Gothic horror teaser prompt for MiniMax H3

This template is here because of a number rather than a preference. We pulled every official MiniMax H3 showcase clip that ships under a licence we can quote, seven in all, and measured the audio instead of describing it: RMS level in half-second windows, 2,340 windows across the set, encoder priming discarded. The result was lopsided enough to be worth publishing. Six of the seven clips never go quiet. The fashion piece, which explicitly asks for VHS glitches and CCTV broadcast interruption, spends its entire fifteen seconds inside a 16.8 dB band and bottoms out at -25.5 dB. The starship, the ramen shot, the crime title sequence, the portal, the voice-edit demo — all of them keep a floor.

This clip is the exception. It opens on roughly 1.6 seconds below -40 dB, then dips again at 8.5, 10.0 and 13.9 seconds. That is not a flat track with a quiet intro; that is pacing. And it is the only clip in the set that reads as horror.

The interesting part is what produced it. We went back through the prompt looking for the audio instruction and there is not one. In 1,366 characters there is no mention of sound, music, score, ambience, silence or quiet. What the prompt does carry is an emotional register stated four different ways — dark romance, gloomy oppression, high-end and restrained, the hook of a short drama — and a list of things it refuses to be: no blood, no cheap horror, no Halloween feel, no modern street feel.

The contrast case is the official crime title sequence, which also refuses horror but does it differently. It writes "not horror, not heavy" and in doing so negates the temperature along with the costume. Its floor is -27.9 dB and it never stops. So the working read — and it is a read, from one positive sample, not a settled result — is that naming a restrained register buys dynamics, while negating the whole mood spends them. That is a testable claim and it is worth testing before you rely on it: same scene, one take that names the register, one that does not.

What is not in doubt is the default. Six clips out of seven, from three different generation modes, with prompts written by different people for completely different briefs, all come back with a soundtrack that never lets go. If you are making anything where a pause is part of the effect — horror, suspense, a product reveal, a punchline — you are working against that default and you should plan for it rather than discover it.

The rest of the prompt is a clean demonstration of what Ref2VA is for. Two images with separate jobs, five setups, two identities held across all of them, and almost no physical description of either lead. "Innocent human female lead" and "vampire noble male lead" is the entire character brief; the faces come from Image 1 and the castle from Image 2. On a vertical frame the shot-scale list does the rest of the work, and it does it in the direction the model is strongest: close. Every diffusion video model degrades at distance, so a brief that lives in medium close-up and tighter is spending its quality budget where the model can actually pay.