ComfyUI MiniMax H3 Speed LoRA and VRAM setup

This tutorial configures MiniMax H3 video workflows in ComfyUI with a Speed LoRA that reduces generation from 20 steps to four or eight. The presenter places its loader between the diffusion model and sampler, then uses a shift value of six rather than the stated default of 12. The setup starts with updates to ComfyUI and Pixaroma Nodes; detailed model installation belongs to episode 29.

In the presenter's tests, Euler with the simple scheduler worked best for text to video and first-plus-last-frame generation. Euler with beta scored best for first-frame generation. Eight steps gave a better balance of speed and generated audio quality than four. The presenter recommends 20 steps without the LoRA when quality matters most. These are results from the demonstrated workflows, not universal performance guarantees.

Pixaroma's monitor displays VRAM, RAM and temperature. A Free VRAM node clears memory after VAE decoding to address cache buildup, but forces models to reload on later runs. One example peaked around 22 GB of VRAM; that is an observation, not a minimum requirement.

Singing and speaking workflows reuse uploaded audio, so the presenter uses four steps. The reference model sometimes changes character details at larger sizes. Gemini supplies detailed prompts, while resolution and duration affect generation time. The examples give local AI users settings to test against their own hardware.