What you'll make: examples and honest expectations
By the end of this lesson you'll know what MiniMax H3 does well, what it can't do, and what you'll have in hand when you finish this course.
Real examples
The videos below are work other people made with MiniMax H3. None of them are ours; each card shows its source and how it was made.
Fox•Fur•Essence Films: Making a short film from Beginning to End w/ Minimax
Evidence AMade with MiniMax H3Made by: Fox•Fur•Essence FilmsOne person shoots a short film from start to finish with nothing but H3 and their own computer. The length comes from editing, not from the model.Karakter sayfası kendi üretimimiz olmalı (Pinterest'ten alınmaz). İş akışının civitai.red bağlantısı kullanılmaz.Open at SourceOpen WorkflowH3 resmî Ref2VA demosu: pembe takım elbiseli adam ve kara kuzu
Evidence AMade with MiniMax H3Made by: MiniMax (resmî)You can also give a character's voice as a reference: a short audio clip carries the same timbre into new lines.Open at SourceOpen WorkflowH3 resmî T2VA demosu: yıldız gemisi köprüsü, hiper uzay sıçraması
Evidence AMade with MiniMax H3Made by: MiniMax (resmî)MiniMax's official text-to-video demo. Sound and music come out of the same generation as the picture.2K adımı açık kaynak değil, yalnız API'de.Open at SourceOpen WorkflowWhat you'll have at the end of the course
When you finish the required path (Modules 0, 1, 2, 3 and 10), these will be yours:
A GPU server with MiniMax H3 installed, in your own cloud account, opened with $300 in free credit. While the server is off, no GPU charges run.
Your first clip with sound, made with the official H3 template.
A method for working out what a clip costs you.
See how a prompt is built right now. You'll write your first clip in Module 3 with exactly this skeleton:
An H3 prompt isn't one flat sentence. Picture, ambient sound and music go in separate fields.
integrated_multimodal_description: A glass of tea on a wooden windowsill in soft morning light. The camera slowly pushes in toward the glass.overall_soundscape: Quiet morning ambience, distant birdsong, a soft breeze.non_diegetic_music: The score is silent for this shot.
H3's limits today
Know these before you buy. Each one is either stated in a source or measured by us.
1. Plan on English dialogue. In our production, characters spoke English. For Turkish we used two routes: English audio with Turkish subtitles, or generating the Turkish voice separately with a voice-over tool. We set up a test of H3 speaking Turkish but didn't record the result, so we're not saying it definitely can't; we haven't measured it yet. Module 4 covers voice in other languages separately.
2. One generation is 5–15 seconds. H3 was trained on clips of 124 to 362 frames. At 24 frames per second, that's roughly 5.2 to 15.1 seconds. Longer video is made by joining several clips in the edit (Module 6).
3. 768p on your own server. The 2K footage you see in the examples is only available through MiniMax's paid API. With the open model on your own server, you generate at around 768p.
4. It holds faces weakly. Keeping the same character from scene to scene takes a reference image. In our measurement, H3 fell behind a closed competing model on face likeness. In a wide shot, silhouette, costume and color carry the identity. We work on this in Module 5.
5. It draws some things broken. Crowds, readable text, and objects that fall or break often came out broken in our production. You'll learn to build your scenes without leaning on them.
6. The GPU is in the cloud, not in your computer. The course runs on a 96 GB GPU in the cloud; our measurements are on that card. We haven't measured H3 on 12–24 GB home cards yet.
How are the lessons taught?
There are no video lessons. Every lesson is written; commands sit in copyable boxes, and under each step there's a "how you know it worked" line. Topics like camera movement come with interactive graphics you play with yourself.
Next lesson: where you can use H3. The license's region rule may affect you too.
Powered by MiniMax H3