Real timing methodology
Generation time is measured from the stored task creation time to its terminal update. It includes provider queue and processing, so it is not presented as pure inference time.
- Use real task records
- Do not publish private prompts
- Separate queue time from inference time
- Prefer a short test before a long render
A practical workflow
A fair speed record starts when the task is stored and stops when a success or failure is stored. Provider load, network calls, safety checks, and output persistence may all contribute. Compare groups of similar jobs rather than treating one fast result as a promise.
- Anonymous records only
- No prompts or user identifiers
- Total time clearly labeled
- Empty state instead of invented samples
Frequently asked questions
Are these benchmark results?
No. This page is a feature and workflow comparison unless a measurement is explicitly labeled as a real task record.
Does H3 Max generate audio?
Yes. The H3 Max workflow includes synchronized native audio. Describe ambience, dialogue, sound effects, or silence directly in the prompt.
What durations and resolutions are available?
This generator offers 5, 10, or 15 seconds at 480P or 768P for H3 Max. MiniMax H3 remains available separately for 2K final output.
How should I start a new idea?
Begin with a five-second draft, compare a few variations, and only move the strongest direction to a higher resolution or 2K finish.