MiniMax H3 prompt guide

How creators write MiniMax H3 prompts: references, dialogue, sound and the shot plan, sorted by the shot you want. A template for each, the checklist, the mistakes, and 61 clips to learn from.

Pick an effectSend us yours

Anime horror-night gag with burned-in subtitles

Shot list · MiniMax H3 per the author · Oct 6, 2026

Six shots have exact cut times, a field of view and a static camera each; the Japanese lines are spoken, and the English subtitles are specified as image-only text with typeface, outline, a single line, a fixed position and a rule that each one leaves at the cut. It reverses the usual subtitle ban on the same principle: every piece of on-screen text is listed exactly, and shots without a line are told to carry none. Gag faces are written as concrete anime conventions (shrinking pupils, shade lines, souls leaving the mouth), with a closing rule that faces return to the normal design.

Dialogue and native soundPerformance and emotionAnimated styles and art direction

Read @MayorKingAI's prompt
2D anime | romantic comedy | cel-shaded, clean line art, dark blue night palette.
Multishot, 6 shots. Cuts only at the specified points.

CHARACTERS
HANA: wavy shoulder-length dark red hair, oversized gray cardigan with one button missing, a bandage on her index finger. 160 cm.
YUTO: neat black hair with a small bed-head tuft at the back, navy hoodie with a pilled sleeve, glasses pushed up on his forehead. 178 cm.

LOCATION
Dark living room at night. They share one sofa under a single fleece blanket. A TV off-frame in front of them is the only light source, flickering blue on their faces. A bowl of popcorn between them.

SUBTITLES
Spoken dialogue is in Japanese. English subtitles are burned into the image as on-screen text only — they are never spoken aloud. Style: white sans-serif, thin black outline, one single line, bottom center of frame, 8% from the bottom edge. Each subtitle appears with its line and disappears at the cut. Subtitle text is shown inside 【】 brackets below; the brackets themselves are not displayed.

Shot 1 (0–2.5s) — Medium front shot of both, FOV 47, static.
Hana sits upright, arms crossed, smug smile, chin raised. Yuto beside her grips a cushion with both hands, already tense.
Hana (spoken, Japanese): 「全然怖くないし。」
On-screen subtitle: 【I'm not scared at all.】
HARD CUT

Shot 2 (2.5–5s) — Close-up on Yuto, FOV 29, slow push-in ending on his eyes.
His pupils shrink to tiny dots, a vertical column of blue shade lines drops over his upper face, sweat drop on his temple. He swallows, the cushion squeezed flat. No subtitle in this shot.
HARD CUT

Shot 3 (5–8s) — Medium front shot, FOV 47, static.
A violent white flash from the TV and a loud sting. Both jump. Clear gag faces: mouths huge and wavy, eyes white, hair standing in spikes. The popcorn bowl flips and popcorn bursts upward like confetti. Two small white ghost-shaped souls slip out of their open mouths.
Hana and Yuto (spoken, Japanese): 「ぎゃああああっ!」
On-screen subtitle: 【AAAAAAH!】
HARD CUT

Shot 4 (8–10.5s) — Wide shot, FOV 63, static.
They are clutching each other tightly, trembling, their outlines drawn with shaky wobbling lines. Popcorn rains down on them. The souls drift back into their mouths. No subtitle in this shot.
HARD CUT

Shot 5 (10.5–13s) — Two-shot close-up, profile, FOV 29, static.
Silence. They slowly realize how close they are, faces a hand apart. Both blush deep red at the same time. A tiny comic sparkle pops between them. They let go and shuffle apart in short jerky movements. No subtitle in this shot.
HARD CUT

Shot 6 (13–15s) — Close-up on Hana, three-quarter front, FOV 29, static.
Normal detailed anime face: cheeks red, eyes sliding sideways, lips pouting. One popcorn kernel stuck in her hair.
Hana (spoken, Japanese): 「…今のはノーカン。」
On-screen subtitle: 【...That one doesn't count.】
She pulls the blanket up over her nose. End on her eyes peeking above the blanket.

LIGHTING
Only light source is the off-frame TV: flickering cool blue 8500K, with a single white flash in shot 3. Deep shadows in the room.

Sound: No score. Production audio only — faint TV dialogue murmur, a sharp horror sting, popcorn scattering, fabric rustle, double scream, then near silence and a soft blanket rustle. All voices in Japanese only.

