MiniMax FastH3 in ComfyUI: three video workflows

This tutorial covers three MiniMax FastH3 workflows in ComfyUI: text to video, first frame to video, and first and last frame to video. The presenter updates ComfyUI and Pixaroma Nodes, then starts ComfyUI without attention enabled at launch because the workflows apply Kitchen Attention and Sparse Attention through nodes. FastH3 uses eight sampling steps. The presenter also notes regional licensing restrictions that may require permission to access the model.

The text workflow exposes orientation, resolution, duration and seed settings. The suggested durations span 4 to 15 seconds; the presenter reports audio problems beyond that range. Google Gemini Gems supply prompt formulas for each workflow, with Claude offered as another option. Examples cover speech, multiple speakers and sound effects.

Tests use a 24 GB VRAM card. The presenter reports that resolution affects generation time more than duration and that the first run takes longer. One small five-second sample takes 44 seconds but looks poor; a larger version takes under two minutes and looks better to the presenter.

The image workflows resize frames to multiples of 32 and can crop the final image to match the first. RunningHub provides a cloud alternative to local AI generation. The presenter tested full HD only with five-second clips, leaving full HD at 15 seconds unconfirmed.