MiniMax H3: The Essentials

Provider: MiniMax (H3 Max and H3 Max Turbo post-trained by Fal)  |  Modality: Text / Image / Video to Video  |  Model IDs: model_minimax-h3, model_minimax-h3-max-t2v, model_minimax-h3-max-i2v, model_minimax-h3-max-turbo-t2v, model_minimax-h3-max-turbo-i2v

MiniMax H3 generates short video clips with native stereo audio from a text prompt, first and last frames, or multimodal references, running 5 to 15 seconds at up to 2K and 24 fps. MiniMax H3 Max is a faster, simpler sibling from Fal: text-to-video and image-to-video only, at 480p or 768p, with no reference images or videos, tuned for stronger prompt adherence and richer aesthetics. Use H3 for brand films, product spots, story beats with dialogue, motion design, and game UI trailers; use H3 Max when you want a fast, cheap, on-model clip from a prompt or a single image. MiniMax H3 Max Turbo pushes that further: the same text-to-video and image-to-video pair, tuned again by Fal for even stronger prompt adherence and faster inference, best when a detailed prompt needs to land exactly as written.

Which Model Should I Use?

Reach for MiniMax H3 when you need synced dialogue, reference-driven consistency, or up to 2K output. Reach for H3 Max when you want a faster and cheaper clip from just a prompt or a single image, and do not need multi-reference input.

How MiniMax H3 Works

Start with a Prompt that describes the scene, camera, and audio. Then choose one input mode:

• Text only. Generate from the Prompt alone. Good for open creative exploration.

• First Frame / Last Frame. Lock the opening shot, and optionally the closing shot. Aspect Ratio follows the first frame. Do not mix frames with Reference Images, Videos, or Audio.

• Reference Images. Pass up to 9 images to guide subjects and style. First 5 are free; additional images are billed. Cannot combine with first or last frame.

• Reference Videos. Pass up to 3 clips (2 to 15 seconds each, 2 to 15 seconds total) to guide motion.

• Reference Audio. Optional, and only with at least one reference image or video.

Native audio (dialogue, SFX, ambience) is generated in the same pass. Put spoken lines in the Prompt when you need clear lip-synced dialogue.

Writing Prompts for MiniMax H3

H3 takes a single Prompt, but it responds well when that Prompt reads like a tiny shooting script rather than one long sentence. MiniMax's own guidance boils down to a simple habit: walk the clip forward in time, and describe three things as you go, the picture and action, the in-scene sound, and any background music. You can write all of it in plain prose.

Walk the clip forward in time

Open with a look, for example "Live-action, cinematic" or "2D anime", then describe what happens beat by beat. If the clip has more than one shot, call the cuts out in order and, when timing matters, add a rough timestamp such as "around the four-second mark, cut to a close-up of the coffee cup". Keep any readable on-screen text in quotes so it renders as written.

Direct the camera in plain words

Name the move, how big it is, and how fast, all in one phrase. Something like "the camera slowly pushes in on the folded letter" is enough. H3 understands the usual moves, push in and pull out, pan, tilt, truck, arc, tracking, a static hold, a slight handheld shake, or a POV, so describe the one you want and let it sit inside the action.

Write dialogue as spoken lines

Put spoken words in quotes and make it clear who is talking, for example: the waitress leans in and says, "We close in five, honey." Short lines lip-sync best. If a voice is narration rather than someone on screen, say so, and note that their lips stay closed.

Describe the sound you want

Because H3 generates audio in the same pass, a sentence or two of ambience goes a long way: rain on glass, a fridge humming, a cup set on the counter. If you want a musical bed, describe it briefly by instrument and mood, for example "a slow, sparse piano, quiet and unhurried". Ask for silence explicitly when you want it.

MiniMax H3 Max shares this same approach: describe the shot in timestamped beats, name the camera move in plain words, and describe the sound you want. See the MiniMax H3 Max section below for its own worked examples.

A short worked example

A text-only beat, written as plain prose:

Live-action, cinematic. A rain-streaked diner window at night, warm interior light. The camera slowly pushes in on a tired waitress as she leans over the counter and says, "We close in five, honey." Around the four-second mark, cut to a close-up of a full coffee cup with steam rising. Ambient sound: steady rain on the glass, the low hum of a refrigerator, distant traffic. Under it, a slow, sparse piano over soft upright bass, quiet and unhurried.

