A minimal ComfyUI workflow for Qwen-Image 2.1 on a low-end PC with 8GBvram. Ten nodes, no custom packs, no reference images, no ControlNet.

WHAT YOU NEED (4 files)

- models/diffusion_models/ -> Qwen GGUF Q4

- models/text_encoders/ -> Qwen3-VL 8B int8

- models/vae/ -> Qwen Image VAE

- models/loras/ -> lightning 4-step LoRA

The GGUF loader is built into ComfyUI — no extra extension needed.

THE FOUR NUMBERS THAT MATTER

Steps 8 · CFG 2.0 · Lightning LoRA 0.5 · euler / simple

Sizes: 1024x1408 · 1408x1024 · 1024x1024.

Speed mode: 4 steps / CFG 1.0 / LoRA 1.0. Twice as fast, but the negative prompt stops working at CFG 1.

PROMPT STYLE — plain language, not tag lists

"a quiet greenhouse at dawn, morning light through fogged glass, rows of potted plants, a wooden bench, soft volumetric light, detailed, cinematic"

What it is unusually good at: legible text inside the image (signs, labels, posters, English or Chinese), photographic depth (glass, haze, reflections), and spatial logic (behind, in front of, reflected in).

TWO THINGS TO KNOW

- CFG 1.0 silently kills the negative prompt. Not weakened — ignored.

- Raise steps before raising CFG. 20 steps at CFG 3 drifts toward photorealism and takes four times as long.

SPEED ON A LOW-END PC

~50 s at 1024x1408 with 8 steps, ~20 s in 4-step mode. First run after a restart is slower.