Minimax H3 Full-Featured 4-Step Accelerated LoRA Workflow

For the online version, select the fp16 model. Generating a 15-second 1280×736 video takes 13 minutes and consumes 320 RH coins.

This workflow significantly boosts generation speed via a 4-step LoRA, cutting runtime from 950s to 330s with negligible quality loss (tested on 480P 10s video, RTX 5060 Ti 16GB + 32GB RAM).
4-step LoRA download link: https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora/tree/main
The pruned_int8 version of the base model is incompatible with this workflow and is not recommended.
Use the fl2va version as the base model, which also supports full-reference generation.
The MiniMax H3 Audio Conditioning (T8) node integrates multiple functionalities and can be configured as needed.
Plugin link: https://github.com/T8mars/comfyui-minimax-h3-blockcache-T8
Runtime reference for RTX 5060 Ti 16GB + 32GB RAM:

| megapixels | Duration (seconds) | Runtime (seconds) |

|---|---|---|

| 0.4 |10 | 330 |

| 0.5 | 10 | 448 |

| 0.6 | 10 | 526 |

| 0.7 | 10 | 785 |

| 0.7 | 10 | 1262 |

| 0.8 | 10 | Unable to run |

| 0.4 | 15 | 500 |

It is recommended to generate prompts with the official skills for better output quality.
Official skills download link: https://github.com/MiniMax-AI/MiniMax-H3/tree/main/skills
It is recommended to try the free online version first. If you are satisfied with the results, you may deploy it locally.
(1000 RH coins upon registration; generating a 15-second 720p video costs approximately 320 RH coins)

Free online H3 workflow:

Free online H3 application:

Description