Your input mode shapes where the description starts. With Reference Images, ground the opening in the reference and develop forward while keeping the subject consistent. With First and Last Frame, describe the motion that carries the opening frame to the closing one. With a Last Frame only, describe the action that leads up to it.

MiniMax H3 Max: A Faster, Simpler Sibling

MiniMax H3 Max is Fal's post-trained variant of MiniMax H3: stronger prompt adherence and richer aesthetics, at 480p or 768p. It takes only a prompt or a first/last frame image, with no reference images, videos, or reference audio inputs, though it does generate its own synced audio in the output. Two models cover it: model_minimax-h3-max-t2v for text-to-video, and model_minimax-h3-max-i2v for image-to-video with optional end-frame control.

How MiniMax H3 Max Works

Both models take one plain-text Prompt describing the scene, action, camera, and mood. H3 Max responds best to long, detailed prompts written as timestamped beats (what happens from 0 to 3 seconds, 3 to 6, and so on) rather than a short caption. Set Duration (5 to 15 seconds) and Resolution (480P or 768P), and optionally raise Prompt Expansion to let the model rewrite a thin prompt into a richer one before generating.

Text to Video

H3 Max T2V also exposes Aspect Ratio (21:9, 16:9, 4:3, 1:1, 3:4, 9:16), so the same model covers widescreen cinematics, square product loops, and vertical social clips.

A veteran wildland firefighter in soot-streaked turnout gear stands at the edge of a smoldering ridge at dusk. 0-3s: she plants her boots on the ash-covered rock, exhaling as embers drift past her face, the orange glow of the distant fire line reflecting off her visor. 3-6s: the camera slowly pushes in as she raises a hand to shield her eyes, wind picking up and blowing ash sideways across the frame. 6-10s: she turns and starts walking down the slope toward a line of parked engines, silhouetted against the smoke-filled sky, the rumble of distant flames and crackling radio chatter building under a low ambient wind. Cinematic wide shot, warm amber and deep orange grade, volumetric smoke, handheld camera with subtle shake.

Duration 10s, Resolution 768P, Aspect Ratio 16:9 · Open on Scenario

Prompt Expansion set to Quality spends extra time enriching a shorter prompt before generation. It paid off on this neon alleyway shot, where the model added rain, drone lighting, and background detail beyond what was explicitly described:

A cyberpunk mercenary in a weathered tactical jacket strides down a narrow neon-lit alley at night. 0-4s: rain-slicked pavement reflects flickering holographic ads as he walks toward camera, breath visible in the cold air. 4-7s: the camera cranes low as he passes a food stall, steam rising and neon signage in Japanese and English glowing pink and cyan overhead. 7-10s: he glances up at a passing drone, its spotlight briefly sweeping across his face before he ducks into a doorway, distant synth music and rain underneath. Moody cyberpunk cinematography, high contrast neon color grade, volumetric rain, slow tracking shot.

Duration 10s, Resolution 768P, Prompt Expansion: Quality · Open on Scenario

Duration goes up to a full 15 seconds, enough room for a small multi-beat story in one clip:

An elderly fisherman in a knit sweater sits on a weathered wooden dock at dawn, mending a torn fishing net draped across his lap. 0-4s: his weathered hands work the twine methodically, mist rising off the still water behind him as the sun begins to break over the horizon. 4-8s: he pauses, looks up toward the open sea with a quiet, reflective expression, gulls calling faintly in the distance. 8-12s: he returns to his mending, a small wooden boat gently rocking at its mooring beside the dock, the light growing warmer as the sun climbs. 12-15s: he ties off the final knot and sets the net down, exhaling slowly as calm water laps against the dock pilings, soft ambient waves and distant gulls underneath.

