Descriptions:
Nerdy Rodent walks through FreeFuse, a set of custom ComfyUI nodes that solves one of the most persistent challenges in local image generation: combining multiple character LoRAs in a single scene without resorting to manual regional masking. The technique uses adaptive token-level routing to automatically generate per-character attention masks, and the video demonstrates it working across three model families — Z-Image Turbo, Flux, and SDXL.
The tutorial begins with the creator training two original Z-Image LoRAs from scratch — one character using just seven images and another using six — and documents the practical lessons learned around training step counts and learning rates. Findings include that 800 steps at a slightly elevated learning rate was insufficient for a cartoony character style, while 1,500 steps with more descriptive prompting produced better results. These firsthand failure-and-fix narratives give the tutorial genuine practical value beyond what any documentation provides.
The ComfyUI workflow is broken down group by group: the FreeFuse setup node, reconditioning with the “compute token positions” custom node, the “collect mask” sampler that auto-generates the spatial attention masks, and the final “apply mask and sample” step using a standard KSampler. The video also covers model-specific settings (some options are SDXL-only and ignored for Flux/Z-Image), making it easy to adapt the workflow regardless of which base model you’re targeting.
📺 Source: Nerdy Rodent · Published February 14, 2026
🏷️ Format: Tutorial Demo
