Download
1 variant available
1760 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
(10)
Aug 13, 2026
MiniMax H3
Initial public release.
- Rebuilt AllModes around one authoritative MODE + duration controller.
- Fixed the mismatched 10-second preset: 1152×640, 243 frames, 24 fps.
- Synchronized FIRST/LAST preprocessing with the resolved canvas.
- Replaced saved local image paths with portable optional upload nodes.
- Hardened image loading so validation and decode use one verified immutable snapshot.
- Repaired Ref2VA's duration mapping and duplicate subgraph backlink.
- Added pinned model receipts, dependency revisions, licenses, checksums and a static security review.
No model weights or generated H3 outputs are included. Final GPU-render verification after the portability rewrite has not been run.
Show more
70 1 2 3 4 5 6 7 8 9
230 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
> LICENSE / AUTHORIZATION NOTICE
> The publisher has received written authorization from MiniMax to use MiniMax H3 and MiniMax H3 Works. The private authorization is not bundled and does not transfer to downloaders. No model weights or generated outputs are included. Users must review the upstream license and obtain any authorization required for their own location and use case.
Two organized MiniMax H3 ComfyUI workflows are included in one archive.
### Public-friendly controls
- Direct generation width × height inputs — no megapixel conversion
- Requested-duration or manual exact H3 frame mode
- Native H3 video/audio visibly locked to 24 fps for A/V sync
- Seed in the main user panel
- Sage and Sol attention can be switched independently
- Choose one step path: Spectrum, EasyCache Fast, EasyCache Draft, or Off
- Native attention + full-step evaluation is the public default: Sage OFF, Sol OFF, step acceleration OFF
### AllModes
- T2V, I2V, L2V and FLF2V in one workflow
- MODE is the only authority for FIRST/LAST image requirements
- Portable image inputs with no saved local filenames
### Ref2VA
- One reference_count activates a contiguous REF1…REFN set, up to nine images
- @REF1…@REF9 aliases are validated against the selected count
- Direct WxH, duration/frames, seed, and match/max reference sizing
Friendly starting canvases: 864×480, 1024×576, 1152×640, and the H3 Base reference 1344×768. Larger aligned values such as 1920×1088 are experimental high-VRAM inputs, not Regenerate-2K.
H3 uses a 17k+5 frame grid. For example, requested 10.0 seconds resolves to 243 frames = 10.125 seconds of native H3 A/V at 24 fps.
The slim archive contains eight files: both JSON workflows, the required h3_personal_controls source, one consolidated README, and the four required license/notice files. Model receipts and dependency revisions are consolidated into the README. Third-party node source and model weights are not included.
Static validation and actual ComfyUI offline prompt-schema validation passed. No /prompt submission or GPU render was performed for the final v2 build. Sage, Sol, Spectrum and EasyCache are optional experimental/performance paths.