Duration 15s (the model maximum), Resolution 768P · Open on Scenario

Aspect Ratio adapts the same model to square product loops, vertical social clips, and ultra-wide cinematics:

Aspect Ratio 9:16, product/marketing

Aspect Ratio 21:9, fantasy/game cinematic

Aspect Ratio 1:1, Resolution 480P, product loop

Aspect Ratio 21:9, travel/lifestyle

Image to Video

H3 Max I2V takes a First Frame image and animates it; the output shape follows that image, so there is no separate Aspect Ratio control. It stayed fully on-model with an illustrated, non-photoreal source, keeping the exact art style while inventing a consistent kitchen environment behind the original flat background:

Source image (First Frame)	H3 Max I2V output

0-3s: the chef throws his head back mid-laugh as he tosses the wok, the tower of orange flame surging upward and the noodles arcing through the air before landing back in the pan. 3-6s: he swirls the ladle through the sizzling noodles, sparks and embers flicking off the flame as steam curls from the wok, his straw hat tilting with the motion. 6-10s: he plants the wok back down with a satisfied grin, the flame settling into a steady roaring plume behind him, a few stray noodles still drifting down into frame. Energetic hand-drawn animation style, bold flat colors, dynamic flame and motion lines, playful comedic energy, upbeat sizzling and crackling fire sounds.

Duration 10s, Resolution 768P, source: illustrated chef still · Open on Scenario

Product stills work just as well. A studio photo of a watch became a slow hero-angle rotation with light sweeping across the dial, nothing invented beyond what the prompt asked for:

Source image (First Frame)	H3 Max I2V output

Duration 8s, Resolution 768P, product commercial style · Open on Scenario

Locking the Closing Shot with Last Frame

Last Frame only works alongside First Frame. Set both and the model animates a transition between the two exact images instead of improvising an ending. Here, a perched dragon still was paired with a separately generated airborne pose as the Last Frame, and the model built a full launch-and-climb sequence that lands on it:

0-4s: the bronze-scaled dragon perched on the cliff edge slowly unfolds its wings, muscles rippling beneath its scales as it crouches lower. 4-8s: with a powerful downward beat of its wings it launches off the rock into the golden hour sky, talons releasing from the stone as dust and loose pebbles scatter below. 8-12s: it banks upward against the sunset, wings fully spread and flapping steadily as it climbs over the misty coastline, a deep guttural roar echoing across the valley.

Duration 12s, Resolution 768P, First Frame: perched dragon, Last Frame: airborne dragon · Open on Scenario

Last Frame only takes effect when a First Frame is also set. Without a First Frame, Last Frame has no visible effect.

A few more I2V examples across verticals:

Classroom robot assistant, education

Lighthouse keeper, film/narrative

Surfer in the barrel, sports/action

Cinematic Showcase

H3 Max is not limited to simple product turns. It follows aggressive, named camera moves (barrel-roll orbits, Hitchcock dolly-zooms, drone launches, crash-zooms) as reliably as it follows plain description. The clips below all came from this article's test batch, full prompts included.

A getaway driver in a black leather jacket grips the wheel of a matte-black muscle car inside a rain-slicked, neon-lit parking garage at night. 0-3s: tires screech as he yanks the handbrake into a full 180-degree drift, sparks flashing off the rear bumper against a concrete pillar. 3-6s: the camera whips into a full barrel-roll orbit around the spinning car, neon pink and blue reflections streaking across the wet floor and windshield. 6-9s: the car snaps straight and launches forward through the exit ramp, tires smoking, engine roaring under screeching rubber and distant sirens. High-octane heist-film cinematography, anamorphic lens flares, high contrast neon and shadow, dynamic orbiting camera.

T2V, barrel-roll orbit camera, Aspect Ratio 21:9 · Open on Scenario

A window washer in a harness stands on a narrow platform against the glass facade of a skyscraper, hundreds of feet above a city at dusk. 0-3s: he leans out to wipe a streak of glass, wind tugging at his jacket, city lights twinkling far below. 3-7s: the camera performs a slow dolly-zoom, pulling back while zooming in, the background plunging into vertiginous distortion as his eyes widen slightly at the drop beneath him. 7-10s: he steadies himself, exhales, and resumes wiping calmly as the skyline glows gold in the fading light, wind and distant traffic hum underneath. Vertigo-style Hitchcock zoom, warm dusk lighting, razor-sharp skyline detail, unsettling scale.

