MiniMax H3 in ComfyUI: setup for 8GB VRAM

This tutorial covers a local AI text-to-video workflow for MiniMax H3 in ComfyUI. The presenter tests an RTX 5060 Laptop GPU with 8GB VRAM and 32GB system RAM, which supports model offloading. Their setup uses Python 3.12, CUDA 13.2 and ComfyUI Nightly. They explicitly leave support in the stable build unconfirmed.

The walkthrough explains model downloads, folder placement and loader selection. The presenter recommends FP8 diffusion weights for Ada Lovelace or Blackwell GPUs and suggests INT8 for older hardware. For the text encoder, they recommend NVFP4 only on Blackwell and INT4 otherwise. Their sampling settings use res_multistep, the simple scheduler and 20 steps.

Memory limits determine clip length and resolution. On this system, 15-second attempts at 0.6 and 0.5 megapixels failed with out-of-memory errors; 0.4 megapixels worked at 864 × 480. The presenter also reports roughly 540p for 10 seconds and 720p for five seconds. These are results from their hardware, not universal limits.

The presenter finds H3 slower than LTX 2.3 but more consistent in producing usable results. Optional interpolation and RTX Super Resolution follow generation. They warn that mismatched frame rates can disrupt timing, and that distant faces can distort at low resolutions; closer framing helped in their tests.