Characters keep the same face, hair and wardrobe in every shot. Gag faces are clear, exaggerated super-deformed expressions, then return to the normal detailed anime face. Clean consistent line art, consistent lighting between shots. Subtitles stay in the same position, size and style in every shot, spelled exactly as written. The room contains only the sofa, the blanket and the popcorn bowl.

Original post ↗

Pick the shot you want

  1. 01 Dialogue and native sound Scenes where people talk, written so the model generates the voices, lip movement and room sound together with the picture. 12 clips →
  2. 02 Performance and emotion Restrained scenes carried by eyes, breath, hands and one telling prop rather than by adjectives. 10 clips →
  3. 03 Reference-driven characters Images, character boards and storyboards used as references, each told exactly what it controls and what it must leave behind. 27 clips →
  4. 04 Kinetic type and title sequences Words that move, assemble and become the scene, from lyric videos to blockbuster title reveals. 23 clips →
  5. 05 Animated styles and art direction Pottery, paper cut-outs, woodblock prints, watercolour and pixel art, each with its own rules of motion. 23 clips →
  6. 06 Camera moves and transitions FPV one-takes, locked phone angles and train-window journeys where the camera has a defined path or a fixed post. 19 clips →
  7. 07 Sport, dance and rhythm Dance, ballet, fashion posing and idol routines where movement is broken into countable moves on a beat. 9 clips →
  8. 08 Product, cars and trailers Spots, parodies, game screens and title PVs where each beat has one job and one rule runs through them all. 7 clips →
  9. 09 Transformation and style shifts Letters becoming objects, forms morphing into forms, and smoke or debris reassembling into something new. 12 clips →
  10. 10 Close combat and impact Duels and battles written as exchanges with a cause, a contact and a reaction. 4 clips →
  11. 11 Chases, racing and fast movement Pursuits and flights built from a route, named obstacles and a camera that rides along. 4 clips →
  12. 12 Giants, armies and epic scale Small figures against huge worlds, and big events shrunk on purpose. 8 clips →

Seven rules that hold for every shot

  1. Give every reference a label and one job

    H3 accepts several images, and often a video or an audio track, in one request. Name each one (Picture 1, Image 2, the storyboard) and say what it controls: the face, the outfit, the location, the first frame or the order of beats. Then say what it must not pass on. A portrait carries its pose and camera angle, a turnaround carries its grey backdrop, and a two-person reference carries who stands where. The cases that hold up use the same label for the same asset everywhere it appears.

  2. Write the soundtrack into the prompt

    H3 generates sound together with the picture, so the prompt is also the sound script. Put each line of dialogue inside the shot where it is spoken, in its original language, attached to a named speaker. List ambient and physical sounds separately, and state whether there is a score and what it is. If you want a quiet room, say there is no music. Leaving sound unwritten does not give you silence; it hands the choice to the model.

  3. Give every line time to land

    Dialogue scenes break down when lines are crammed together. Write a small action before each line and a beat after it, and keep the word count modest for the running time; about two words per second is a sensible ceiling for natural delivery. One author running H3 locally reported dropped and blended words when a scene carried too many lines for its length. If a scene needs more words, make it longer or split it into two clips.

  4. Describe voices in words

    Introduce each speaker with a voice description: age, pitch, texture, pace and accent, plus how it changes during the scene. Keep one stable speaker label per voice, and mark off-screen speech as off-screen, with the visible character's lips closed. Attaching a read-aloud voice sample tends to make the delivery sound dubbed; a written description lets the model perform the line in the space it is in.

  5. Plan the shots like an editor

    Most strong H3 prompts are shot lists: numbered shots with cut times, a framing and a lens for each, and a fixed axis that says which side of the frame each character holds. Dialogue scenes follow film grammar, from a prop or two-shot opening through singles and over-the-shoulders to a held close-up. If you want one continuous shot, say so plainly and define the frame edges; short clips can add cuts of their own when the prompt leaves room for them.

  6. Lock the camera harder than seems necessary

    H3 tends to pull an unusual camera back toward the angle of the reference image, and to straighten perspective on faces at high resolution. To hold a high angle, a low angle or a phone POV, state the camera rule first and mark it as the top priority, repeat it in every time block, and list the fallbacks you do not want (eye level, side angle, pull-back) by name. Make the subject's pose follow the camera, not the other way round.

