FLUX 3 Klein: ComfyUI setup and base vs distilled

This walkthrough covers FLUX 3 Klein in ComfyUI, with GGUF and W4A8 downloads for users who want to run models locally. The speaker describes a 5.6-billion-parameter open-weights model for image generation and editing that accepts up to 10 reference images and produces one output. It uses Qwen 3 VL 4B Instruct as its text encoder.

The setup uses supplied workflows and model dropdowns. GGUF users must swap the loader used in the W4A8 workflow. The speaker gives download sizes of roughly 5 to 11 GB for the native variants and 3.3 to 6 GB for the smaller options; these are file sizes, not measured VRAM requirements. A prompt enhancer remains disabled after an illegal memory access error.

The comparison uses the base model at 25 and 50 steps alongside the distilled variant. The stated workflow settings are CFG 3.5 for base and CFG 1 for distilled, with up to 16 steps for the latter. The speaker generally prefers the detail at 50 base steps, though some simpler scenes look similar and one fish-seller example favors distilled.

Editing examples use base at 25 steps for a handle color change, a prompted anime conversion and a two-reference hat edit. The speaker reports good detail retention in the color and hat edits, but notes a likeness change in the anime result. These are observations from the examples, rather than a general quality benchmark.