Testing Image Generation models

AI imagery: from prompting to directing

AI image tools are developing quickly, both in what they can produce and how we work with them. Two recent experiments brought this into focus for me: comparing how different models responded to the same brief, then exploring a more hands-on approach to editing the results.

‍

Finding the right starting point

With so many image generators available — Midjourney, ChatGPT Image, Gemini, Firefly, Flux and others — I wanted to see how differently they would interpret the same task.

I had a specific image in mind, so I gave more than ten AI models exactly the same prompt and three reference images.

‍

‍

The results were surprisingly varied. Some followed the brief closely. Others changed the model, moved or removed objects, or interpreted the references in unexpected ways. One made the model about two-and-a-half times the size she should have been.

The practical lesson was that different models suit different jobs. Learning their strengths and building a toolkit gives you more options when a particular approach isn’t delivering.

But choosing a model is only the starting point. The next question is how effectively you can refine what it produces.

Taking control of the edit

Using the same reference images, I explored GPT Image 2.5 in ChatGPT, this time focusing on the editing experience.

Working with annotation, drawing and mark-up tools felt very different from text prompting alone. I could sketch approximate forms, identify specific areas and indicate colours directly on the image.

‍

‍

For this kind of work, I found it more intuitive and substantially faster than constructing the image manually. It suggested a production workflow in which design teams could spend less time building individual elements and more time on art direction, judgement and refinement.

There are limitations. Successive edits can cause quality to deteriorate or introduce unexpected changes. When that happened, I found it more effective to start a fresh generation chain with updated inputs and prompts. Reducing complexity helped more than continuing to add instructions.

Together, the experiments showed me how much both model choice and working method matter. A useful toolkit gives you options; visual editing gives you a more direct way to guide the result. The tools will keep changing, but deciding what an image needs — and recognising when it is right — remains central to the work.

‍