MiniMax H3 in ComfyUI: Turbo and video extension tutorial

This tutorial covers MiniMax H3 video workflows in ComfyUI, with Turbo LoRAs for the first/last and reference models. The speaker discusses four- and eight-step variants and warns that faster sampling can affect generated audio. For lip-sync work with supplied audio, the workflow can send the original recording directly to the video combine node.

The local AI workflow includes Sage Attention patches, Comfy Kitchen Attention, low-VRAM attention and chunked feed-forward options. The speaker presents these as ways to reduce memory use or improve speed, without giving benchmark results. Model Sampling MiniMax H3 controls shift values rather than reducing memory use or generation time.

Custom keyframes place reference images at chosen positions, including the middle of a clip. The examples also extend generated footage and an existing video using the first/last model. The speaker describes an approximately 15-second generation limit and demonstrates successive extensions of five and seven seconds.

The technical walkthrough covers shared sampler settings, input-image masks and separate conditioning for each extension prompt. It recommends a motion context length of 39, then trims the context frames before assembly. Overlap blending joins the image batches, while a separate assembly node handles concatenated audio. Trimming makes the final result slightly shorter.