MiniMax H3 community optimizes with turbo LoRAs and sparse attention

Community members share workflows using turbo LoRAs (e.g., Larry's Turbo, lightx2v) to cut steps from 20 to 8, enabling 27-second clips in ~9 minutes on an RTX 5090. One dev reports native 720p→1440p second sampling on a single RTX 4090 at 112s/223s/334s with auto-scheduled sparse attention.