Effective Hailuo prompts describe one readable shot: a clear subject, a visible action, a controlled setting, and an intentional camera movement. MiniMax also documents 15 bracketed camera commands for specific text-to-video, image-to-video, and first-and-last-frame models. This guide shows how to use those commands without claiming results that have not been generated and reviewed.
Independent-site notice: MiniMax-AI.chat is an independent educational website and is not MiniMax, Hailuo, or an official MiniMax Open Platform property. The examples below are editorial guidance based on MiniMax’s official platform documentation.
Last verified: August 13, 2026. Verification status: the Hailuo V1 syntax, parameter names, routing, and documented camera-command compatibility below were rechecked against official MiniMax references. MiniMax now lists
MiniMax-H3as the current video model and Hailuo 2.3, Hailuo 2.3 Fast, and Hailuo 02 as legacy. H3 uses the V2 endpoint, whose reference does not document the legacy 15-command compatibility list. No paid generation task was run for this article. The prompts are test-ready examples, not tested-output or quality claims.
The MiniMax-AI.chat demo is text-only, accepts up to 1,000 characters, and does not render video. To evaluate motion or visual quality, submit the prompts through your own MiniMax Open Platform account and record the result. The official API prompt fields covered here allow up to 2,000 characters.
A practical Hailuo prompt structure
Start with the minimum information needed to stage one shot. A useful order is:
- Subject: identify the main person, object, animal, or scene.
- Action: use a visible verb and one primary action.
- Setting: state the location and the spatial relationship that matters.
- Light and look: name only the visual qualities that materially affect the frame.
- Camera: add one documented bracket command when the chosen model supports it.
- Continuity constraint: describe a single shot, stable subject, or absence of a scene cut when that is important.
[Camera command] Subject + visible action + setting + light/look + continuity constraintExample structure:
[Push in] A ceramic teapot releases a thin ribbon of steam on a walnut table, soft morning window light, one continuous product shot.This is a test-ready prompt. It defines what to inspect—forward camera movement, product stability, steam motion, and shot continuity—but it does not promise that a generation will satisfy every instruction.
Choose the mode and model before writing camera syntax
| Generation route | Documented model and command status | Prompt implication |
|---|---|---|
| Current multimodal V2 | MiniMax-H3; the V2 reference does not document the legacy 15-command list | Express camera intent in natural language and validate the result; do not assume Hailuo bracket syntax is supported. |
| V1 text-to-video | MiniMax-Hailuo-2.3, MiniMax-Hailuo-02, T2V-01-Director | Write the full composition because there is no source frame. Hailuo 2.3 Fast is not listed for text-to-video. |
| V1 image-to-video | MiniMax-Hailuo-2.3, MiniMax-Hailuo-2.3-Fast, MiniMax-Hailuo-02, I2V-01-Director | Let the source image carry appearance and composition; use the prompt mainly for motion, camera, and continuity. |
| V1 first-and-last-frame | MiniMax-Hailuo-02 | Describe the transition and motion between two required frames. This is a legacy-model route. |
| V1 subject-reference | The S2V-01 endpoint reference checked for this guide does not document the 15-command compatibility list. | Do not promise bracket-command behavior for S2V-01. Use clear natural-language motion and verify it in your account. |
Other model IDs appear in the text-to-video and image-to-video endpoint model lists, but the camera-command sections name the compatible models above. Do not assume bracket syntax works on an unnamed model. The MiniMax Video API guide provides the full mode matrix and task workflow.
