MINIMAX REFERENCE WORKFLOW
MiniMax H3 Max Reference: What the Reference Mode Actually Does
Reference mode guides a new shot with images, motion clips and audio. It is different from animating a fixed first frame. Learn how to prepare the inputs, then try MiniMax H3 or H3 Max below. The studio starts with MiniMax H3, and keeps H3 Max as a separate model choice.
Open the MiniMax studioSongLee is an independent studio using third-party MiniMax models through APIMart. Model capabilities and provider availability may change.
Output resolution
H3 · 768P or 2K
H3 Max · 480P, 768P or 1080P
Requested duration
H3 · 4–15 seconds
H3 Max · 5–15 seconds
Reference inputs
H3 · Images, videos and audio
H3 Max · Images, videos and audio
Try MiniMax H3 reference-to-video
Select H3 or H3 Max, upload your own permitted references, describe one shot and review the credit estimate. Multi-reference requires Pro or Max; image-to-video and text-to-video are also available. This is a credit-based tool, not a free MiniMax trial.
Video result
Your generated video will appear here
Choose a MiniMax model and upload permitted media. Review the credit estimate and plan limits before generating.
One reference image, one clear action
This example uses an original AI-created ceramic vase image. The reference supplies the product and setting; the prompt asks for a slow camera push-in and subtle flower movement. It demonstrates the input-to-output workflow, not guaranteed exact matching across every generation.
Example prompt
Use the vase in reference image 1 as the subject. Preserve its lavender ceramic material, asymmetric shape, three pale flowers and stone pedestal. A slow smooth camera push-in in a bright studio, subtle natural movement of the flower stems, soft daylight, one continuous shot.
After uploading your own image, adjust the subject and setting to match it. The product and motion presets also need suitable reference media.
MiniMax H3 vs H3 Max: choose the right model
Both models support text, first/last frames and multimodal references through our provider. Their duration and resolution options differ; selecting H3 Max does not silently call H3, or the other way around.
| Setting | MiniMax H3 | MiniMax H3 Max |
|---|---|---|
| Output resolution | 768P or 2K | 480P, 768P or 1080P |
| Requested duration | 4–15 seconds | 5–15 seconds |
| Reference inputs | Images, videos and audio | Images, videos and audio |
| First / last frames | Separate image-to-video mode | Separate image-to-video mode |
| Sound | Native model audio | Native model audio |
What each reference input controls
Images give the model visual cues such as a subject's appearance, clothing, a product or a setting. Use sharp, consistent references and say which image supplies which detail. Reference images influence a new composition; they do not pin the exact first frame.
Videos provide motion and camera cues. Use a short clip showing the movement you need, with one main action. Audio provides sound or speech cues and must accompany at least one reference image or video; audio alone is not supported.
The provider accepts up to 9 reference images, 3 videos and 3 audio clips, with at most 12 reference files combined. Each video or audio clip must be 2–15 seconds, and each media type must total no more than 15 seconds. SongLee also applies your plan's lower reference limits. Images here accept JPG, PNG or WebP up to 15 MB; for MiniMax use MP4 video up to 50 MB and MP3/WAV audio up to 15 MB.
Reference-to-video vs image-to-video vs text-to-video
Choose reference-to-video when appearance, motion or audio should guide a newly composed shot. Upload the reference media in Multi-reference mode and explain how each input should be used. A reference image is not an instruction to start on that exact image.
Choose image-to-video when your picture should become the starting frame, or when you need first/last frame control. Upload one or two pictures in the frame fields. The image determines the aspect ratio in this mode, even if another ratio is selected.
Choose text-to-video for a scene described entirely in words. Do not mix first/last frame images with reference images, videos or audio in the same MiniMax request. The studio separates these workflows and validates the request before reserving credits.
Reference strength and identity drift: common fixes
There is no reference-strength slider in this integration. Control the result by simplifying your inputs and explicitly assigning their roles in the prompt. Start with one clear subject and one action, then add a motion or audio reference only when you need it.
If identity drifts, use references with the same person, outfit, lighting and viewpoint. Ask the model to preserve specific visible details. Conflicting faces or costumes, a tiny subject and rapid camera changes make the task harder; extra references do not automatically improve consistency.
If the shot ignores your first image, check the selected mode. Reference mode can recompose the scene; switch to first-frame image-to-video when the opening frame matters. For distorted hands or movement, try a shorter clip and slower action before increasing resolution.
Prepare a reference shot in three steps
1. Collect only media you own or have permission to use. Pick a clear subject image; optionally add a brief motion clip and permitted audio. Keep all inputs focused on the same intended scene.
2. Write the shot in order: reference role, subject, action, camera and lighting. For example: use image 1 for the subject, video 1 for the camera movement, and keep the courtyard background stable. Describe a manageable action instead of a whole sequence of scenes.
3. Select the model, duration and supported resolution. Check the displayed credits before generating. Reference video adds input-duration credits, and extra images can add credits too. Credits are reserved at submission and released if the task fails. Review the final video for visual errors before publishing.
MiniMax H3 Max reference questions
Input limits checked against the provider documentation on October 6, 2026: MiniMax H3 · MiniMax H3 Max