Summary
Nerdy Rodent delivers a practical ComfyUI tutorial focused on MiniMax H3’s newer capabilities, walking through several distinct workflows including multi-segment video extension with re-referencing, voice-consistent two-character generation, and exploratory camera movement prompting.
The tutorial demonstrates how to maintain character consistency across extended video sequences — even when the camera zooms in and back out, losing costume detail in intermediate segments — by using a frame from an earlier segment as a reference image for later extensions. Voice consistency across characters is shown working reliably even when input footage is in black and white. A separate section covers the new Sparse Linear Attention (SLA) node from Plague Kind Node, which delivers approximately 25% inference speedup on an Nvidia 3090, with or without the matching SLA LoRA from LiteX2V.
The video also explores PDD acceleration LoRAs and the LiteX2V SLA LoRA, comparing outputs with and without each enabled. Nerdy Rodent uses a modular “rodent method” workflow structure with color-coded groups, making the ComfyUI graphs easier to follow and update. All workflows are built around the int8 version of the MiniMax H3 model using Comfy Kitchen Attention as a drop-in alternative to Sage Attention that requires no additional installation. Useful for anyone building multi-scene AI video productions in ComfyUI.
📺 Source: Nerdy Rodent · Published August 29, 2026
🏷️ Format: Tutorial Demo
