Image Generation
The picture comes out close to what you had in mind. Do not expect an exact match. You refine it with prompts round after round, and precision does not always come. It is hardest with text inside the picture, with a specified number of objects, and with how objects are placed relative to each other.
Models draw best what they already know well: objects and scenes that appeared often in their training data. A rare concept comes out worse.
Researchers found that pattern in 2024 after testing 34 models, image models among them. None of the services in this chapter were in that test, so what carries over to your chat is the tendency, not its size.
Professional diagrams come out worse. A wiring diagram, a chart, or a drawing looks plausible in a picture, but the connections, labels, and details in it may be wrong. Do not put an image like that into your work until a specialist has checked it.
How fast this is changing is visible in the GenExam benchmark. Researchers gave models 1,000 exam problems that required a drawing, across 10 subjects from physics to economics. A picture counted only if every detail was right and no labels were wrong.
OpenAI's GPT-Image-1, released in April 2025, solved 13 percent of them. Google's Nano Banana Pro came out in November and solved 70 percent.
The problems in the test are academic. It gives you no numbers for your own work diagrams, but it shows that in seven months models got noticeably more accurate at drawing them. Keep an eye on what new versions can do.
The technology is moving quickly. Some services already place furniture in an empty room from a photo of it and pick out decor while keeping the perspective and lighting of the original shot.
Try this in Google Gemini. Photograph your own room, send the picture into the chat, and ask for furniture to be placed in it or for what is already there to be rearranged. The new version of the room arrives as an image in the same conversation.
Compare services on one task. Photograph a room with an empty wall and ask for decor suggestions, or shoot your yard and ask to see it renovated.
Send the same photo with the same request into each of the three. The results show which of them to trust with work like this.
The same limitation applies here: ask to move the armchair in that picture, and something you never asked to touch will change along with it.
It is worth trying. A task that did not work half a year ago may go through on the first prompt today. Repeat your old failed attempts on new versions.
If generating images is part of your job, ask your model which services exist for it and how they differ. There is no point listing them here: the list will turn over while the book sits on a shelf.
One thing will hold. A general-purpose chat is weaker than a service built only for images: it passes your request to a separate model and steers that model for you. An image service hands the steering to you.
For professional work, when you need many images, there is one more option: a local model on your own computer. It has no service limits, and the photos and descriptions do not go into someone else's cloud. How that setup works is covered in the chapter "Where the Model Runs".
Among local image models there is Stable Diffusion from Stability AI, with files downloaded from Hugging Face.
The license terms for version 3.5 allow free commercial use if the company's annual revenue is no more than $1 million. Companies above that ask Stability AI for a separate license.
This option is for experienced users. You install and configure the software yourself, and you need a powerful graphics card. The medium version of Stable Diffusion 3.5 needs 9.9 gigabytes of video memory according to its developer, and that is before the memory used to process the text of your prompt.