MiniMax H3 in ComfyUI: video extensions and SLA tests

This ComfyUI tutorial builds longer MiniMax H3 videos by joining separately prompted segments. The speaker demonstrates a 33-second sequence using the int8 first-last model, then reuses a frame from the opening segment to help restore character details after a close-up. Separate previews let users work on each segment before combining the result.

Further examples extend an existing video, combine two video references for character appearance and voices, and pair a masked face image with an audio clip. These are demonstrations of generated likeness and voice consistency, rather than guarantees for every input. The workflow also uses camera prompts to generate imagined views beyond a still image.

For local AI performance, the speaker compares Lightx2v SLA and non-SLA LoRAs with Sparse Linear Attention enabled or bypassed. On the speaker's Nvidia 3090, enabling the SLA node produced about a 25% speed improvement in the shown comparison. The speaker attributes the gain to the attention node rather than the LoRA alone. A separate PDD acceleration comparison took the same time at eight steps and produced an audio hiccup.

The technical walkthrough covers reference conditioning and extension controls. It specifies a 124-frame input limit for existing videos, warns that results may degrade above roughly 160,000 tokens, and notes that the PDD node permits four, six, or eight steps.