All 15 documented MiniMax camera commands
| Movement family | Exact commands | What to evaluate |
|---|---|---|
| Horizontal camera translation | [Truck left], [Truck right] | The camera moves sideways while its viewing direction can remain broadly consistent. Check parallax between foreground and background. |
| Horizontal rotation | [Pan left], [Pan right] | The camera turns horizontally from a position. Check whether the view reveals adjacent space rather than translating sideways. |
| Physical move toward or away | [Push in], [Pull out] | Check for perspective change and depth, not only subject enlargement. |
| Vertical camera translation | [Pedestal up], [Pedestal down] | The camera position rises or falls. Inspect changing height and parallax. |
| Vertical rotation | [Tilt up], [Tilt down] | The camera angles upward or downward from a position. Inspect the revealed vertical area. |
| Optical framing change | [Zoom in], [Zoom out] | Check subject scale with less perspective change than a physical push or pull. |
| Intentional instability | [Shake] | Check whether controlled vibration supports the scene instead of obscuring the subject. |
| Subject following | [Tracking shot] | Check whether the camera follows the moving subject while preserving readability. |
| Locked camera | [Static shot] | Check that environmental or subject motion occurs without an unwanted camera move. |
MiniMax also accepts natural-language camera descriptions, but the documentation says explicit commands provide more precise control. The table reproduces the documented capitalization and spelling. Because the reference does not define case-normalization behavior, preserve that exact syntax in your tests.
Simultaneous and sequential movements
For simultaneous movement, place multiple commands inside the same brackets. MiniMax recommends no more than three commands in one combination:
[Pan left,Pedestal up] A climber reaches the ridge as the valley opens below.For sequential movement, place commands in the intended order through the prompt:
[Push in] A watch rests on dark stone under a narrow light. Condensation moves across the surface. [Pan right] Reveal the engraved crown, one continuous shot.Combined movements increase the number of things that can go wrong. Establish each movement alone before testing a combination, and use two commands rather than the documented maximum of three unless the shot genuinely needs more.
Test-ready Hailuo prompt examples
Each example states a motion hypothesis to inspect. It does not report an output. Keep the model, source image, duration, resolution, and optimizer setting fixed when comparing variants.
1. Static product shot
[Static shot] A cobalt perfume bottle on wet black stone, soft rain ripples in the background, a narrow rim light outlines the glass, one continuous shot.Inspect: whether the frame remains locked while water and reflections move; whether the bottle shape stays stable.
2. Physical push-in
[Push in] A violin maker adjusts a bridge at a wooden workbench, warm side light, shallow depth of field, slow deliberate hand movement.Inspect: forward perspective change, hand coherence, and whether the camera motion remains smooth.
3. Pull-out reveal
[Pull out] Begin close on an astronaut's reflective visor, then reveal a silent greenhouse filled with tall plants, cool lunar light, one continuous reveal.Inspect: growing field of view, scene continuity, and whether newly revealed geometry remains coherent.
4. Pan across a location
[Pan right] A quiet night market opens for the evening, paper lanterns glow above the stalls, vendors arrange fruit, realistic ambient movement.Inspect: horizontal rotation, the readability of newly revealed stalls, and the absence of an unintended scene cut.
5. Truck for foreground parallax
[Truck left] Move beside a row of wildflowers as a cyclist passes on a distant path, late-afternoon light, foreground petals create clear parallax.Inspect: sideways translation and relative movement between the flowers, cyclist, and background.
6. Tilt reveal
[Tilt up] Start on rainwater crossing a stone plaza, reveal a tall museum facade through light mist, overcast daylight, slow motion.Inspect: upward rotation from ground to facade. Compare with a separate [Pedestal up] variant to determine whether the model changes camera height rather than angle.
7. Pedestal movement
[Pedestal down] Descend from the canopy of a greenhouse to a botanist labeling seedlings below, soft diffused light, stable vertical move.Inspect: vertical camera-position change, foreground parallax, and the botanist’s identity stability.
8. Zoom-versus-push comparison
[Zoom in] A red fox pauses at the edge of a snowy forest, pale dawn, fine snow falls, the fox remains centered.Inspect: subject enlargement and perspective. Run a second version that changes only [Zoom in] to [Push in]; do not assume those two moves are visually equivalent.
9. Tracking a moving subject
[Tracking shot] A runner in a blue jacket follows a coastal trail from left to right, wind moves the grass, the camera follows at a steady distance.Inspect: subject following, framing distance, limb coherence, and background motion.
