MiniMax H3.
2K, sound included, tested prompt by prompt.

MiniMax H3 generates video from a text prompt, from a first frame and an optional last frame, or from up to nine reference images. Everything comes out at 2K and 24 frames per second with a stereo audio track in the same file. Clips run from 5 to 15 seconds, and the prompt field takes up to 7,000 characters, so a full shot list with sound design fits in one request.

This guide is written from generations on the live model rather than from a spec sheet. Every clip below is one of those tests, at the resolution it came out of the model. Where a technique earns its words, there is a template for it. Where a widely repeated technique made no measurable difference in our tests, that is what we say.

• 5 to 15 seconds in one generation, at 2K and 24fps.
• Native stereo audio in the same file: room tone, foley, music, and speech.
• Six aspect ratios from 21:9 through 9:16, all at 2K.
• Up to nine reference images, or a first frame and a last frame. Not both.
• Prompts up to 7,000 characters. 3 credits per second of output.

Habits built on Seedance and WAN mostly carry over, and two of them turned out to be doing less work here than expected. Each row below was tested as a matched pair: same subject, same seed, same duration, with only the technique under test changed.

A note on the two negative results, because they contradict guidance published elsewhere. Our reference tests used inputs that could only fill one slot each, which is the easy case. Role language most plausibly matters when two references could plausibly be the same thing, such as two people or two locations, and we did not test that. Likewise, our negative lists banned things the model was not inclined to add. Both techniques are free, so there is no reason to drop them. They are just not the load-bearing part of a prompt here.

H3 fills gaps confidently. Leave the camera unspecified and it will direct itself, often cutting between two setups inside five seconds. That is a good deal when you want a finished-looking clip and a bad one when you need a specific shot. Name the parts you care about and let it invent the rest.

<Subject> <performs the primary action> in <scene and environment>, at <time of day>.
<Light direction, materials, colour, and texture>.
Begin on <opening shot size and framing>, then <the one camera move you want>.
Audio: <the sounds that belong to this action, in the order they happen>.

A barista in a narrow espresso bar tamps a portafilter, locks it into the group head, and starts the shot, at sunrise.
Low warm sunlight comes through a window on the left, steam catches the light, brushed steel and matte ceramic surfaces.
Begin on a medium shot of her hands at the machine, then slowly push in to a close-up of the espresso stream.
Audio: the grind of the burr, the clack of the portafilter locking in, the hiss of the pump, and quiet morning room tone.

Neither clip is bad, and that is the finding. The structured prompt did not buy prettier pictures, it bought obedience. Use the formula when the shot has to match something already in your head, and a one-liner when you are happy to be surprised.

For anything longer than a single action, write the clip as consecutive timed beats. Give each beat one primary change and say what should be visible when it ends. The end state is the part that does the work: it gives the model a checkpoint to hit rather than a mood to sustain.

<One line naming the setting and the overall event.>
[0-3 seconds] <Opening shot. One primary change.> End with <a state a viewer could point at>.
[3-7 seconds] <Next shot. One primary change.> End with <a state a viewer could point at>.
[7-10 seconds] <Final shot. One primary change.> End with <the closing state>.
Audio: <what runs underneath, and what happens at the specific moments>.

Where fifteen seconds runs out

We ran the 15-second ceiling with five beats of three seconds each. The first four landed in order and on time. The fifth, which required cutting the base free with a wire and lifting the finished vase away, did not complete before the clip ended.

• Budget about four seconds for any beat that involves a prop change or a hand-off. Three is enough for a camera change but tight for an action.
• Expect the last beat to be the one that gets squeezed. Put the shot you care about most in the middle, not at the end.
• One primary change per beat. Two changes in one beat collapse into whichever is easier to render.
• Write end states as things a viewer could point at: an empty surface, a tool in a named hand, a door that is now closed.

Legible text is the thing H3 is most obviously good at, and it is the biggest practical difference from the other video models we have tested. Both halves of our test produced clean, correctly spelled type. What the prompt controls is which words appear and how they are set.

A title card reads "<THE EXACT STRING>" in <weight, case, and family>, <position in frame>.
Do not misspell it, do not add any other text, do not add subtitles.

• If a word has to be readable, type the word. Describing it gets you text, but not your text.
• Name the typographic treatment: condensed, all-caps, serif, tracked wide. It follows these.
• Say where it sits in frame. "Centred" and "lower third" both land.
• Add the do-not line. It costs nothing and it is the one place a negative earned its space in our tests.

Attach up to nine images and H3 composites them into one shot: a face from one, a product from another, a location from a third. Published guidance says the highest-leverage habit is telling the model what each image is for. We tested that twice and could not measure a difference.

The earlier two-image test with a person and a location came out the same way. Our reading is that when each reference can only plausibly fill one slot, H3 works out the assignment on its own. Role language is free and we still recommend writing it, because the moment two references could fill the same slot the model has to guess, and a sentence is cheaper than a re-run.

