Image generation is the most commoditised corner of AI. Chasing the best tool is a waste of time; matching the tool to the job is where the value is. Four jobs cover almost all business use, and the legal questions matter more than which tool you pick.
| Job | Right tool | Why |
|---|---|---|
| Quick utility images | ChatGPT / Gemini built-in | Already in the tool you pay for |
| Crafted visuals | Midjourney | The quality benchmark most people use |
| Licence-comfortable work | Adobe Firefly | Trained on licensed stock, enterprise indemnity |
| Volume / in-house | Flux, Stable Diffusion lineage | No per-image fees, full control |
Quick utility images
For diagrams, a picture for a blog post or a rough hero image to brief a designer, use whatever you already pay for. ChatGPT and Gemini both generate decent images inline, let you refine them in conversation, and handle text inside an image far better than they used to. For anything internal or short-lived, convenience beats specialist quality, and it costs nothing extra.
Crafted visuals
When the image is the product (campaign art, distinctive site imagery, anything a designer would once have mocked up), Midjourney is still the benchmark for visual quality and control over style. It takes some learning. The design skill now goes into prompting, choosing between outputs and iterating, and a saved house-style recipe is what keeps the outputs looking like one brand instead of a mood board.
Licence-comfortable commercial work
If your legal team asks where the training data came from, Adobe Firefly is the usual answer. It’s trained on Adobe’s licensed stock, it sits inside the Creative Cloud tools designers already use, and enterprise customers get indemnification. The images are competent rather than the best available. What you’re buying is the paperwork, and for some firms that’s exactly the right purchase.
Volume and in-house generation
At scale, or where prompts and outputs have to stay on your own infrastructure, the open-weight families (Flux and the Stable Diffusion lineage, mainly) run on your own hardware with full control and no per-image fees. It’s the image version of the open-weight LLM decision: real control and real upkeep, worth it for volume or data reasons rather than fashion.
The legal and honesty risks
Three risks come up, most common first:
- Resemblance risk. A tool’s commercial licence doesn’t protect an output that resembles a copyrighted character, artwork or real person. Don’t prompt towards any of them.
- Weak ownership. UK law on copyright in purely AI-generated images is unsettled, so don’t build anything that depends on owning a raw generation outright. Human editing and composition strengthen your position.
- Faking the record. Generated images presented as real photos of your team, premises, work or results are on the same side of the line as fabricated reviews, and platforms increasingly label synthetic media anyway.
Illustrate as much as you like. Just don’t fake the record.