MiniMax H3 Video Generation — multimodal references, any length from 5 to 15s, 2K output
MiniMax H3 takes images, video and audio as references in the same prompt, outputs 2K, and lets you pick any whole-second duration from 5 to 15. On XinYu AI it shares one credit balance with every other model.
Available builds
- MiniMax H32K output; any whole second from 5 to 15; text-to-video, image-to-video, reference-to-video and multimodal reference; end-frame supported
What it's good at
Images, clips and audio as references together
H3's multimodal reference isn't image-only — video segments and audio tracks go into the same prompt to set tone and pacing.
Any length from 5 to 15 seconds
Not just 5s and 10s presets: every whole second in between is selectable, so the cut doesn't need padding or a hard trim.
Pin the end frame
Give a start and an end frame so the shot lands where you want it — which makes the next shot easier to join.
Why run it on XinYu AI
One balance, every model
Image, video, audio and text draw from the same credit balance — no separate subscription per model. Credits never expire.
Chain it on one canvas
A generated image becomes the next step's reference; a clip can be trimmed, extended and scored right there — no export/import round trips.
Drive it from MCP or the CLI
Generate straight from Claude or Cursor, or wire the CLI into your own scripts and automation. The browser is optional.
FAQ
How long can a MiniMax H3 clip be?
Any whole second between 5 and 15 — not a fixed set of presets. Output resolution is 2K.
What does "multimodal reference" actually mean?
One prompt can carry image, video and audio references at once rather than a single still — video sets motion and pacing, audio sets tone.
Does H3 support text-only generation?
Yes. Text-to-video, image-to-video and reference-to-video all work, and you can pin an end frame.
Anything to know when using reference images or clips?
H3's upstream moderates the reference media you send, and non-compliant references are rejected with credits returned automatically. If references keep getting rejected, switch to text-only generation or to another video model on the canvas.
Try it now
New accounts start with free credits — no card up front. Image, video and audio share one balance, and credits never expire.