MiniMax H3 Ref2V
Prompt Guide:
Direct the Clip.

Fill the six sections in order — subject definitions, a one-line summary, a retention note per reference, a detailed description of shots and actions, an overall soundscape, and an optional music cue. Give every attached file an explicit job and label (Image 1 the face, Video 1 the movement, Audio 1 the voice), name the camera as a plain verb with its speed, and quote any spoken line inside a tagged span with its language. Plain direction outperforms keyword lists.

References are ordered, and their position in the upload is their label in the prompt. Images carry identity, wardrobe, and style; the first video carries the movement and camera feel you want transferred; standalone audio carries voice timbre. Up to 9 images, 3 video clips, and 3 audio files are allowed with a 12-file ceiling, and at least one visual must anchor the render — audio alone cannot be the only input.

Up to 9 reference images, 3 reference videos, and 3 standalone audio clips, with 12 files total as the hard cap. Each video and audio clip runs 2-15 seconds and the combined reference-video seconds stay under 15. A standalone audio file must accompany at least one image or video so the model always has a visual first frame.

Yes — write the spoken line the way a script reads, tag it with the language, and H3 lip-syncs it into the picture with synchronized stereo ambience in the same render pass, across the model's supported dialogue languages. Keep the clip short enough that the line lands, and prefer the voice-transfer route when you already have a take you want kept.

Upload the exact movement as Video 1 — a walk, a steadicam push, a product turn — trimmed to the seconds where the move is isolated, then confirm it in the brief with a sentence like a named push in matching Video 1. The clip demonstrates amplitude and speed text cannot, so showing it beats describing it. When the scene content changed a lot, pair the clip reference with a plain-language camera verb so the transfer is not read loosely.

In Moosky credits, at 100 to the dollar, scaling with clip length and resolution: about 10 credits per output second at 768P and about 18 at 2K, plus every second of your reference video billed at the output resolution you pick and any images beyond the free first five. The composer shows the exact total for your bundle before anything renders, and failed generations refund automatically — there is no subscription.

Start at 768P for drafts — it is H3's native short-edge around 768 and cheap to iterate, so you can rough a whole sequence and check that the motion and identity transfers read. Escalate only the take that works to 2K, whose short edge is 1440, for the delivery file. That ladder is the cheapest way to converge on a final without paying full rate on every experiment.

Yes — attach the approved clip as Video 1 and prompt only the change while naming what to preserve, so the performance and transferred movement carry and you pay for output seconds once instead of rebuilding from the reference sheet. This is generative editing, not a video editor: it cannot cut, trim, or speed-ramp, so restart from the sheet when the scene concept itself was wrong.

Yes. Every recipe renders on the live MiniMax H3 Ref2V composer on Moosky, which is the same full-reference engine with native stereo audio and the ordered Image, Video, Audio, and Subject labels the six-section brief cites. Paste a recipe into the composer, adapt its wording to your files, and see the credit price before you run it.

This page is the tool lane: a commercial prompt recipe library that sells the reference-to-video composer and deep-links you into it so you can run the six-section brief and watch the cost first. The sibling workflow guide keeps the informational how-to — building the reference sheet, tuning the transfer, running delta edits step by step. Read the guide to learn the method, then load your kit and generate from here.