Summary
This tutorial from the Nerdy Rodent channel marks a practical turning point for ComfyUI users: FP8 models are being superseded by INT8 convolutional-rotation models (“conrot”), which deliver both better image quality and faster generation speeds, including on older hardware like the RTX 3090.
The video demonstrates the quality gap across several leading image models — Crea 2, Quen Image Edit, Z Image, and Bugoo Edit — by running the same prompts through INT8 and FP8 (or BF16) variants side by side. In nearly every case the INT8 version wins on text rendering fidelity and fine detail while generating in roughly half the time. The presenter also introduces the “lenient switch” workflow pattern, a modular ComfyUI design that lets users flip between model types and refiner stages without rebuilding the graph.
Beyond the head-to-head comparisons, the video walks through a practical multi-model pipeline: Quen handles pose-conditioned generation, Z Image and Crea 2 serve as sequential refiners, and Ultimate Upscale or Seed VR handle final resolution boosts. The result is a reusable workflow template covering text-to-image, image editing, and upscaling entirely within INT8 models — a useful reference for anyone optimizing a local ComfyUI setup for speed and quality.
📺 Source: Nerdy Rodent · Published July 18, 2026
🏷️ Format: Tutorial Demo
