MiniMax H3
Separate ordinary H3 generation from the two exclusive reference workflows.
Use H3 for a scene and optional still
The minimax_h3 preset exposes Text and Image inputs, not audio or video inputs. The ordinary H3 service accepts text and up to two image references; use it for a scene that does not need a speech track.
- Place Video Generation and select minimax_h3; write a short scene with one subject, one action, and a stable camera.
- Connect a useful image, choose its @ thumbnail, and run a single shot before expanding the prompt.
Use Context IR without losing the brief
The H3 canvas has an Enhance Prompt action backed by MiniMax Context IR and a 7,000-character prompt limit. Enhancement is a drafting aid, not a substitute for checking continuity constraints.
- Enter “A baker closes the shop at dusk; one slow push-in, flour on apron,” then press Enhance Prompt if the route is available.
- Read the replacement text and remove invented people, cuts, or colors before running the video.
Reference mode: one still and one sound
minimax-h3-reference is a different preset. Without video references, its server requires exactly one image and one audio reference plus a nonempty prompt; a lone image will fail.
- Select minimax-h3-reference, connect one portrait to Image and one recorded line to Audio, then select both in @.
- Describe how the visible subject reacts to the line; check the preview for lip motion and for the identity of the selected portrait.
Reference mode: videos with their own audio
The other reference mode accepts one to three videos, each containing its own audio track. Their combined source duration cannot exceed 15 seconds; images and separate audio are forbidden in this mode.
- Connect a short video with audible sound to Video and choose it in @; remove previously selected Image and Audio items.
- If video inspection reports no audio stream, use a source that actually contains sound rather than attaching a separate Audio node.
Fix a reference rejection
The reference service checks duration, file type, source accessibility, and video audio tracks before submission. Its output ratio follows the selected image or video rather than an independent ratio control.
- Use the displayed 4–15-second choice for reference output; shorten source videos if their combined runtime exceeds the limit.
- If the server rejects the input combination, clear all @ selections and rebuild only one of the two valid modes.