Use Image 1 for <what it defines>: preserve <the specific features that must survive>.
Use Image 2 for <what it defines>: preserve <the specific features that must survive>.
Use Image 3 for <the location>: <the elements of the set that matter>.
Do not use <the parts of any reference that should not appear, such as a studio backdrop>.
<The action, the light, and the shot.>

On PixelDojo the references are images. The underlying model also reads reference video and reference audio for motion transfer and voice, which the tool does not expose today.

Everything above is enough for a single shot with one or two references. Once a prompt carries several assets and a named cast, it needs a token system and a fixed layout. This section is the structure we use for those, written from the model documentation rather than from our own A/B battery.

Reference tokens

Refer to attached assets with compact tokens: Image1, Video1, Audio1. No space before the number. The number is the upload position, so a token means whatever sat in that slot when you attached it. Fix the order before you start writing and never renumber mid-prompt. If you swap an asset, swap the file rather than the token.

• Map every character and every voice one at a time. "The two women, respectively" is the fastest way to get them crossed.
• Use the token every time you mean that asset. Falling back to "the reference photo" halfway down a long prompt reopens the guess.
• One job per asset. If a single image has to supply both a face and a room, say which part supplies which.
• Say what to exclude as well as what to keep. A studio backdrop, a white product sweep, or a watermark is part of the file and needs ruling out by name.

Pick the task contract first

Before writing a word, decide which of the narrow contracts the shot is. Each one has a different rule about what the references are allowed to control. Mixing two contracts in one prompt is where most reference work falls apart.

Lock the stable truths before the shot list

Anything that has to stay identical for the whole clip belongs above the shot list, in labelled sections. The shot list then only carries what changes. Subject count, faces, wardrobe, props, voices, and who stands where are the six that drift if you leave them to the beats.

[REFERENCE USE]
Image1 = <one job>. Keep <features>. Ignore <what must not travel>.
Image2 = <one job>. Keep <features>. Ignore <what must not travel>.

[IDENTITY / CONTINUITY LOCKS]
<How many people are in frame, and no more.>
<Each face, tied to its token.>
<Each garment, described in words as well as shown.>
<Each hero prop, and which hand or surface holds it.>

[SCENE]
<Location, time of day, weather, and the light.>

[DIALOGUE]
<Character>: "<the exact line>"

[SCREEN GEOGRAPHY]
<Who is on the left, who is on the right, and what they look at.>

Screen geography is the one people skip. Naming who stands where, and where each pair of eyes goes, is what keeps eyelines from breaking when the camera moves.

Then the shot list, then the finish

The shot list follows the same timed-beat rules as earlier in this guide: consecutive ranges that do not overlap, one primary event each, and a stated end state. Give each range a framing, a camera behaviour, an action, any line of dialogue, and what is true when it ends. For a single continuous action, plain prose is still fine. Brackets are for sequences, not for decoration.

After the shot list, close with the finishing sections. These describe the whole clip rather than any one beat, so they sit at the bottom where they cannot be read as a step.

• [ACTING]: the observable behaviour, written as in the performance section above.
• [LIGHT AND IMAGE]: light direction, contrast, palette, lens feel, and grain.
• [CAMERA]: the one move you want, or the refusals that keep the frame still.
• [PRODUCTION SOUND]: the bed, the event sounds in order, and anything banned.
• [NEGATIVES]: name real failure modes, not generic quality words. Identity drift, wardrobe swaps, extra people, broken eyelines.

[REFERENCE USE]
Image1 = Nadia, lead. Keep the face, the freckles, and the dark curls tied back. Ignore the grey studio backdrop.
Image2 = Theo, second character. Keep the face, the beard, and the wire-rim glasses. Ignore the office behind him.
Image3 = the record shop interior. Keep the wooden browsing bins, the yellow pendant lights, and the tiled floor.
Video1 = motion source only. Copy the slow turn of the head and the reach across the bin. Do not use its actors, its clothing, or its location.
Audio1 = Nadia's speaking voice. Nadia speaks in this voice. Keep her lip sync accurate to the line below.

[IDENTITY / CONTINUITY LOCKS]
Exactly two people are in frame for the whole clip. No extras, no passers-by.
Nadia has the face from Image1. Theo has the face from Image2.
Nadia wears a rust corduroy jacket over a cream tee. Theo wears a charcoal knit sweater.
The hero prop is one sleeved vinyl record. It stays in Nadia's hands until she passes it to Theo.

[SCENE]
The record shop from Image3, late afternoon. Low sun comes through the front window on the right and the pendant lights are already on.

[DIALOGUE]
Nadia: "This is the pressing I told you about."

[SCREEN GEOGRAPHY]
Nadia stands on the left at the browsing bin. Theo stands on the right by the counter. They look at each other when they speak and at the record when it changes hands.

