MiniMax H3 Video Generation Cut From 4.5 Minutes to 79 Seconds on RTX 5090
The log shows local AI video generation on one consumer GPU has reached the point where a full music video can be produced in under an hour, with face and lip-sync checks confirming the quantization holds up.
Reporting from 1 source: GameBusiness.jp.
The author spent three weeks speeding up local MiniMax H3 video generation on a single RTX 5090. A 5-second 1280x704 clip dropped from 4.5 minutes to 79 seconds using the official Turbo LoRA, a 4-step FastH3 distillation, and INT8 attention built into ComfyUI. A 28-segment music video that took 3.8 hours now completes in about 40 minutes. Half the gains came from community tools, half from self-discovered fixes, including a tensor reorder required for FastH3 conversion.
The author first built a 46-line profiler to find where time went. Sampling consumed 84 percent of generation time, VAE decode 11 to 17 percent, and video export 2 percent. That breakdown mattered later: cutting attention in half made VAE decode the next bottleneck.
The gains came from two directions. Official and community tools supplied the int8 ConvRot weights, the Turbo LoRA, and a 4-step FastH3 distillation from UC San Diego's Hao AI Lab. The author's own work included measuring quantization error at 0.90 percent, not the 0.17 percent the conversion tool's comments claimed, and finding that FastH3's LoRA needed a tensor swap when converting from diffusers to ComfyUI format.
Quality checks held: face consistency moved from 0.322 to 0.311, with the same-person line at 0.28. One caveat: text-to-video with the same seed no longer produces the same clip, because quantization changes the random number path.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.
Sources
- GameBusiness.jp 4分半かかっていたAI動画が79秒に。MiniMax H3をRTX 5090で3週間かけて速くした全記録（CloseBox）