless earbuds, tightening her smartwatch strap, picking up a stainless-steel water bottle, and stepping onto the walking trail. Dialogue: “A fresh morning starts right here.” ⸻ Scene 2 (2–5 sec) Front-facing selfie shot as she walks confidently with warm sunrise light filtering through the trees behind her. Gentle breeze moves her hair naturally. Dialogue: “The best part of the day is before the world wakes up.” ⸻ Scene 3 (5–8 sec) Quick cinematic montage: shoes landing softly on the path smartwatch counting steps sunlight shining through leaves birds taking flight dew sparkling on grass close-up of her relaxed smile Dialogue: “Even a short walk makes everything feel lighter.” ⸻ Scene 4 (8–12 sec) Elegant slow-motion shots of her stretching beside the lake, taking a deep breath, laughing naturally, and enjoying the warm morning glow. Dialogue: “A little movement… a lot more peace.” ⸻ Scene 5 (12–15 sec) She looks back toward the camera with a bright smile, gives a friendly wave, and continues walking toward the sunrise as the camera slowly pulls away. Dialogue: “Same time tomorrow?”

At a glance

Is this model right for you?

Image mode requires at least a first frame or a last frame. Reference audio cannot be used alone and must be accompanied by a reference image or video. Every upload has its true format validated, and signed duration, geometry, and FPS are checked before task creation.

What it does best
A video generation entry point centered on a 2K tier, flexible first/last-frame input, and controlled multimodal references.
Who it is for
A good fit for high-resolution concept clips, combined character and motion references, first-to-last-frame transitions, and multi-asset short videos.
Why use it on Kyeo AI
Kyeo includes media authenticity, count, duration, and additional-image cost in the contract enforced before credits are reserved.

Key facts

Quickly evaluate whether this model fits your business use case.

Reference images
Up to 9
Reference videos
Up to 3; ≤15 seconds total
Reference audio
Up to 3; cannot be used alone
Additional image charge
From image 6 onward: 4 credits each

Other names

The same model may have different names in official documentation, community discussions, and API integrations. This section brings them together for you.

MiniMax H3 video
Hailuo H3
MiniMax H3 2K video

Common questions

These practical questions focus on the task, input conditions, and result requirements you should confirm before choosing.

How do I use MiniMax H3?
How much does MiniMax H3 cost?
What reference media does MiniMax H3 support?
How do MiniMax H3 768P and 2K differ?
How are MiniMax H3 input images priced?

Decision guide

Still deciding? Start with these key decision points.

1
For 2K, compare H3

H3 provides a defined 2K tier in this group, with a correspondingly higher per-second rate.

2
Pair audio with a visual reference

A request containing reference audio but no image or video is rejected before task creation.

3
Limit images after the fifth

Each additional image costs 4 credits, so retain only assets that contribute distinct information.

Popular comparisons

Compare common alternatives on the same task to clarify differences in inputs, controls, and cost.

How do the resolutions and rates of MiniMax H3 and Seedance 2.0 Mini differ?
How do the reference-media capabilities of MiniMax H3 and HappyHorse 1.1 compare?

Model comparison table

Not sure which model to choose? Compare the key differences between this model and its alternatives at a glance.

Reference mode
MiniMax H3
9 images / 3 videos / 3 audio clips; 2K available
Seedance 2.0 Mini
9 images / 3 videos / 3 audio clips; up to 720p
HappyHorse 1.1
Up to 9 images; no reference video or audio

Real-world experience

Practical guidance based on public sources, current on-site limits, and representative tasks.

Media count directly affects the cost formula

H3 is not priced solely by seconds; every input image from image 6 onward also increases preauthorization.

The narrower format range is a product-safety boundary

Kyeo enables only JPEG, PNG, and WebP formats that the current path can identify reliably, rather than presenting unverified formats as usable.

Capabilities

2K video generation

Choose 2K for text, image, or reference scenes.

Multimodal references

Combine up to 9 images, 3 videos, and 3 audio clips.

First- or last-frame input

Use a first frame, a last frame, or both in image mode.

Use cases

High-resolution concept clip

Use the 2K tier to test a brand-film or visual-proposal direction.

Combined motion and character reference

Assign separate roles to a character image and a motion video.

Build toward a target last frame

Upload only the desired final frame and describe how the shot should arrive there.

Prompt tips

Refer to visual elements explicitly

Explain how the person, clothing, or setting in each image should appear in the video.

Align audio with action

State when the sound should occur and which action it accompanies.

Control detail at 2K

Prioritize subject, material, lighting, and camera direction instead of stacking conflicting style terms.

Why choose it

Provides a defined 2K output tier.
Flexible first- and last-frame input.
Clear multimodal-reference limits.
Additional-image charges are visible before task creation.

What to know first

Output is limited to 15 seconds.
Reference audio cannot be used alone.
Images from image 6 onward increase the cost.
Kyeo does not enable image formats that the current path cannot identify reliably.

Adjustable settings

Quickly evaluate whether this model fits your business use case.

Generation scene
Optional
Parameter type: Select
Default: text
Text
First frame / last frame
Multimodal references
Resolution
Optional
Parameter type: Select
Default: 2K
768P
2K
Video duration
Required
Parameter type: Integer
Default: 6 seconds
4–15 seconds

Credit cost

32–406 credits

768P and 2K cost 8 and 13 credits per second, multiplied by output duration plus verified total reference-video duration. The first 5 input images add no separate charge; images 6–9 add 4 credits each. Kyeo rounds the total up before task creation.

Credit-saving tip

Before generating images or videos in bulk, run an A/B test with the same input on this model and similar alternatives. Confirm the result first to avoid wasting credits.

Frequently asked questions

Related models

Still deciding? Take a quick look at these similar alternatives.

Seedance 2.0 Mini

Seedance 2.0 Mini combines text, first/last-frame, and multimodal-reference generation in one entry point. Reference video enters the pre-task pricing formula, so both media duration and output duration must be verified before task creation.

AI video model
15–150 credits

Seedance 2.5

Seedance 2.5 combines a 30-second boundary, higher media counts, and video-editing reference input in one strict scene contract. Automatic duration is not underestimated at the 5-second default; Kyeo preauthorizes against the 30-second maximum.

AI video model
102–5,700 credits

HappyHorse 1.1

HappyHorse 1.1 routes by image count: no image selects text mode, 1 image selects first-frame image-to-video, and 2–9 images select reference mode. All three modes share the same resolutions, durations, and per-second rates.

AI video model
68–435 credits

Sources

This page is based on vendor materials, model-family documentation, and Kyeo AI integration settings.

Source note

Explore MiniMax H3: 768P or 2K, 4–15 seconds, first/last frames, 9 images, 3 videos, and 3 audio clips, 32–406-credit preauthorization, and media limits. 768P and 2K cost 8 and 13 credits per second, multiplied by output duration plus verified total reference-video duration. The first 5 input images add no separate charge; images 6–9 add 4 credits each. Kyeo rounds the total up before task creation.

Last updated: 2026-10-06
MiniMax H3 video current interface notes
Integration

Used to check the input modes, visible controls, media limits, and availability of the version currently connected to this site. The page and workbench show the actual options.