The pipelines are structurally different
| Capability | Local H3-Base | Official full workflow |
|---|---|---|
| Base generation | Local | Hosted or hybrid |
| Context-IR preprocessing | Build your own prompt processing | Hosted module available |
| Native target | 768-pixel short edge | Up to 2K through regeneration |
| Stereo audio | Yes | Yes |
| Data boundary | Can remain local | API request for hosted stages |
What can change in the result
Context-IR is designed to resolve relationships among prompts, images, videos, and audio before H3-Base sees the request. A raw local prompt can be effective, especially when it follows the published prompting guide, but it does not automatically reproduce that hosted interpretation layer.
Regenerate-2K then uses the base video and original context to generate a higher-resolution result. MiniMax says this can recover details that a normal super-resolution system would have to guess, including small text and fine visual information. Because it is generative, evaluate identity, geometry, timing, text, and audio again after regeneration.
When local wins
Local is most compelling for high-volume 768p iteration, workflow research, and environments where an external media-processing call is unacceptable. The cost shifts from per-second billing to hardware, electricity, storage, maintenance, and operator time.
- Many prompt experiments with no per-run API charge.
- Sensitive source media can stay inside the controlled environment.
- Model revision, seed, and runtime can be pinned for reproducibility.
- Developers can change scheduling, placement, attention, and pipeline code.
When the API wins
The API is usually the simpler final-delivery path when 2K is required. It also provides the closest reference for determining whether a local quality difference comes from the open base model, prompt processing, runtime, or a low-bit repack.
- No model download or accelerator fleet to manage.
- Direct access to the official preprocessing and 2K path.
- Fast time to first output and predictable per-generation billing.
- Better fit for bursty use that cannot keep expensive hardware busy.
A hybrid production workflow
- Develop the prompt and reference roles locally at a preview size.
- Confirm motion, composition, dialogue timing, and audio at local 768p.
- Select only the clips worth finishing.
- Use the official regeneration endpoint for selected outputs when policy and budget permit.
- Run a final visual, audio, text, and rights review on the 2K file.