T2V, Hitchcock dolly-zoom, Aspect Ratio 16:9 · Open on Scenario

0-4s: the camera begins a slow 360-degree orbit around the model as the liquid mercury dress ripples and reforms in constant slow motion, silver highlights sliding across its surface like living metal. 4-7s: as the orbit completes, droplets of mercury lift briefly off the hem and hover before falling back into the fabric, rim light catching every ripple. 7-9s: she tilts her head slightly, the dress settling into a final rippling wave down its length, a low resonant hum and soft metallic chime underneath. High-fashion editorial cinematography, dramatic rim lighting, full uninterrupted orbital camera move, dark and glossy.

I2V, full 360-degree orbit, from a liquid-mercury fashion still · Open on Scenario

T2V, drone skim into vertical launch, bioluminescent wave

T2V, chase camera whip-pan, wingsuit canyon dive

I2V, macro crash-zoom to vertigo pull-out, lava cake

A few more from the same batch, spanning scale from zero-gravity to extreme macro:

T2V, full orbital barrel roll, zero-gravity EVA

T2V, low-angle push-in to whip-pan reveal, storm chaser

T2V, slow dolly-in, self-folding paper crane

T2V, tracking shot cut to onboard POV, race chicane

I2V, deep glide to upward tilt reveal, ice cave

Examples (MiniMax H3)

All clips below were generated with MiniMax H3 (model_minimax-h3) in Public Data.

Reference Images: desert fashion campaign

Four stills (car, portrait, handbag, brand mark) passed as Reference Images for a premium editorial brand film.

Desert Fashion Campaign · Reference Images · Open on Scenario

Text to video: kitchen creature

Prompt-only smoke test: lived-in kitchen at dusk with soft hand-drawn luminous creatures and quiet room-tone audio.

Hand-Drawn Kitchen Creature · Text to Video · Open on Scenario

Reference Images: game equipment UI

Character key art plus HUD panels as references for an interactive game inventory reveal.

Interactive Game Equipment UI · Reference Images · Open on Scenario

Dialogue: vertical family confrontation

9:16 short drama with explicit spoken English lines in the Prompt and Reference Images for the two-shot and close-up.

Vertical Family Confrontation · Dialogue · Open on Scenario

Dialogue: cafe reunion

9:16 reunion scene with clear synced dialogue under rainy cafe light.

Cafe Reunion · Dialogue (9:16) · Open on Scenario

Parameters (MiniMax H3)

These are the dials you will touch most often on MiniMax H3.

prompt

Required. Describe the video: subject, action, camera, lighting, and audio. For dialogue, write the spoken lines in English inside the Prompt so lip sync has a clear target. Max length 7000 characters.

firstFrameImage

Optional image used as the opening frame. The video shape follows this image. Required if Last Frame is set. Cannot combine with Reference Images, Videos, or Audio.

lastFrameImage

Optional closing frame. Only works when First Frame is also set. Cannot combine with reference media.

referenceImages

Optional. Up to 9 images (max 30 MB each) to guide subjects or style. First 5 are free; additional images are billed. Cannot combine with First Frame or Last Frame.

referenceVideos

Optional. Up to 3 videos to guide motion (each 2 to 15 seconds, 2 to 15 seconds total, max 50 MB each). Billed per second of uploaded duration. Cannot combine with First Frame or Last Frame.

referenceAudio

Optional. Up to 3 audio clips (each 2 to 15 seconds, 2 to 15 seconds total, max 15 MB each). Requires at least one reference image or video.

duration

Clip length in seconds, from 5 to 15. Default 5. Longer clips cost more.

aspectRatio

Output shape: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16, or Adaptive (default). Ignored when First Frame is set, because the shape follows that image.

