ComfyUI INT8 setup: Boogu, Krea 2 and Flux Klein

This tutorial covers INT8 image generation in ComfyUI with Boogu, Flux Klein, Krea 2 and Z-Image Turbo. The speaker explains how INT8 integer quantization differs from FP8 floating point and stresses that speed gains depend on the model and GPU. These are local AI workflow demonstrations, with timings from the speaker's 24 GB VRAM card rather than general hardware benchmarks.

Setup includes updating ComfyUI and Pixaroma nodes, placing diffusion models, text encoders and VAEs in their folders, and refreshing node definitions. The speaker reports using ComfyUI 0.27 and recommends that version or newer. Cache troubleshooting covers Ctrl+Shift+R and disabling the browser cache through developer tools when node buttons are missing.

Boogu's turbo example uses four steps and CFG 1, without negative prompts. Its first run takes 12 seconds and the repeat takes 2 seconds. Oversized outputs can duplicate subjects. In editing examples, the speaker finds Boogu preserves placement but struggles with pose changes, while Flux Klein handles those changes better.

The tutorial also covers masked inpainting, prompt enhancement and Krea 2 image-to-prompt workflows. Pixaroma's Save Image node supports custom folders and filenames; PNG can retain workflow metadata, while JPG cannot. The closing comparison favors Flux Klein for speed and Krea 2 for image quality in the speaker's tests.