  7. Allow only the text you list

    Creators often use H3 for kinetic type, title reveals and lyric videos, and it will also render text you never asked for. Keep a text ledger: the exact words, in order, with typeface, colour and an outline rule for contrast. Ban everything else, including captions of the dialogue. Make each piece of text come from something already on screen, and tie its timing to a beat or a spoken line.

Newest in the guide

Each plays from the creator's own post on X. Our note sits under it; the prompt belongs to them.

A 1970s studio take lip-synced to an audio reference

Shot list · MiniMax H3 per the author · Oct 28, 2026

Two references get separate jobs: the image fixes the singer's face, with hair and wardrobe written out, and the audio supplies the vocal he must lip-sync from the first second to the last in every shot where his mouth is visible. Five three-second shots each have one framing and one step of rising intensity, and the period look is defined as a film stock with a separate clean-print rule, so the 16mm feel does not bring scratches or flicker with it. Handheld behaviour is written as specific imperfections (slight wobble, small reframes, focus hunting) rather than as a mood.

Reference-driven charactersSport, dance and rhythmCamera moves and transitions

A Korean magazine edit in thirteen cuts and seven looks

Shot list · MiniMax H3 per the author · Oct 23, 2026

Each of the thirteen numbered cuts carries a look number, one action, one set element and one exact line of type, so outfit, framing and text are scheduled in the same row. Outfits change on the cut and never by morphing, one model is fixed from the reference, and the last half second is held completely still on the finished cover. A long avoid list closes the prompt, covering face and hair drift, garbled text and extra people.

Kinetic type and title sequencesReference-driven charactersProduct, cars and trailers

Read @Yuki26856052's prompt
MiniMax H3 × Ima StudioTHE RED EDIT — BOLD LOOK. SOFT POWER. STYLE: Korean fashion magazine × K-pop comeback teaser × cinematic motion typography. A luxurious red magazine comes alive through 3D editorial installations, paper transitions, and dynamic layouts. CHARACTER: Use the uploaded reference image. Only ONE adult Korean female model, preserving exactly the same face, features, skin tone, slim proportions, and long dark brown-black hair throughout. Mature, glamorous, confident Korean idol aesthetic. No identity drift. WORLD: One unified red editorial studio with giant cream magazine pages, burgundy torn paper, red silk, geometric panels, chrome lettering, silver reflections, film strips, and contact sheets. Palette: scarlet, burgundy, cream, black, silver. 55% live-action fashion film, 45% dimensional motion graphics. Different typography and camera compositions across cuts. 7 LOOKS: Burgundy tweed jacket, cream top, red mini skirt. Scarlet fitted mini dress, black heeled boots. Cream top, high-waisted red skirt, ruby necklace. Black corset top, burgundy skirt, black boots. Red cropped leather jacket, black top and mini skirt. Silver-red metallic top, fitted black skirt. Scarlet off-shoulder couture gown, structured waist. 13 CUTS: 01 | 0–0.8s: Extreme eye close-up, raises gaze, burgundy paper. Large serif: "THE RED EDIT". 02 | 0.8–1.7s: Look 1, waist-up shoulder turn, cream page. Condensed text: "SCARLET". 03 | 1.7–2.6s: Look 1, beauty close-up, brushes hair, red editorial line. "SEOUL MUSE". 04 | 2.6–3.7s: Look 2, full-body, steps forward, giant scarlet panel. "RED SIGNAL". 05 | 3.7–4.8s: Look 3, folding magazine page reveal, adjusts necklace. Thin serif: "CRIMSON TOUCH". 06 | 4.8–5.9s: Look 4, seated pose, weight shift, black platform, burgundy silk. "BOLD MOOD". 07 | 5.9–7s: Look 5, three-quarter turn, brushes hair, silver reflective set. Chrome: "ICON GIRL". 08 | 7–8.1s: Look 5, chest-up, lowers chin then raises gaze, giant red typography wall. "강렬하게". 09 | 8.1–9.2s: Look 6, knee-up, subtle waist shift, black-and-white contact sheets of herself. "AFTER DARK". 10 | 9.2–10.3s: Look 6, burgundy torn-paper reveal, hand on waist, cream background. "OWN THE LOOK". 11 | 10.3–11.4s: Look 7, steps toward camera, minimal red-white editorial grid. "SEOUL MUSE". 12 | 11.4–12.7s: Look 7, slow knee-up push-in. Magazine pages, film strips, red silk, contact sheets, and chrome frames assemble around her. "THE RED EDIT". 13 | 12.7–15s: FINAL COVER. Graphics snap into an elegant Korean luxury magazine cover. She stands slightly right, turns, raises her chin, and looks into camera. Text: "THE RED EDIT / SEOUL MUSE / BOLD LOOK. SOFT POWER. / VOL. 2026". Final 0.5s completely still. MOTION: Slow, elegant model gestures; fast rhythmic graphics. Hard cuts, page flips, paper folds, typography slams, and contact-sheet flashes. Change outfits instantly between cuts, never morph. AUDIO: 120–124 BPM premium K-pop instrumental, punchy kick, clean bass, hi-hats, paper snaps, metallic clicks. No vocals or dialogue. AVOID: Extra people, face drift, hair changes, outfit morphing, outdoor scenes, dancing, static slideshow, repetitive framing, flat stickers, cheap graphics, cyberpunk, garbled text, logos, watermarks, subtitles. RESULT: One girl, seven looks, thirteen dynamic red luxury magazine spreads, ending in a polished THE RED EDIT cover.