Parameters (MiniMax H3 Max)

T2V and I2V share most of the same controls. Differences are called out below.

prompt

Required on both models, up to 50,000 characters. Write it as a full, detailed paragraph, ideally broken into timestamped beats (0 to 3s, 3 to 6s, and so on) describing action, camera, and mood. See the firefighter and fisherman examples above.

firstFrameImage

I2V only, required. The opening frame of the video. The output's shape follows this image, so there is no separate Aspect Ratio control on I2V. See the chef and watch examples above.

lastFrameImage

I2V only, optional. The closing frame of the video. Only takes effect when firstFrameImage is also set. See the dragon launch example above.

duration

5 to 15 seconds on both models. Longer durations cost more. The fisherman example above used the full 15 seconds for a small multi-beat story in one clip.

resolution

480P or 768P, default 768P, on both models. Higher resolution costs more. 480P is a fast, cheap option for drafts and quick product loops (see the sneaker example above); 768P is the better default for hero shots.

aspectRatio

T2V only, default 16:9. Choose from 21:9, 16:9, 4:3, 1:1, 3:4, or 9:16. See the demo grid above for 9:16, 21:9, and 1:1 in practice.

promptExpansionMode

Disabled, Balanced (default), or Quality, on both models. Disabled skips prompt rewriting entirely. Balanced spends about a second on it. Quality spends up to about 30 seconds letting the model enrich a shorter prompt before generating, which added convincing background detail in the cyberpunk alley example above.

seed

Optional on both models. Leave blank for a random seed, or set one for reproducible results.

Use Cases

• Brand and fashion films: Multi-still Reference Images for campaign continuity across mood, talent, product, and end card.

• Product and ecommerce: Orbit and hero reveals for chairs, eyewear, sneakers, and pack shots with native ambience.

• Story and dialogue shorts: Vertical 9:16 beats with explicit spoken lines for social and drama concepts.

• Game and UI trailers: Character plus HUD references for inventory reveals and concept trailers.

• Motion design and stylized spots: Text-only or image-guided clips for neon, anime, and editorial looks.

• Film and teaser language: Space opera and mystery teasers with orchestral or drone beds generated in-pass.

• Fast product and social loops (H3 Max): the watch, sneaker, and perfume I2V examples turn a plain product still into a hero rotation without a physical shoot, in a fraction of the time and cost of a reference-driven H3 render.

• Bringing illustrated or concept art to life (H3 Max): the wok chef example shows H3 Max preserving a non-photoreal art style rather than pulling the output toward photorealism.

• Marketing and advertising concepts across art styles (H3 Max): flat-illustration ads, retro synthwave spots, and stop-motion or claymation-style pieces all hold their intended style rather than defaulting to realism.

Tips for Better Results

• Prefer Reference Images for showcase work. Generate clean stills first (one scene per still), then pass those asset IDs into Reference Images.

• Write dialogue as quoted lines in the Prompt. Keep lines short and assign speakers so lip sync stays readable.

• Do not mix modes. First/Last Frame and Reference Images/Videos/Audio are mutually exclusive.

• Match Aspect Ratio to the story. Use 9:16 for vertical dialogue; 16:9 for brand and product films. Adaptive is fine when you are exploring.

• Describe audio explicitly. Name room tone, music bed, SFX, and whether dialogue should lead the mix.

• Avoid packing multiple unrelated scenes into one Prompt without references. Split concepts across stills or separate runs.

• Write prompts as timestamped beats. Breaking the action into 0-3s, 3-6s, and so on gave consistently more coherent motion than a single unstructured description across every example in this batch.

• Give every beat a clear end state. Actions that trail off without a visible resolution risk the model losing track of the object; the wok chef and dragon examples both close each beat on a concrete pose.

• Ground I2V prompts in the actual source image. The wok chef and watch examples describe exactly what is already in the still (the straw hat, the flame, the dial layout) rather than a generic scene.

• Use Last Frame for a specific ending, not just a starting point. Pairing a matched First and Last Frame, as in the dragon example, gets a coherent full sequence instead of an improvised close.

• Reach for Prompt Expansion: Quality on thinner prompts. It filled in convincing environmental detail on the cyberpunk alley clip beyond what was explicitly written.

• Match resolution to the job. 480P was plenty for the sneaker product loop; save 768P for hero shots where detail matters.

• Push duration to 15 seconds for anything with more than one beat. The fisherman example used the full range to fit a small four-beat story in a single clip.

Known Limitations

• Duration is capped at 15 seconds. There is no schema-supported extend-to-30s path on this model.

