Short-form video is unforgiving: a couple of seconds to hook someone, a vertical frame only, and no room for wasted motion. We wanted to know whether Minimax H3 Max — hiapi's text-to-video model — could go straight from a prompt to a vertical clip. We ran it for real through hiapi's /v1/tasks API, and the answer had a catch worth knowing before you spend render budget on it.
What We Actually Generated
We sent this exact prompt to minimax-h3-max in text-to-video mode (reused nowhere else — write your own for your own footage):
A florist's hands wrap a bouquet of ranunculus and eucalyptus in brown kraft
paper on a sunlit marble counter, then tie it with twine in a neat bow; soft
morning window light, shallow depth of field, close-up vertical framing,
warm cozy mood, gentle paper-rustling motion
Request body:
{
"model": "minimax-h3-max",
"input": {
"prompt": "...",
"duration": 6,
"resolution": "480P",
"aspect_ratio": "9:16"
}
}
We asked for 9:16. Here's the real output the API returned, byte-for-byte, no edits:
Notice the frame: it came back landscape, not the vertical clip we requested.
The Catch: Text-to-Video Always Renders Landscape
This isn't a fluke on our end — hiapi's own model description for Minimax H3 Max says it directly: "Generate 5-15 second videos at 480P or 768P from text or first/last-frame images. Text-only generation uses the default landscape canvas." We tested it to confirm: send aspect_ratio: "9:16" with a pure text prompt, and the model silently ignores that field and renders landscape anyway.
image-to-video is listed alongside text-to-video in the model's task capabilities, but there's no frame-orientation steering happening in pure text mode. If you need a different canvas from a text prompt, you fix it after the render — not in the request.
For anyone building a short-form pipeline on this model, that's the one thing worth knowing before you burn a render: don't set aspect_ratio in text-to-video mode expecting vertical output. Plan for a post-processing step instead.
Turning It Vertical With One ffmpeg Command
Cropping the landscape frame to 9:16 would cut off half the bouquet. Instead we used a blurred-background pillarbox: scale a heavily blurred copy of the source to fill the vertical canvas, then overlay the original frame centered on top.
ffmpeg -y -i video.mp4 -filter_complex "\
[0:v]scale=720:1280:force_original_aspect_ratio=increase,crop=720:1280,gblur=sigma=20[bg];\
[0:v]scale=720:-1[fg];\
[bg][fg]overlay=(W-w)/2:(H-h)/2[out]" \
-map "[out]" -c:v libx264 -pix_fmt yuv420p -movflags +faststart video_vertical.mp4
One pass, no re-render, no lost frame content. Here's the result — same clip, now a real 720x1280 vertical file:
That's the version you'd actually publish to Reels, Shorts, or TikTok.
Pricing: What a Clip Like This Actually Costs
Minimax H3 Max bills per output second, tiered by resolution (confirmed on hiapi's pricing page):
- 480P — $0.072 / second
- 768P — $0.115 / second
Duration runs 5-15 seconds. Our 6-second 480P test clip cost $0.432; the separate 16:9 cover image for this post (a vertical clip can't serve as a site cover) added $0.008. Total: $0.44 for one publish-ready short.
At that price, 480P is cheap enough to iterate on hooks and framing before committing to a 768P final render — you can run four or five 480P drafts for less than the cost of one 768P clip.
FAQ
Can Minimax H3 Max generate vertical video directly?
Not from a text prompt. Text-to-video always renders landscape regardless of the aspect_ratio value you send. You need a post-processing step (a crop or, better, a blur-pad overlay) to get 9:16.
How much does a 6-second clip cost? $0.432 at 480P ($0.072/sec x 6s), or $0.69 at 768P ($0.115/sec x 6s).
What's the difference between 480P and 768P beyond price? 768P is the higher-fidelity tier at roughly 1.6x the per-second cost of 480P — worth it for a final publish render, overkill for testing prompts and framing.
Does image-to-video avoid this orientation issue? It's listed in the model's task capabilities, but we didn't test first/last-frame image inputs for this piece — our finding here is specific to pure text-to-video mode.
Try It Yourself
The Minimax H3 Max model page has the live Playground if you want to test a prompt before writing any code, and our Minimax H3 Max API guide covers authentication and the full request/response shape end to end.