[SHOT LIST]
[0-4 seconds] Medium two-shot, locked off. Nadia turns her head toward Theo and reaches across the bin. End with the record lifted clear of the bin in her right hand.
[4-8 seconds] Slow push in to a medium close-up on Nadia. She says her line and holds the record up. End with the sleeve fully readable in frame.
[8-12 seconds] Hold the medium close-up, then reframe right as she passes the record. End with the record in Theo's hands and Nadia's hands empty.

[ACTING]
Nadia's gaze stays on Theo through the line and drops to the record as she passes it. Her thumbs stay on the sleeve edges. Theo's shoulders drop as he takes it and he exhales once.

[LIGHT AND IMAGE]
Warm low sun from the right, soft fill from the pendants, deep shadows in the bins. Muted rust and amber palette. Shallow depth of field on a 40mm lens, fine grain.

[CAMERA]
One slow push in and one reframe, both on a tripod head. No handheld, no zoom, no whip pan.

[PRODUCTION SOUND]
Quiet shop room tone, the rustle of sleeves in the bin, one soft card-stock slide as the record leaves its slot, and Nadia's line clean over the top. No music.

[NEGATIVES]
No identity drift between Nadia and Theo. No wardrobe changes. No third person entering frame. No broken eyelines. No subtitles or captions.

Read it back before you spend the credits

• Does every attached asset have exactly one named job, and are the parts you do not want ruled out by name?
• Does every identity, voice, prop, and edit target have exactly one owner? Two owners means the model picks.
• Are the timed ranges consecutive, non-overlapping, and long enough for the action inside them?
• Do the camera, light, and acting sections contradict anything the references already establish?
• Are the negatives real failure modes rather than words like "bad" or "low quality"?

Attach a start frame and H3 animates from it. Attach a start and an end frame and it fills the motion between them, which is the most precise control the model offers. Aspect ratio comes from the image in this mode, so the ratio picker is ignored.

• Write the prompt as the motion between the two frames, not as a description of either one. The frames already say what things look like.
• Use it for transitions where you know both ends: a poster that animates, a product that rotates into a hero angle, a character who exits a set.
• Frames and reference images are mutually exclusive. Pick one or the other.

H3 reads cinematography vocabulary directly, and it defaults to movement. If you say nothing about the camera, expect a slow drift and continuous reframing. The reliable way to stop that is not to ask for a static shot but to list the moves it should not make.

Locked-off static <wide / medium / close> shot.
No push in, no handheld, no zoom, no dolly. The frame never moves.

When you do want a move, name one and describe what changes on screen. "Slowly push in to a close-up of the espresso stream" was executed as written. Bare film-school terms with nothing visible attached are the weakest form of camera direction on every model we have tested, this one included.

Emotion words are a summary of a performance, not a description of one. Write what a camera could see instead: where the eyes go, what the hands do, what the breathing does, what stays still.

<Subject> <is in the situation>.
<What the hands do>, <where the eyes go>, <what the shoulders or posture do>, and <one breath or pause>.
<Shot size>, <camera behaviour>.

Every clip arrives with a stereo track, so sound is something you direct rather than something you receive. Name the sounds in the order they happen, say what runs underneath, and say what should not be there.

Audio: <what runs underneath for the whole clip>, <the specific event sounds in the order they happen>, and at <the moment> <the sound that marks it>.
No music.

One honest limitation: we verified that every clip carried a real audio track and measured how the levels behaved, but we did not check by ear whether spoken dialogue matched requested words. Treat speech and lip sync as untested here rather than endorsed.

There is no negative prompt field. Exclusions go in the prompt itself, as plain sentences. We tried twice to show that they do something and could not.

The earlier boat-on-a-lake pair behaved the same way. So the practical advice is narrower than "negatives are unusually effective": use them for the things the model likes to add on its own, which in our testing meant on-screen text and camera movement. Do not count on a negative to remove a subject the scene implies. If a person should not be in the shot, describe an empty place rather than a place with the person banned.

Six aspect ratios, all rendered at 2K: 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16. Vertical is a real 1440 by 2560 rather than a cropped landscape, so it is worth generating social assets natively instead of reframing afterwards.

Because every generation is 2K, a rejected clip costs the same as a good one. Test the look and the sound at 5 seconds, then re-run the prompt you like at the length you need.

• Is the camera specified? If not, expect it to drift. If you want it still, list the moves it should not make.
• Is any on-screen text typed out exactly, with the treatment and position named?
• Is the performance written as behaviour a camera could see, rather than as an emotion?
• Does each timed beat carry one primary change and an end state a viewer could point at?
• Is the last beat one you can afford to lose? It is the one that gets compressed.
• Is the audio named in order, with anything unwanted ruled out?
• If references are attached, are the features that must survive named in words as well as shown?
• Is the length right? You pay 3 credits a second either way, so prove the prompt at 5 seconds first.