MiniMax Hailuo 03

One node, three endpoints

MiniMax ships Hailuo 03 as three separate APIs. This is one node, and what you wire decides which API the run goes to:

• nothing but a Prompt goes to text-to-video
• an Image goes to image-to-video, and that frame becomes the first frame
• anything in Reference Images or Reference Audio goes to reference-to-video

References win over a start frame. If you wire both, the run goes to reference-to-video and your start frame is added to the end of the reference list rather than being dropped.

Addressing references in the prompt

Reference assets are numbered by their position in each list, and you refer to them in the prompt by name: "Image 1", "Image 2", "Audio 1". A prompt that never names them leaves the model to guess what they are for.

The caps are 9 images and 3 audio clips, and 12 files combined. Anything past a cap is trimmed from the end of that list. Reference audio has to be between 2 and 15 seconds per clip.

Audio cannot be the only reference you wire. MiniMax refuses a run that has audio and nothing else, so pair it with at least one image.

No reference video, on purpose

The underlying API also takes reference video clips, and this node does not expose them. MiniMax bills reference footage per second of input at the same rate it charges for the video it renders, so a short render conditioned on a long clip costs more to run than it charges. Nothing measures the length of a clip before you are charged, so the port is left off rather than shipped at a price that is wrong. Reference images and audio are unaffected.

Aspect ratio only applies to two of the three

Aspect Ratio is used for text-to-video and reference-to-video. It is ignored for image-to-video, where the output follows the shape of the image you wired. Reference-to-video also accepts "adaptive", which lets the model pick from the references.

End Frame needs a start frame

End Frame gives you first-to-last keyframe control, and it is only read on the image-to-video path. Wiring an end frame on its own does nothing, because reaching that path requires an Image. Wire both.

Resolution and cost

Hailuo 03 only renders at 2K, so there is no resolution setting and no cheaper tier to drop to. Billing is per second of output at a flat rate, which makes this one of the more expensive video nodes in the catalog: the default 5 seconds costs roughly five times what a 1 second render would, and Duration goes up to 15. Duration is a whole number of seconds, minimum 5.