MiniMax H3 in ComfyUI: four-step video setup

This tutorial builds a MiniMax H3 workflow in ComfyUI for generating video and sound together from text or a single image. The presenter combines Light X2V Turbo LoRA for four-step sampling, a sparse attention speed patch and an 11 MB Semantic Bridge adapter intended to improve prompt following.

The download walkthrough covers a roughly 21 GB pruned INT8 model, a roughly 16 GB Qwen 3 VL text encoder, audio and video VAE files, and custom nodes. The Semantic Bridge needs manual installation. A smaller W6A8 model saves about 5 GB; the presenter reports similar speed and quality in his tests.

The local AI setup uses PyTorch built for CUDA 13 and requires NVIDIA driver 580 or higher. The presenter reports that switching from CUDA 12.8 reduced one sampling run from four and a half minutes to 1 minute 37 seconds. Other five-second examples sampled in about a minute on his test GPU.

Memory tests peaked around 13 GB. Text-to-video kept full speed under a 12 GB limit, while an 8 GB limit caused substantial slowdown. The Windows portable setup section adds --vram-headroom 1 to the launch command to reserve 1 GB.

The examples use ERSD with SGM uniform. For image-to-video, the presenter recommends a full-size source with a matching aspect ratio and 0.6 megapixels on a 12 GB card. His 1-megapixel example peaked just above 12 GB and sampled in about two minutes.