How to Use MiniMax H3:
Step-by-Step Tutorial for Text, Image & Reference to Video

Write one observable outcome before selecting a style: a product stays recognizable as the view changes, a portrait turns toward the viewer, or a character picks up an object. The outcome should be specific enough that another person could watch the clip and tell whether it happened. “A beautiful cinematic video” leaves that decision almost entirely open.

Separate the essential result from optional decoration. For a product shot, the bottle shape may be essential while drifting dust is optional. For a spoken line, completing the sentence may matter more than a complex camera move. If both cannot fit comfortably into the available clip, keep the requirement that makes the video useful.

Also identify the intended destination. A portrait-oriented social clip and a wide website banner need different compositions. Note where a face, label or hand must remain visible after cropping.

Check what your service offers first

Confirm the exact model name, input modes and audio options in the service you intend to use. Do not prepare a multi-reference sequence solely because a general article mentions references. First check whether the service accepts the files and roles your idea needs.

Done when: You can say in one sentence what a viewer must see or hear, and which crop must keep it visible.

If your idea can tolerate an invented scene, text may carry enough information. If a particular composition must open the clip, a suitable image provides a more concrete starting point. If the ending is the important constraint, prepare that ending deliberately rather than adding it as an afterthought to a first-frame workflow.

MiniMax’s video generation guide documents text, first/last-frame and reference workflows. The shorthand below also appears in its prompt-writing reference. Match the input role to your goal before uploading.

In the official API, frame inputs and reference inputs are separate modes and cannot be mixed in one request. Reference mode accepts up to nine images, three videos and three audio clips, with a combined limit of twelve files; each video or audio clip runs 2–15 seconds, and together they stay within 15 seconds. MiniMax-H3-Max does not yet offer reference to video (listed by MiniMax as coming soon). More inputs can introduce more disagreements. A product photograph that accurately defines a bottle may also contain a background you do not want. Decide whether it should supply the entire opening or just the product appearance, then use a workflow that supports that distinction.

Inspect the source at the scale you need

Open the original image and examine the important part. If the face is hidden behind hair, or the label is already unreadable, write down that limitation before requesting a result that depends on it. Check the crop, orientation and empty space around moving subjects.

For first-and-last-frame work, put the images side by side. List what changes: hand position, prop state, camera angle, light, background. A small intended action becomes a much larger transition if all five change at once. Keep deliberate changes, remove accidental ones, and describe the path between the remaining differences.

Done when: You have named the mode, and every file you plan to upload has one role you can state in a sentence.

Make a short event list before polishing the prompt. For a cup interaction, that could be approach, contact, lift, pause and release. Consider which steps must be visible for the viewer to understand the event. If the cup is already in a hand at the start, the approach and first contact may be unnecessary. If an object would seem to move before anything touches it, make contact and lift separate beats.

Read any spoken line aloud at the desired pace. Then leave room for the visual action before or after it. If the service offers a limited duration, shorten the scene or the sentence rather than assuming the model will fit every beat naturally.

Decide whether the camera stays still, moves through space, or changes framing with a zoom. Separately decide how the subject moves. Then list the audible events you actually need. The prompt field accepts up to 7,000 characters; a first version is easier to judge when it describes one shot. Keep the detailed language work in the prompt-writing guide and draft in the homepage editor; here the goal is a coherent action before entering it anywhere.

For multiple shots, identify why each cut exists. A cut that adds no useful information is another transition to manage. Start with a single shot when it can deliver the outcome, then add another only when the story needs it.

Done when: The prompt reads as a beginning, one visible change and an ending, and every sound has a source.

Using text or frame inputs? Skip to Step 5. This step applies to reference to video, where up to nine images, three videos and three audio clips, twelve files in total, stand in for parts of the description.

Prepare a plain-language role note for each asset: “this image supplies the jacket,” “this image supplies the room,” or “this clip supplies the pace of camera movement.” Use the actual labels assigned by the service when entering a prompt. Writing a filename in a text box does not upload that asset or assign its role.

MiniMax’s full-reference guide distinguishes visible subjects, frame anchors, video sources and audio sources. It also separates reusing audio from referring to its characteristics.

Choose a winner when sources disagree

Suppose one picture shows your character wearing a green coat and another shows the same character in a brown coat. If the second picture is only a pose reference, state that the green coat should remain. Otherwise, there are two equally plausible wardrobe instructions. Make the same decision for hair, material, lighting and background when they matter to the shot.