10. Controlled shake
[Shake] A rescue vehicle crosses uneven ground through heavy rain, headlights cut through spray, the vehicle stays readable in frame.Inspect: whether vibration communicates rough terrain without destroying subject readability. Use shake sparingly; stronger wording is not a guarantee of a better result.
11. Simultaneous two-axis move
[Pan left,Pedestal up] A dancer turns in a sunlit atrium as the camera reveals the upper glass roof, flowing fabric, one continuous shot.Inspect: whether both horizontal rotation and vertical translation are visible. Test each command separately first so a failed combination can be diagnosed.
12. Image-to-video preservation prompt
[Truck right] Preserve the exact package shape, label placement, and tabletop composition from the input image. Move slowly as soft reflections travel across the surface, one continuous product shot.Inspect: source-image fidelity, label stability, parallax, and whether the prompt adds motion without restaging the product. “Preserve” is an instruction to evaluate, not a guarantee.
How prompt writing changes by mode
Text-to-video: describe the shot, not a screenplay
With no source frame, establish the subject, location, action, light, and camera. Avoid packing several locations or time jumps into a short clip. A six-second request is easier to evaluate when it contains one scene and one primary motion.
Image-to-video: spend words on motion
The first frame already defines appearance and framing. Describe how the subject, environment, and camera should move, plus the parts that should remain stable. The official image rules are JPG, JPEG, PNG, or WebP under 20 MB, a short edge over 300 pixels, and an aspect ratio from 2:5 to 5:2. Hailuo 2.3 Fast belongs to this route.
First-and-last-frame: write the transition
Both endpoints of the shot are supplied as images, so the prompt should describe the movement and continuity between them. The dedicated reference lists only legacy MiniMax-Hailuo-02; it does not list Hailuo 2.3 or 2.3 Fast. Read the first-and-last-frame API guide before preparing a production payload.
Subject-reference: prioritize identity and action
The documented subject-reference route uses S2V-01. Describe the referenced character’s action and scene clearly, but do not advertise the 15 bracket commands as supported on this route without an official compatibility statement. The S2V-01 subject-reference guide explains its distinct input schema and availability caveat.
Prompt optimizer settings and controlled comparisons
prompt_optimizer defaults to true in the referenced video endpoints. That can be useful for ordinary generation, but an optimizer can introduce another changing variable. Set it to false when you need to compare two precise prompt formulations, then document the setting with the result.
fast_pretreatment defaults to false. The text-to-video reference documents it for Hailuo 2.3 and Hailuo-02; the image-to-video reference documents it for Hailuo 2.3, Hailuo 2.3 Fast, and Hailuo-02. It only applies when prompt optimization is enabled. Treat it as an API configuration, not a substitute for a clear prompt.
A repeatable A/B test for MiniMax video prompts
- Lock the route: use the same mode and exact model ID.
- Lock the media: reuse the identical source image or frames; record a file hash if possible.
- Lock API settings: keep duration, resolution, optimizer, and pretreatment settings unchanged.
- Change one variable: for example, compare
[Push in]with[Zoom in], or a natural-language move with one bracket command. - Save evidence: record prompt text,
task_id, model, settings, input asset, and generation timestamp. - Score defined criteria: use the same rubric for camera fidelity, subject stability, action fidelity, continuity, and visible artifacts.
- Repeat when budget permits: generative outputs vary. Multiple runs help distinguish a repeatable tendency from one sample, but do not describe them as statistically conclusive without a suitable design.
| Record | Variant A | Variant B |
|---|---|---|
| Exact prompt | Paste before running | Paste before running |
| One changed variable | Baseline | Name the change |
| Model / mode | Record exact ID | Must match A |
| Duration / resolution | Record values | Must match A |
| Optimizer / pretreatment | Record booleans | Must match A |
| Task ID | Record after submission | Record after submission |
| Camera fidelity, 1–5 | Score after review | Score after review |
| Subject stability, 1–5 | Score after review | Score after review |
| Continuity, 1–5 | Score after review | Score after review |
| Artifact notes | Observed evidence | Observed evidence |
Do not invent a negative_prompt parameter
The MiniMax T2V, I2V, first-and-last-frame, and subject-reference request schemas checked for this guide do not document a separate negative_prompt field. Do not copy that parameter from another provider and assume MiniMax accepts it. State important constraints in the main prompt—such as “one continuous shot” or “the product remains centered”—and evaluate whether the output follows them.