Original post ↗

A hand-drawn tiny baker in a live-action kitchen

sections · MiniMax H3 per the author · Sep 30, 2026

The scale is built into every beat: each one pairs the small drawn baker with a full-size object he has to deal with (a flour bag, a stick of butter, a mixer attachment), so there is always something to measure him against. The style is stated as a hybrid of a hand-drawn character and a photoreal kitchen, and the camera list favours macro and low angles that keep the size contrast in frame. The beats carry no times, so pacing is left to the model; the post credits the concept to @AIwithSynthia.

Giants, armies and epic scaleAnimated styles and art direction

Read @airina_xyz's prompt
Subject
A miniature hand-drawn 2D animated boy with messy brown hair, round glasses, a blue apron, striped socks, and a tiny silicone spatula, baking cookies inside a realistic full-sized kitchen.
Style
Mixed-reality hybrid animation combining playful 2D hand-drawn doodle character art with a photorealistic live-action kitchen. Whimsical stop-motion physics, expressive line animation, exaggerated scale, warm cinematic lighting, and playful cooking sound effects.
Camera & Framing
Macro close-ups, low-angle tracking shots, overhead kitchen perspectives, dramatic push-ins, and shallow depth of field emphasizing the tiny baker against oversized kitchen objects.
Audio & Scene Breakdown
Visual: The tiny baker stands beside a giant bag of flour, trying to push it open with his little spatula. A burst of flour explodes around him like a snowstorm.
Visual: He scrambles onto the edge of a huge butter dish and slides down a stick of butter into a giant mixing bowl.
Visual: Inside the bowl, he dodges chunks of chocolate chips falling from above like boulders while cookie dough is being stirred around him.
Visual: He grabs onto a massive electric hand mixer attachment and spins in circles as the dough swirls beneath him.
Visual: He leaps out onto a baking tray and uses an oversized spoon to shape cookie dough blobs, getting comically stuck in one of them.
Visual: He rides a rolling pin like a balancing beam while trying to flatten one runaway cookie.
Visual: The camera pulls back to reveal a tray of freshly baked cookies. The tiny baker proudly places one chocolate chip on top of the biggest cookie—then the cookie suddenly cracks and tilts, sending him sliding off.
Audio: Playful kitchen ambience, flour poofs, dough squishes, mixer whirring, tray clinks, tiny footsteps, comedic impact sounds, subtle upbeat music, no dialogue.

Original post ↗

One ribbon that reveals a character from feet to face

Continuous action chain · MiniMax H3 per the author · Sep 25, 2026

One effect gets one job: a luminous ribbon that never disappears is the reveal mechanism, passing up the body region by region in a named order while the camera stays in a single take. The reference sets identity and rendering, and palette, effects and music are left for the model to derive from the character, so the same text can be reused with other characters. Light and sparkle are told to peak at the face before a pull-back to the first full-body view, which gives the one-take a build and a payoff.

Camera moves and transitionsReference-driven characters

Read @aimikoda's prompt
Create a high-impact single-take transformation-style character PV using  @[character ref]. Preserve identity, proportions, outfit and original rendering style exactly. Let the character determine the full visual language: palette, lighting, effect design, motion style and music. Treat the sequence as a spectacular stylized reveal of the existing character design, not a redesign or costume change.

