MiniMax Reference-to-Video: Combine Character, Motion, and Style Inputs

Assign a role to every reference

Write a short note explaining what each asset contributes. A portrait may guide character appearance. A landscape may guide location. A video may guide motion or camera movement. An audio sample may guide sound or rhythm. Identify the primary reference when two assets overlap. This role assignment prevents the model from treating every detail as equally important. It also creates a review standard: the subject should follow one source while the camera follows another. Without that structure, an output can borrow the wrong property and still look superficially consistent with the collection of inputs.

Official context: H3 V2 API reference; Current video-generation guide.

Remove references that conflict with the brief

A reference set should support one coherent target. If one image shows warm daylight and another shows harsh neon lighting, explain which lighting should govern the shot or remove the irrelevant input. A slow motion reference and a rapid editing reference can pull pacing in different directions. Avoid adding files simply because they are attractive. Each additional input should resolve a meaningful ambiguity. A smaller well-defined set is often easier to evaluate than a large mood board. This is a production recommendation, not a claim that the model cannot accept multiple sources.

Follow the current input rules

Use the documented reference roles for images, videos, and audio. The current V2 API separates reference generation from first-frame and last-frame generation, so do not combine those mode-specific roles in one request. Validate the allowed file types, size, duration, and count against the live reference before submission. A failed request may result from those constraints rather than the creative description. Keep technical validation separate from visual review. This makes the workflow easier to debug and avoids repeatedly changing the prompt when the actual issue is a file that the endpoint cannot accept.

Describe relationships rather than listing files

A useful prompt might ask for the subject's appearance from the first image, the environment from the second image, and the gentle camera movement from the video. Specify what should not transfer, such as the reference video's actor or background. This example is original editorial wording. Its purpose is to establish relationships. Avoid a prompt that merely says use these references, because the model has to guess the role of each source. Clear relationships make feedback more precise. You can identify whether identity, environment, or movement missed the intended reference instead of judging only the overall visual impression.

Keep the first test simple

Use one subject, one environment, and one primary motion for the initial experiment. Avoid adding dialogue, rapid cuts, and a complex transformation until the core reference relationships work. A simple shot provides a clean test of whether the model preserved the important characteristics. If the subject drifts, strengthen the identity direction or choose a more suitable reference. If the motion is wrong, revise that portion while keeping other inputs stable. This controlled approach helps the team learn from each candidate and reduces the cost of diagnosing a failure with many possible causes.

Review each dimension independently

Create a checklist for identity, location, movement, framing, sound, and continuity. Compare the output with the relevant source for each dimension. A clip may follow the camera reference while changing the subject's clothing. Another may preserve the subject but ignore pacing. Record those outcomes separately. Do not treat a strong result in one dimension as evidence that the whole brief passed. Review the entire duration and inspect difficult moments, such as turns or interactions. Independent dimension checks help you select usable candidates and explain revision requests in terms that a designer or client can understand.

Respect the intended use of source material

Use references that you are entitled to provide and that fit the project's approved creative direction. For a brand project, distinguish an inspiration reference from a request to reproduce another brand's protected assets. If a person's likeness or voice is involved, establish the permission and intended use before generation. Keep those records with the project. This is part of managing source assets, not a substitute for reviewing applicable terms. A reference workflow should preserve clarity about where inputs came from and what the resulting content is allowed to represent, especially when several contributors exchange files.

Create a reproducible handoff

Save the input manifest, prompt, selected model, mode, settings, and accepted clip. Include a small preview of each reference so another person can understand the relationship quickly. Record which details were deliberately changed and which were required to remain stable. When delivering to an editor, identify any known limitation that needs finishing work. A reproducible handoff helps the next contributor make variants without rediscovering the original intent. It also prevents an old reference set from being mistaken for the current approved direction when the project evolves over several rounds of feedback.