Hi MiniMax-H3 maintainers — thank you for releasing H3 and the public prompt-writing guidance.
I maintain Omni-Rewriter, a small Apache-2.0 community project exploring a typed prompt-expansion workflow for image and video generators. We added an H3-oriented profile based only on public materials:
user intent → typed draft → deterministic validation → bounded repair → H3-oriented text
The current H3 path covers the public T2VA / I2VA / L2VA / FL2VA / Ref2VA routing, timestamped shot grammar, dialogue tags, media-role checks, and explicit first/last-frame constraints. Prompt expansion is kept separate from generation. We do not claim to reproduce or approximate the private H3-Context-IR implementation, and we do not present the project as an official MiniMax SDK.
We have also exercised optional H3-Base generation through a local SGLang deployment and published same-seed RAW vs expanded examples. These are community examples rather than a quality leaderboard or a universal improvement claim.
Links:
Would this kind of community prompt-validation tool be useful in a community resources or prompting section? If so, I would be glad to open a very small README PR in whatever location and wording you prefer. If it is outside the repository's scope, no worries at all — feedback on any incorrect H3 assumptions would still be very helpful.
Thank you for your time and for making the model and guidance available.
Hi MiniMax-H3 maintainers — thank you for releasing H3 and the public prompt-writing guidance.
I maintain Omni-Rewriter, a small Apache-2.0 community project exploring a typed prompt-expansion workflow for image and video generators. We added an H3-oriented profile based only on public materials:
user intent → typed draft → deterministic validation → bounded repair → H3-oriented textThe current H3 path covers the public T2VA / I2VA / L2VA / FL2VA / Ref2VA routing, timestamped shot grammar, dialogue tags, media-role checks, and explicit first/last-frame constraints. Prompt expansion is kept separate from generation. We do not claim to reproduce or approximate the private H3-Context-IR implementation, and we do not present the project as an official MiniMax SDK.
We have also exercised optional H3-Base generation through a local SGLang deployment and published same-seed RAW vs expanded examples. These are community examples rather than a quality leaderboard or a universal improvement claim.
Links:
Would this kind of community prompt-validation tool be useful in a community resources or prompting section? If so, I would be glad to open a very small README PR in whatever location and wording you prefer. If it is outside the repository's scope, no worries at all — feedback on any incorrect H3 assumptions would still be very helpful.
Thank you for your time and for making the model and guidance available.