The entire PV must be one continuous shot with no cuts. Use one flowing camera move that travels around the character in a stylish, dynamic way: smooth orbiting, rising motion, push-ins, slight reframes, parallax shifts and elegant angle changes, all within a single uninterrupted take. The movement should feel premium, controlled and cinematic, gradually building energy.

Use one continuous luminous ribbon as the main effect. The ribbon should remain alive for the whole sequence, never disappearing, constantly flowing around the character like a single uninterrupted stream of energy. It should travel across and around the body, tracing the silhouette, wrapping the legs, circling the waist, sliding past the torso, orbiting the shoulders and rising toward the face. Its path should feel intentional and graceful, as if it is guiding the viewer through the reveal.

Let the ribbon function as the reveal mechanism. As it passes each body region, it illuminates and reveals details in sequence: feet, legs, hips, waist, torso, shoulders, neck, jawline, eyes, then the full figure. The ribbon should leave behind glowing contour accents, colored reflections, spark trails and brief graphic echoes, creating a chain of revealed detail as it moves upward. Make the effect feel continuous and elegant, not fragmented.

Push the visual richness with more color and more light play. The ribbon should shift through layered colors that suit the character, such as vivid pinks, cyans, golds, violets, reds, teals or other fitting tones. Add dynamic lighting interactions throughout: sweeping highlights, moving reflections, lens flares, glossy gleams, rim-light pulses, flare streaks, refracted sparkles and soft color blooms. Let the light dance across skin, fabric, hair and accessories in a way that constantly refreshes the image.

Keep the character subtly active within the shot. Use confident body language and small elegant movements: a slight step, a turn of the hips, a hand gesture, a shoulder roll, a head turn, a lift of the chin, hair and fabric reacting with graceful follow-through. The ribbon should respond to these motions, tightening, whipping, looping or widening as the character moves. Make the effect feel driven by the character’s presence.

Build steadily toward the face. As the ribbon rises, let the camera move closer and the lighting become more intense. Around the eyes and face, increase sparkle density, glow intensity and color layering. This should feel like the emotional peak of the reveal before the final payoff.

At the end of the shot, pull or arc outward just enough to reveal the first full head-to-toe hero view. The ribbon completes one final elegant loop around the full body, then detonates into a spectacular colorful burst of light: expanding ribbons, flare bursts, sparkling fragments, radiant arcs, glowing rings and a brilliant energy bloom. This final explosion should be the visual climax, but the full figure must remain clearly readable inside it. After the burst, briefly hold the completed hero image.

Sync the entire take to a rising, glittering, high-energy musical build with layered whooshes, shimmering accents, bright flare sounds and one huge final impact. Keep the aesthetic graphic, luminous, colorful and stylized rather than realistic fire, smoke or destruction.

No cuts, no text, no typography, no captions, no logos, no redesign, no outfit replacement, no unrelated props.

Original post ↗

How we chose the clips

The examples are public posts on X, collected in August 2026 with a second batch from September and October 2026. A post became a case only if its thread holds a readable video prompt and the author's own text ties the video to MiniMax H3. We left out duplicate posts, prompts the author said came from someone else, near-identical variants of one template, posts that compare several models in one video, and posts we did not think suitable for a general audience. That left 61 cases from 14 creators. Model labels come from the authors' own posts, not platform logs; a few cases where the H3 link is only implied, for example through MiniMax's design agent or a caption cut short in our copy, are marked as not stated. We did not check the videos against their prompts frame by frame, and nothing here is a measured success rate.

Every case plays from the author's own post on X, and every prompt belongs to its author. Where an author wrote the prompt in English and we could match our copy word for word against their post, we quote it in full with credit and a link; otherwise we link to the post rather than translate it. The case notes, templates and checklists are our own analysis of what the prompts do structurally, so you can learn the method and write your own.

By creator: @aimikoda (14), @PixelAigc (10), @MayorKingAI (7), @airina_xyz (7), @CharaspowerAI (4), @techhalla (3), @umesh_ai (3), @LoveUolanda (3), @ou_zhen599 (3), @egeberkina (2), @AIwithJessica (2), @Yuki26856052 (1), @ponyodong (1), @eternityspring (1). Examples checked Oct 29, 2026.

If you are one of these creators and want a post left out or credited differently, write to [email protected] and we will change it the same day. Made a shot that worked? Send us the prompt and the video it made.

The week's best new dramas and AI movies, in one email.

We are putting together a weekly email: the new dramas and AI movies worth your time, and any free-coin deals we find. Leave your address and you will get the first issue when it starts. No spam.