Source-based review · Updated August 5, 2026
MiniMax H3 Review: Quality, Native Audio, Reference Control, and Limits
MiniMax H3, also known as Hailuo 03, is a multimodal video model that combines text, images, video, and audio references. Its most distinctive capability is planning video and native stereo sound together rather than treating audio as a separate finishing step.
How this review was prepared
This is an evidence review of current provider documentation, official ComfyUI workflows, published prompt formats, and attributed community tests. It does not claim that every linked example was generated by Hailuo3.me, and it does not convert one community run into a guaranteed benchmark.
MiniMax H3 at a glance
| Generation modes | Text, Image, and Reference to Video |
|---|---|
| Reference inputs | Images, video clips, and audio files |
| Hosted output | Up to 2K |
| Duration | 4 to 15 seconds |
| Audio | Native stereo sound generated with the video |
| Framing | 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, plus adaptive in Reference mode |
Where H3 is strongest
- Native audiovisual scenes with dialogue, ambience, and score.
- First-and-last-frame transitions with a defined destination.
- Reference roles for identity, motion, camera, and voice.
- Short product films, narrative scenes, and motion concepts.
- Detailed natural-language direction across multiple shots.
Where to be careful
- Dense prompts can overload a short four-to-six-second clip.
- More references can introduce conflicts instead of control.
- Local ComfyUI performance varies sharply by configuration.
- Readable text and product details still need review.
- Every final result needs a full watch-through with sound.
Video quality and motion
H3 is designed for cinematic movement, multi-shot direction, and consistent audiovisual storytelling. Published examples show promising subject motion and spatial sound, but quality still depends on the prompt, references, duration, resolution, and number of retries. Treat an impressive clip as evidence of what is possible, not the expected output of every request.
Native stereo sound
H3 can coordinate dialogue, visible action, ambience, sound effects, and audience-only music inside one timeline. The prompt format distinguishes scene sound from non-diegetic music and can assign stable IDs to speakers. This is more controllable than adding “cinematic audio” as an isolated keyword.
Image and reference control
Image mode can begin from one frame or move between opening and ending frames. Reference mode can combine multiple media types. The key is role clarity: use one source for identity, another for motion, and another for sound only when each contribution is explained. See the MiniMax H3 prompt guide for reusable formats.
Local ComfyUI versus hosted H3
Open weights make local experimentation possible, but the model stack remains large and render time depends on GPU memory, system RAM, storage, quantization, and workflow settings. Hosted H3 is better for creators who prefer managed infrastructure and a focused interface. The ComfyUI setup guide covers that tradeoff in detail.
Primary sources and examples
Kie MiniMax H3 model page
Current hosted modes, input counts, aspect ratios, duration, resolution, and output contract.
Official workflowComfyUI MiniMax H3 examples
Official local workflow guidance and model-package requirements for H3.
Official prompt formatMiniMax H3 prompt-writing docs
Published format for shots, camera direction, dialogue, sound, and reference alignment.
Curated galleryMiniMax H3 examples and settings
Attributed first-party and community examples with clear evidence labels.
Create in your browser
Turn your direction into a MiniMax H3 video
Use Text, Image, or Reference to Video with clips up to 2K, native stereo sound, and private creation history.