MiniMax H3 Reference to Video: Character Consistency Guide

MiniMax documents reference generation using images, video and audio. Its full-reference writing guide separates referenced subjects, retention decisions, the scene timeline and sound. A practical starting point is a brief asset list with a clear responsibility for each file.

For a character scene, an image might define appearance while a video suggests a walk. A second image could define the location. Decide whether the walk reference’s clothing or background should carry over; leaving that unresolved creates competing instructions.

Start with the smallest set that explains your scene. More files are useful only when they supply missing information. Two character pictures with different costumes may make the outfit less clear unless you explicitly select which one governs it.

For an original clay robot, list the features that make it recognizable: a square orange head, two round dark eyes, a short blue body and wide feet. Pick a source image in which those features are visible. If the next shot needs the robot’s back, consider whether your assets explain that view.

Separate identity requirements from optional scene decoration. A new room should not imply a new head shape. A different lighting setup can alter the apparent color, so compare shape and proportions as well as surface color when reviewing results.

Keep a copy of this checklist with each accepted shot. For a sequence of clips, compare the final frame of one with the opening of the next. A character that looks convincing in isolation can still change proportions at the edit.

Prepare one image of your robot and, optionally, a short motion reference you are entitled to use. Map the asset names below to your provider’s actual upload labels. These names are descriptive placeholders, not universal syntax or a complete structured rewrite.

This original draft is untested. It is designed to make appearance, motion and environment separate review questions. First try the character image alone if the walking action can be described adequately in words.

Original prompt draft — adapt to your inputs; not a tested generation.

Use the character image for the clay robot’s square orange head, round dark eyes, short blue body and wide feet. If supplied, use the motion clip only for its slow alternating steps, not its actor, clothing or background. Place the robot in a miniature white-tile hallway. In one full-body shot, the robot walks toward a red door, stops beside it and raises its right hand. Keep the head shape, body proportions and clay texture consistent. The camera stays still. Add soft foot taps and quiet indoor ambience; no speech or music.

Decide whether you need a specific vocal quality, an existing sound event or simply a description of ambience. The robot scene needs only footsteps, so an extra audio file may contribute nothing to the brief.

If you add a sound reference, describe the aspect it should influence and the moment it belongs to. Also decide whether sound from a motion reference should be used. An unrelated music track inside a walking clip should not silently become the plan for your scene.

Review with headphones after checking the picture. Mark whether the sound starts with the action, continues after it, or masks something important. Revise the sound instruction separately from the character design when the visual result is already useful.

If the robot inherits the motion actor’s outfit, revise the role assignment: appearance comes from the image, steps come from the clip. Do not immediately add another costume image. First remove ambiguity from the material you already supplied.

If the background follows the character image, describe the intended hallway independently and limit the image’s role to the robot. If a turn changes the face, reduce the requested turn for a baseline and inspect whether the source actually shows enough information for that viewpoint.

When several things fail, rank them. Fix identity first if that is the purpose of the scene; then adjust the action, camera and sound. Save each prompt and result so that a successful revision can be distinguished from an unrelated change in settings.

Write a short continuity note after selecting a shot: which hand is raised, where the robot faces, and where the door sits in the frame. Carry that state into the next brief. Reusing the character image alone does not describe the action state at a cut.

Use a simple contact sheet of the opening and closing frames to compare silhouette, wardrobe and screen direction before assembling the edit. Then watch the joined clips at normal speed. A noticeable mismatch is a reason to revise the affected shot, not proof that every asset must be replaced.

Choose endpoint control when a precise opening or landing is the main requirement. Choose references when selected characteristics need to guide a new composition. Check the selected service’s supported modes before building a workflow around a particular combination.

Official documentation supports the input roles and prompt conventions discussed here. Preparation tips, examples and review checklists are our editorial recommendations. We have not run comparative tests or generated the example prompts.

• MiniMax video generation documentation
• MiniMax official full-reference prompt guide

Sources checked September 13, 2026. Follow your selected provider’s current interface and requirements. Read our editorial method.