Draw Things: MiniMax H3 video workflows and settings

This tutorial covers MiniMax H3 in Draw Things, which the speaker describes as a free, open source image and video generator for Mac, iPad and iOS. H3 and the separate reference model share a Qwen encoder. Together, the speaker says, they occupy about 100 GB, with extra disk space needed during installation.

For first/last-frame video, the starting image goes on the canvas and the ending image on the mood board. Both should match the render aspect ratio to avoid distortion. The prompt needs to describe the camera and character movements between them; the speaker reports quick fades or wipes when that direction is missing. Switching from edit to generate enables text-to-video, while a single-frame render produces an image.

The settings discussion covers guidance at 1.0, no negative prompting and a recommendation of 50 steps. The speaker reports clean results at 18 to 22 steps without an acceleration LoRA, while warning that acceleration can reduce quality and variation. Output runs at a fixed 24 fps.

The reference workflow uses numbered image references in prompts. At recording time, Draw Things accepted reference images but lacked audio or video inputs and checkpoint model imports. The speaker also reports trouble animating at the claimed 2K resolution and recommends 256-by-256 tile decoding for large still renders. The closing section covers Storyflow frame calculations and mood board controls.