Local image generation

What you’ll learn

The main local image models and what each is good for, the tools that run them, how much VRAM you need, and the licensing detail that catches people out.

The models

Local image generation is dominated by a few families:

• SDXL is the workhorse. It runs on 8GB and up, produces native 1024x1024 images, and has by far the deepest library of community fine-tunes and LoRAs, which makes it endlessly customisable.
• FLUX.1 produces excellent images with genuinely readable text, at the cost of needing more memory, around 12GB and up.
• Qwen-Image is a newer, permissively licensed option that is strong at text in images and designed to fit under 16GB with quantisation.

The licensing catch

This is the part to get right before you build anything. Image model licences vary sharply, and one of the most popular models is non-commercial.

• FLUX.1 [dev] is licensed for non-commercial use only. You can generate images for personal and research use, but selling the output, or using it in a business, needs a separate licence from Black Forest Labs.
• FLUX.1 [schnell] and Qwen-Image are Apache 2.0, so they are safe for commercial work.
• SDXL uses a RAIL licence, which permits commercial use with a list of prohibited uses attached.

If your work is commercial, reach for schnell or Qwen-Image, and read the licence on anything you build a product around.

The tools

Image models are run through a separate application, not a chat interface:

• ComfyUI is the most powerful and flexible, built around a node graph you can save and reuse. It supports new models quickly, at the cost of a learning curve.
• InvokeAI leads with an artist-facing canvas for editing and inpainting, alongside model management.
• Fooocus is the gentlest starting point for SDXL, hiding most settings, though it is now in maintenance only, so it is not the base for a long-lived setup.

What can go wrong

• Assuming any local image model is free to use commercially. FLUX.1 [dev] is not. Check the licence before you sell anything.
• Underestimating VRAM. SDXL is forgiving, but FLUX and Qwen-Image want more. Quantised versions and offloading help, at some cost to speed.
• Chasing quality and ignoring ecosystem. SDXL is not the newest, but its huge library of fine-tunes often makes it the most practical for a given style.

Next steps

For the memory side, see how much memory do I need?. The image generation models and image generation tools in the catalogue cover the full set, each with its licence stated plainly.