• Reference Audio requires at least one reference image or video.

• First Frame / Last Frame cannot be combined with Reference Images, Videos, or Audio.

• Aspect Ratio is ignored when First Frame is set.

• Provider jobs can occasionally fail with a temporary internal error; retry the same payload.

• On-screen readable logos and brand text can drift; keep brand marks abstract or out of frame when fidelity matters.

• Audio is real but ambience/SFX-oriented, not reliable for dialogue. H3 Max does generate synced stereo audio matching the scene (footsteps, engines, wind, crowd noise, and similar), confirmed by measuring output levels across this batch. It is less consistent at rendering clear spoken dialogue than the base MiniMax H3 model, so write prompts expecting ambience and SFX first, with dialogue as a bonus rather than a guarantee.

• No reference images or reference videos. H3 Max only accepts a text prompt (T2V) or a first/last frame image (I2V). For multi-image style or subject consistency, use the base MiniMax H3 model instead.

• I2V has no Aspect Ratio control. The output shape always follows the First Frame image, so plan the source image's orientation before generating.

• Last Frame is inert without First Frame. Setting only Last Frame has no effect; both must be set together.

• Prompt Expansion: Quality adds latency. It can spend up to about 30 seconds rewriting the prompt before generation even starts, on top of generation time.

MiniMax H3 Max Turbo: Faster Still

MiniMax H3 Max Turbo is Fal's second pass on H3 Max: the same two models, model_minimax-h3-max-turbo-t2v and model_minimax-h3-max-turbo-i2v, tuned again for even stronger prompt adherence and faster inference. The parameter surface is identical to H3 Max (Duration, Resolution, Prompt Expansion, plus Aspect Ratio on T2V and First/Last Frame on I2V), so everything in the sections above about writing prompts as timestamped beats applies directly. Turbo is the pick when a detailed prompt needs to land exactly as written, fast.

Prompt Detail Changes Everything

The same subject, shot three ways with Prompt Expansion disabled, shows how much the model actually listens. A one-line prompt gets a generic, forgettable read. One paragraph of detail (bike type, camera move, lighting, sound) gets a materially closer result. A full timestamped shooting script gets a shot-for-shot match, down to the exact camera move scripted for the closing beat.

Simple: "A motorcyclist riding down a desert road at sunset." No scene, bike, or camera detail specified. · Open on Scenario

Detailed: one paragraph naming the bike, jacket, camera move, and lighting. · Open on Scenario

Muito detalhado: a timestamped 3-beat script. The clip ends on the exact scripted beat, a wide aerial pull-back with the bike a small silhouette and its shadow stretched across the rock. · Open on Scenario

Prompt Expansion: Quality Turns a Short Brief Into a Full Scene

Prompt Expansion set to Quality took a two-sentence brief and filled in the atmosphere, steam, sparks, and neon signage on its own, at 480P for a quick, inexpensive social clip:

Quality-mode Prompt Expansion from a short brief, Duration 8s, Resolution 480P · Open on Scenario

Across Genres, Fast

H3 Max Turbo holds up across very different visual languages in the same batch: sci-fi, epic fantasy, 2D anime, emotional drama, and on-screen text.

Sci-fi EVA repair, slow orbital camera drift · Open on Scenario

Epic fantasy creature reveal, wide dramatic camera · Open on Scenario

2D anime style, dynamic rooftop chase, vertical framing · Open on Scenario

Emotional drama, slow push-in · Open on Scenario

On-screen text rendered exactly as scripted, slow orbital camera · Open on Scenario

MiniMax H3 Max Turbo I2V: Animating Stills, Fast

Turbo's image-to-video side takes the same First Frame plus optional Last Frame inputs as H3 Max I2V, and stays faithful to the source, whether it is illustrated art or a plain studio product shot.

Illustrated art style preserved, gentle character animation · Open on Scenario

Product rotation reveal from a single studio still · Open on Scenario

Travel photography brought to life with a slow drone push · Open on Scenario

Turbo is not a different feature set from H3 Max, just a faster, more literal read of the same prompt. When you want the model to hedge less and follow more, reach for Turbo.