Choosing an image tool

Three kinds of product, priced very differently

Bundled inside a chat assistant. Convenient, included in a subscription you already have, weakest control. Fine for an illustration for a slide.

A dedicated hosted service. Midjourney, Adobe Firefly, Ideogram, Recraft and others. Better output, more control, a monthly fee or a credit balance. Each has a house style you will recognise once you have looked at a hundred images.

Open models on your own machine or rented hardware. Stable Diffusion and its successors, Flux, and others. Free after setup, unlimited, fully controllable, and the only route with no content restrictions, which is a benefit and a responsibility.

What separates them in practice

Text in the image. Historically the great weakness, now the clearest difference between products. Some render a shop sign or a poster headline correctly; some still produce plausible-looking nonsense. If your work involves any words in the picture, this alone decides the choice, and it is testable in one prompt.

Consistency across images. A character or a product that looks the same in six pictures. This is what separates a usable set from a collection of one-offs, and the mechanisms differ: reference images, character features, or training a small adapter on your own pictures.

Editing an existing image. Inpainting a region, extending the canvas, removing an object, changing one element while keeping everything else. For commercial work this matters far more than generating from nothing, because most real jobs start from a photograph.

Style range. Some products push everything toward a glossy house look that is attractive and immediately recognisable. Others stay neutral and do what you ask. Neither is better; they suit different work.

The free path, properly

You can do all of this without paying anyone.

• Fooocus: the easiest way to run an open image model locally. Effectively one download, a simple interface, good defaults.
• ComfyUI: the powerful one. A node graph, a steep first hour, and complete control including inpainting, upscaling and multi-step workflows.
• Krita with the AI diffusion plugin: free painting software with generation built into a real editing tool, which is the most practical combination for anyone who actually retouches images.
• GIMP and Photopea for editing afterwards. Photopea runs in a browser, opens PSD files, and costs nothing.

A machine with 8 GB of video memory runs current open image models comfortably. Less than that works with smaller or quantised variants, more slowly.

The commercial question

Before using generated images in paid work, check three things.

The service’s terms on commercial use. Most paid tiers permit it. Some free tiers do not, and some make your generations public by default, which is a problem for client work.

The model’s licence, if you are running it yourself. Image models carry a wider range of licences than language models, including some that restrict commercial use.

Whether an indemnity is offered. Some providers train only on licensed or owned material and offer legal cover for business customers. If you are producing advertising or packaging, this is worth more than image quality.

And the unresolved part, stated as unresolved: whether training image models on scraped work was lawful is being litigated in several countries with different starting points, and whether a purely generated image can be copyrighted at all depends on jurisdiction. That affects what you can protect, not usually what you can make.

The two things that decide quality more than the model

A reference image. Nearly every capable tool now accepts one, and supplying a photograph of the product, the room or the person removes most of the guessing. A generated image built from a reference is far closer to what you wanted than the same tool given only words, and this is the single largest improvement available to a beginner.

Iterating rather than re-prompting. People generate, dislike it, rewrite the whole prompt and generate again. The productive habit is to keep the image you nearly liked and change one thing (inpaint the hand, extend the background, adjust the light) because you are refining a result rather than rolling dice a second time.

How to choose in one afternoon

Take three real jobs from your own work, not test prompts. Run each through two hosted services’ free trials and one local setup. Judge on four things in this order: did it get the text right, could you fix the one thing that was wrong, does the style suit the client, and what will a hundred of these cost.

The last question is the one people skip, and it is why professionals who generate images all day end up running an open model locally regardless of which hosted product they preferred.