How Do I Use Reference Images with the MiniMax H3 API?

Overview

Reference images are how you hold a character, style, or product steady across shots. In the MiniMax H3 API you attach them through the content array alongside your text prompt, so confirm the accepted count, format, and how each image is tagged in the current schema before building a multi-shot workflow around it.

If reference inputs are part of your production workflow, also account for any reference image charges in MiniMax H3 pricing when estimating the cost of each generation.

Source note: Verified 2026-08-06 against the MiniMax official H3 blog, MiniMax Video Generation API docs, and Hugging Face MiniMax-H3 model page.

Use the slots with intent rather than volume. Pick references that agree on identity, lighting, and framing, because conflicting images pull the result in different directions. Where you can label roles — this image is the character, this one the product, this one the scene — do so explicitly, since a clear role helps the model apply each reference to the right part of the output. A few clean, purposeful references usually beat the maximum number of noisy ones.

Keep the images at a supported resolution and free of clutter you do not want carried into the video, then let your prompt handle the action and camera while the references hold the look.

The request itself runs asynchronously like any other: submit with the references attached, store the task ID, poll for completion, and save the output. Test with one reference first so you can see how strongly it steers the result before adding more.