Your first local images with ComfyUI

Image subscription credits are the clearest rental trap in AI: pay monthly, wait in queues, and lose access to your own style history. Local diffusion flips all three.

Pick your engine first

• SDXL (7GB VRAM): the ecosystem king. Thousands of community fine-tunes and LoRAs for every style.
• FLUX.1-schnell (12GB VRAM, Apache-2.0): 4-step generations, best-in-class prompt adherence and text rendering.

On a 12GB card run both; on 8GB start with SDXL; on CPU, honestly, wait for your GPU upgrade.

Install ComfyUI

ComfyUI models generation as a node graph. Intimidating for one evening, irreplaceable by the second week.

git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
pip install -r requirements.txt
python main.py

(Windows users: the portable zip from the repo’s releases is even easier.)

Load a model

Drop the model file into models/checkpoints/:

• SDXL: sd_xl_base_1.0.safetensors
• FLUX schnell: grab the ComfyUI-ready GGUF/NF4 variant for lower VRAM

Open http://127.0.0.1:8188 and load the default workflow (or the FLUX template from the template library).

Your first prompt

photograph of a mechanical hummingbird, brass and glass, studio lighting, shallow depth of field

Settings that rarely disappoint:

Setting	SDXL	FLUX schnell
Steps	25–30	4
CFG	6–8	1.0 (guidance 3.5)
Sampler	dpmpp_2m + karras	euler + simple

The local advantages nobody advertises

• Unlimited iterations. Seed-hunt all night; marginal cost is electricity.
• LoRA stacking. Combine style adapters like layers of flavor.
• Total privacy. Your prompts and outputs never touch a server.
• Upscale pipelines. 2-pass hires fixes and tiled upscalers run while you sleep.

Where to next

• FLUX.1-schnell and SDXL in the model directory
• VRAM calculator to check what fits your card