Constraints are guidance, not enforcement. Use post-generation review for brand accuracy, safety, identity, text rendering, and legal clearance.
Prompt troubleshooting checklist
| Problem to investigate | First checks | Controlled next test |
|---|---|---|
| Bracket command appears ignored | Confirm the exact model is named in the official compatibility list; preserve spelling and brackets. | Use one command, one action, and prompt_optimizer: false. |
| Camera motion is ambiguous | Check whether the prompt mixes pan with truck, tilt with pedestal, or zoom with push. | Run two variants that differ only in the camera command. |
| Too much movement | Remove combined commands and extra scene actions. | Start with [Static shot], then add one movement. |
| Image-to-video restages the scene | Check source-image quality and whether the prompt unnecessarily redescribes appearance. | Keep only motion, camera, and preservation instructions. |
| First and last frames mismatch | Check dimensions and composition. The output follows the first frame, and a differently sized last frame is cropped. | Prepare matched frames and use legacy Hailuo-02. |
| Request is rejected | Check the 2,000-character limit, model-mode pairing, image type, size, dimensions, and aspect ratio. | Submit the smallest valid payload from the relevant endpoint reference. |
Why these are not labeled “tested Hailuo prompts”
A credible tested-example claim needs the exact prompt, model ID, input asset, API settings, task evidence, output, date, and evaluation method. This article verifies documented syntax and prepares prompts for that process, but it does not fabricate output evidence. After running your own tests, publish the full method and representative failures alongside successful outputs rather than presenting one selected clip as a universal result.
Hailuo video prompt FAQ
How long can a MiniMax video prompt be?
The official T2V, I2V, first-and-last-frame, and S2V references checked here document a maximum of 2,000 characters for the prompt field. The independent text-only demo on MiniMax-AI.chat has a separate 1,000-character limit.
How many camera commands can I combine?
MiniMax recommends no more than three commands in one simultaneous combination. For reliable diagnosis, start with one, validate it, then add a second only when the shot requires it.
Does Hailuo 2.3 Fast support text-to-video prompts?
The legacy V1 references list Hailuo 2.3 Fast under image-to-video, not text-to-video. Within a Hailuo V1 integration, use standard Hailuo 2.3 for text-to-video. For a new text-to-video integration, evaluate current MiniMax-H3 on V2 first.
Should I leave prompt optimization enabled?
It defaults to enabled and can help ordinary generation. Disable it for a controlled A/B test when you need the submitted wording to remain the deliberate variable.
Are natural-language camera instructions supported?
Yes. The official references say natural-language descriptions can be used, while explicit bracket commands provide more precise control on the named compatible models.
Related implementation guides
- MiniMax Hailuo 2.3 model guide
- MiniMax Hailuo 2.3 Fast model guide
- MiniMax Hailuo-02 legacy model guide
- MiniMax API error-code guide
Official MiniMax sources
- Current model catalog and legacy status
- MiniMax-H3 Video Generation V2 create reference
- Video generation modes and workflow
- Text-to-video prompt and camera-command reference
- Image-to-video input, prompt, and camera-command reference
- First-and-last-frame reference
- Subject-reference S2V-01 reference
Scope: prompts influence a generative system but do not guarantee a particular camera path, identity match, continuity level, or production suitability. Review every output before use.
More on Hailuo: see the Hailuo AI hub for every model compared side by side, or Is Hailuo AI free? for what each access route actually costs.