MiniMax H3 Spicy: What You Actually Control in NSFW Video
The first time you send a request to a MiniMax H3 Spicy endpoint, the body looks too short. Four or five fields, no seed, no negative prompt, nothing that resembles the long parameter lists on comparable video models. Nothing is missing. This family gives you a small number of knobs on purpose, and knowing exactly which ones you have changes how much work the prompt has to do.
Key Takeaways
- t2v: aspect_ratio, duration, prompt, size — all required.
- i2v: five fields, and no aspect_ratio — the input image sets the frame.
- ref2v: adds images, videos and audios as optional arrays, capped at nine each.
- duration is every integer from 5 to 15, not an even-step enum.
- size offers 768p and 2k only. No seed and no negative_prompt anywhere in the family.
The whole control surface, three endpoints
Endpoint | Fields |
minimax-h3-t2v-spicy | aspect_ratio (21:9 / 16:9 / 4:3 / 1:1 / 3:4 / 9:16) · duration · prompt · size |
minimax-h3-i2v-spicy | duration · image · prompt · size |
minimax-h3-ref2v-spicy | everything in t2v, plus images ≤9 · videos ≤9 · audios ≤9 |
Six aspect ratios is a wide spread for a video family, running from cinema-wide 21:9 through square to full vertical 9:16, which covers social delivery without a crop step.
Why i2v has one field fewer
The missing aspect_ratio on image-to-video is an inference, not an omission. The input image already fixes the frame. Accepting a ratio field alongside it would create a request that can contradict itself, so the endpoint does not offer one.
The practical consequence: on i2v, framing decisions move upstream into whatever produced the still. Generate the source image at the ratio the final clip needs, because there is no parameter to correct it afterwards.
Duration is every integer from 5 to 15
Several video families restrict duration to an even-step enum — LTX 2.3, for instance, runs 6 to 20 in twos. H3 Spicy accepts each whole second across its range, so 7-second and 11-second clips are legal requests rather than rounding targets.
That matters when clips are being cut together. A sequence timed to a specific beat can be generated at its real length instead of generated long and trimmed, which removes a step and avoids paying for frames that get discarded. The 5-second floor and what it costs covers the lower bound and its billing in detail.
ref2v: nine of each
The reference endpoint takes three optional arrays, each capped at nine entries: images, videos and audios. Against the neighbours, that symmetry is unusual.
Model | Images | Videos | Audios |
MiniMax H3 Spicy ref2v | 9 | 9 | 9 |
Wan 3.0 Prime Spicy ref2v | 10 | 5 | 5 |
Seedance 2.5 Spicy ref2v | 30 | 10 | 10 |
Seedance takes far more still references; Wan weights its budget toward images too. H3 is the only one of the three that treats all three reference types as equally important. For a scene assembled from multiple sources — character stills, a motion reference, an audio reference — the H3 budget does not force you to spend it mostly on images. The ref2v reference workflow covers how those inputs combine, and H3's native audio covers the audio side of this family.
No seed, no negative prompt
Both absences change working habits.
No seed means no reproducibility. A generation that lands well cannot be recovered by re-sending the same request. Generate in batches and keep what works, rather than planning to iterate on one result.
No negative prompt means exclusions have to be written as inclusions. Instead of listing what should stay out of frame, describe the frame completely enough that the unwanted element has no room. "Bare shoulders against a plain dark wall" does the job that a negative prompt listing background clutter would have done elsewhere.
These endpoints are available through Siray.ai alongside the rest of the NSFW lineup, on one API key, so comparing control surfaces across families is a model-string change.
Create your free Siray account and start generating with MiniMax H3 Spicy now.
A Note on Compliance
Siray's NSFW models generate legal, consensual adult content only — CSAM and any illegal content are refused with zero tolerance. Uncensored means no excess moderation within the law, not the absence of law. Every example here assumes adult subjects, and any reference image, video or audio is assumed to be used with the consent of the people in it. MiniMax H3 models are third-party models available through Siray's API, not Siray-hosted models.