Quick answer: the H3 video atlas is ten rendered masters, each with the exact request behind it, fifty prompts in total, measured file metadata and dated run logs, published under CC BY 4.0. It is a dataset for checking behaviour, not a gallery, a benchmark or a quality ranking.
Every model has a demo reel. Reels are curated to look good, which makes them useless for the question people actually have: what does this thing output when I ask for something ordinary?
We published the opposite artefact for H3 — a small dataset where every clip ships with the request that produced it and the metadata of the file that came back. It is deliberately unglamorous, and it is the most useful thing we have published about this model.
What is actually in it
| Component | Contents |
|---|---|
| Masters | 10 rendered clips, published at full quality |
| Run logs | one generation.json per sample: prompt, tier, aspect ratio, duration, seed where used |
| Measured metadata | pixels, frame rate, real duration, audio track, cost per sample |
| Prompts | 50 in total, including the ten that were rendered |
| Method | how each sample was recorded, so you can repeat it elsewhere |
| Licence | CC BY 4.0 |
The community section is a separate discipline: fourteen cases are reproduced verbatim, with credit, and we do not re-host anyone's footage. A prompt reconstructed by watching a video describes the output, not the prompt, and passing that off as someone's work is a habit we wanted to avoid.
The findings that justify the format
Three numbers from this dataset contradict what a spec sheet would lead you to expect:
- "1080P" returns 1920×1056, not 1920×1080 — a pixel-area cap, consistent across ten samples and three aspect ratios.
- 768P's native canvas is 1344×768, not 1366×768 and not 1280×720.
- Durations quantise to 24 fps frames, so a five-second request renders about 5.17 seconds and a six-second request about 6.58 — a 1.6× overshoot at the one-second end.
None of those are visible in a gallery of pretty clips. All of them change what you build: overlay templates authored at 1920×1080, NLE presets that assume multiples of 16, batch concatenation, per-second budgeting.
How to reproduce it against another provider
The method is the portable part, and it fits in four steps:
- Pick three tiers and three aspect ratios; generate at least ten samples total, with fixed prompts.
- Record the request exactly as sent, including the tier and any reference inputs.
- Read the returned file with a probe rather than trusting the response body — container metadata for pixels, frame rate, duration, and the audio stream list.
- Publish the pairs. A number without its request is an anecdote.
Two details people skip and then argue about: record wall-clock time from submit to playable file (queue time is invisible in the response), and note the audio sample rate, because a text-to-video job with no audio input still came back with a 32 kHz track on every sample here.
What the dataset is not
- Not a quality comparison. Nothing here ranks models by how good they look; that is a judgement call and it is yours.
- Not a benchmark. Ten samples is enough to establish pixel and timing behaviour, not enough to state failure rates.
- Not a promise of stability. Providers change lanes. If the numbers move, the dated run logs are the reason the change is detectable.
- Not a licence to reuse the footage as stock. CC BY 4.0 covers the artefacts; it does not grant rights to whatever a prompt depicts.
Why publish something this unflattering
Because the alternative is worse. A blog that only shows good outputs produces readers who bounce off the first 1056-pixel surprise, and a support queue full of "why is my video the wrong size" — we wrote about that surprise for exactly this reason.
The practical effect of publishing the awkward numbers is that the corrections come from readers instead of from tickets. Two of the comments on our last post were people reporting the same behaviour on other providers' lanes, which is more useful than agreement.
FAQ
Can I reuse the footage commercially? CC BY 4.0 covers the artefacts. It does not grant rights to whatever a prompt depicts.
Is this a benchmark? No. Ten samples establish pixel and timing behaviour, not failure rates.
How do I reproduce it against another provider? Fix three tiers and three aspect ratios, record the exact request, and probe the returned file instead of trusting the response body.
Where this runs
The atlas is published under CC BY 4.0. https://callirra.com/minimax-h3 plays lighter transcoded copies of the ten samples and lists their prompts, measured metadata and the current tier table. The full masters, fifty-prompt set and per-sample measurement logs were hosted in the atlas's GitHub repository, which is temporarily unavailable; those complete downloads are not currently served by the preview page.
I work on Callirra, so this is our dataset about our own lanes — which is a bias worth naming. If you run the same method against another provider and get different numbers, open an issue with the request JSON and the file's metadata. Contradictions are more valuable than confirmations, and they are the only thing that keeps a dataset honest.