Two environment references can also compete. If one places the subject in a bright kitchen, a separate dark warehouse reference needs a narrowly defined role or a deliberate transition. Removing one source may be clearer than writing a longer compromise. Use these as reference assets within reference mode, not as a mixture of first-frame and reference inputs.

Listen to sound references independently

If a reference is intended to supply only a delivery style, decide what the new words should be. If you want the original soundtrack reused, check whether the service supports that operation and whether you have the necessary right to use it. A video containing sound is not, by itself, an instruction to retain that sound in the output.

Before submitting, read the role notes without looking at the assets. If “reference one” changes meaning halfway through, rewrite the notes. If two files provide the same detail, keep both only when you can explain what the second one adds.

Done when: When you read your role notes without the files, no two files claim the same job.

Use these official model specifications to plan the shot, then select the corresponding controls in your service. These are MiniMax’s published API options, not a promise of settings in the generator planned for this site.

The official API reference requires an explicit ratio for text to video: 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16. Image to video follows the input image using adaptive ratio, and first and last frame images each need an aspect ratio between 2:5 and 5:2. Reference generation can use adaptive sizing or an explicit ratio. A number written into a prompt does not select one of these output settings.

• Model and mode: check the displayed model and the selected input workflow.
• Assets: verify upload completion, preview the files, and confirm opening, ending or reference roles where those controls exist.
• Composition: if aspect ratio or crop controls are available, make sure they leave essential subjects visible.
• Timing and sound: compare the available duration with your event list, and verify any audio option required by the brief.
• Estimate and submission: read the actual estimate and applicable account requirements before confirming an attempt. Keep a copy of the prompt outside the page.

After submitting, use the service’s displayed task status. If processing takes longer than expected, check whether an attempt already exists before pressing submit again. If an error appears, preserve its message and the task reference, if provided. A rejected upload, an unavailable account feature and an unsatisfactory generated clip are different problems and need different responses.

Once an output is available, save its reference alongside the inputs so the review can be tied to the correct attempt.

Done when: The model name, mode, uploads, ratio and duration on screen match your brief, and the prompt is saved outside the page.

First watch at normal speed without pausing. Is the intended event understandable? Then inspect the important transitions: first contact with an object, a turn of the face, a cut, or the final pose. A single attractive still does not tell you whether the action remains coherent between frames.

Make a separate pass for appearance and geometry, another for camera and crop, and another for audio. Compare against the exact source inputs when identity or endpoint composition matters. Write observations as visible or audible facts before assigning a cause. “The cap changes width halfway through” is an observation; “the reference is too weak” is a hypothesis.

Match what you noticed against the troubleshooting table below, which pairs each symptom with the published limit or writing habit most likely behind it and the one change to try first.

Done when: You have one written observation that names where in the clip the problem is visible or audible.

Save the exact prompt, source files, available settings and the output together. Then choose the mismatch that matters most to your required outcome. If a label becomes unreadable, extra atmosphere is not the priority.

This copyable record is a personal working template. Fill only the fields your service actually exposes. Compare saved files at the same size and orientation; a preview, download variant or playback setting can make one result look different. Do not include account credentials or private billing details.

Attempt name:
Date:
Provider and exact model label:
Selected input mode:
Input filenames and the role of each file:
Exact prompt:
Settings actually available and selected:
Required outcome:
Result link or saved filename:
Observation and location in the clip:
One change for the next attempt:
Outcome of that change:
What remains uncertain:

Make one substantial change between attempts when practical. Keep unrelated wording stable so you can explain what you were testing. If you change the image, action, camera and audio at once, a better result may be welcome, but it will not tell you which edit helped.

Review the new output against the same required outcome, then record both improvements and remaining problems. One successful attempt is not proof of consistent performance; evaluate more than one result before describing a method as reliable.

Know when to simplify or stop

Stop adding instructions when they begin to contradict the original purpose. A simpler shot with a readable action may serve your project better than a complicated sequence whose key event is hidden. If the feature you need is unavailable in the selected service, choose an available approach rather than trying to invent a control through text.

For a ready-to-adapt brief, open the sixteen starter prompts; to adapt one to your own product or person, follow the prompt guide. Keep this page alongside your own attempts as a checklist, not a guarantee of a particular result.

Done when: The next attempt differs from the last in exactly one deliberate way, and the record says what that was.