MiniMax H3: ComfyUI speed comparison with SageAttention

This tutorial compares a standard MiniMax H3 ComfyUI workflow with a version using a SageAttention KJ node. The presenter runs both workflows in the cloud and reports generation times of 21 minutes and 10 minutes. The stated setup uses an RTX 4090, with both workflows set to produce eight seconds of video at one megapixel and 25 steps.

The presenter acknowledges using different prompts for the two generations. They report no noticeable visual quality difference, but this comparison does not establish identical quality under matched inputs. The timing result is a reported outcome from this demonstration, rather than a general performance guarantee.

The workflow edit places the SageAttention KJ node between the diffusion model and the basic guider, with attention set to auto. The presenter credits Michael with discovering the change and says they will update the downloadable cloud workflow to enable it.

For local AI users, the main limitation is installation coverage: the video does not explain how to install SageAttention on a personal computer. The presenter had not yet installed it locally, so their expectations for RTX 40 and 50 series cards are predictions rather than demonstrated local results.