How to create an image with AI: the complete guide

The short answer

This is how you create a first image:

1. Pick a model. Each model is good at something different, and the table below helps you choose by task. All models are in the studio and on the models page
2. Describe what is in the picture. Who or what is in the center, where it happens, the light and the style. Write in English, and keep in Hebrew only text that has to appear inside the image
3. Attach images when needed. A photo of a product, a character’s face or an image you want to edit
4. Choose the aspect ratio and resolution. The exact price shows before you send, and that is the amount that leaves your balance
5. Check and refine. The image is saved to your library the moment it is ready. When something is off, change one detail at a time

To give a sense of price: right now an image in Nano Banana 2 Lite costs ₪0.090.94 crNano Banana 2 Lite, 1 image, a 1K image in Nano Banana 2 costs ₪0.191.88 crNano Banana 2, 1K · 1 image, and a medium quality image in GPT Image 2 costs ₪0.222.21 crGPT Image 2, 1K · 1 image · Medium quality. Each extra image in the same request is charged on its own.

Which model fits which task

No single image model leads at everything. This is the split by what the makers state, what you can pick here and our own experience:

On an independent voting leaderboard from October 2026, GPT Image 2.5 Flare was 2nd of 82 for text to image, and GPT Image 2.5 Sunburst 1st of 59 for image editing. A ranking measures the general preference of voters, not your task and not Hebrew, so the best test is to run the same prompt in two models and compare. The two best-known models are compared in the guide Nano Banana Pro vs GPT Image 2.

Nano Banana 2 or Nano Banana Pro

Google’s two image models are similar but not the same. Nano Banana 2 is the fast model, which Google describes as high quality at a mainstream price, and Nano Banana Pro is the quality model for complex jobs such as infographics. Both take up to 14 images and go up to 4K, and Nano Banana 2 also offers long formats such as 4:1 and 8:1 for banners. Right now a 1K image costs ₪0.191.88 crNano Banana 2, 1K · 1 image in Nano Banana 2 and ₪0.424.22 crNano Banana Pro, 1K · 1 image in Pro.

More models worth knowing

• Seedream 5.0 Lite: editing by example. It learns a change from a before and after pair and applies it to another image
• FLUX.2 [flex] and FLUX.2 [dev]: control over advanced settings such as the step count. Dev makes images from text only here
• FLUX.1 Kontext [max]: the premium Kontext, which its maker says follows edit instructions more closely
• Qwen Image 3 and Qwen Image 3 Pro: small readable text in dense layouts. Pro takes up to 3 input images here
• Recraft V4.1 Pro: the high resolution version of Recraft V4.1, with a 2048 pixel base
• Ideogram 3.0: the previous Ideogram, with preset styles, color palettes and editing of a marked area
• GPT Image 1.5: an older OpenAI model scheduled to shut down on December 1, 2026, so for new work GPT Image 2 is the better pick

How to write an image prompt

A good image prompt describes in plain sentences what you see, the way you would describe a photo to someone who has not seen it:

• The subject. Who or what is in the center, and what it is doing
• The place. Where it happens and at what time of day
• The light and style. For example soft window light, a studio photo, a watercolor illustration or a 3D render
• The camera. A close-up, a top-down shot or a low angle
• Text, if needed. Write the exact words inside quotation marks, and say where they go

This is what a prompt built this way looks like. It is an example we wrote ourselves:

A ceramic coffee mug on a wooden table by a window, soft morning light, steam rising from the cup, a small plant in the background. Close-up product photo, shallow depth of field, warm and calm mood.

One concrete detail beats a lot of general words like "beautiful" or "amazing". The aspect ratio, resolution and number of images are chosen in the settings, not in the prompt. When the result is off, change one thing at a time and keep the wording that worked. Some models, such as Qwen Image 3, also have a "Negative prompt" field, where you list only the things that should not appear, without the word "no".

Editing an image in words

Instead of starting from scratch, you can upload an existing image and write what to change, such as "swap the background for a beach" or "put a red coat on the character". Nano Banana Pro, GPT Image 2, FLUX.1 Kontext [pro] and Seedream 5.0 Pro all edit from a written instruction.

• One change at a time. One clear edit comes out more precise than several changes together
• Say what stays. For example "keep the face and the lighting as they are"
• Go back to the original when needed. The maker of FLUX Kontext notes that many edits in a row can add visual flaws, and when that happens, go back to the original image or the last good version
• Mark an area when you can. In the GPT Image models you can attach a mask that marks which area to change, and according to ByteDance Seedream 5.0 Pro understands marks and sketches drawn on your image

Combining several images

You can also combine several images into one, for example a character from one image in a room from another. Write in the prompt what to take from each image and how to combine them. Nano Banana Pro and Nano Banana 2 take up to 14 images, GPT Image 2 up to 16, Seedream 5.0 Pro up to 10 and FLUX.2 [pro] up to 8. Each model’s maximum shows next to the images field in the studio.

Product photos for an online store

Product photos are one of the most common uses: upload a photo of the product to an editing model and ask for a new background, light and setting, while saying clearly that the product itself must stay as it is.

1. Photograph the product well. A sharp, well lit photo gives the model more detail to keep
2. Attach several angles. Nano Banana Pro, GPT Image 2 and Seedream 5.0 Pro take several input images, so you can attach the product from more than one side
3. Describe the scene. For example the product on a white marble table in soft daylight, and add that the product, the logo and the colors must stay exactly as in the photo
4. Make one image and check it. Only when the product comes out true to the original, make a series with the same wording

For a catalog image on a transparent background, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst support transparent backgrounds, and the file has to be PNG or WEBP, since JPEG does not keep transparency. Right now a 1K image in Nano Banana Pro costs ₪0.424.22 crNano Banana Pro, 1K · 1 image.

The same face across several images

To have the same character appear in several images, attach a clear photo of the face as a reference image in every generation, and say in the prompt that it is the same person. According to Google, Nano Banana Pro is built to keep the look of up to 5 people, and according to ByteDance, Seedream 4.5 keeps the facial features, light and color tone of your reference image.

Use the same reference image and the same description of the character in every image of a series, and change only the scene. Google notes that character consistency is not always perfect even in Nano Banana Pro, so check every image. The full method, including the move from images to video, is in the guide how to keep the same character.

Hebrew text inside the image

In our experience, the new GPT Image models write the best Hebrew inside an image today: GPT Image 2, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst. Nano Banana Pro and Nano Banana 2, Google’s image models from the Gemini family, also write good Hebrew. No maker publishes a language list that includes Hebrew, so the advice rests on real use, and in the other models Hebrew text has not been tested.

This is how to raise the odds of clean Hebrew:

• Write the exact Hebrew text inside quotation marks, and the rest of the prompt in English
• Prefer short text: a headline or a few words, not a paragraph
• Say where the text goes and in what style, for example a large headline at the top of the poster
• Pick a higher resolution for small text: according to Google, in Nano Banana Pro small text is often blurry at 1K
• Check every letter before you use the image, and when the text is long or has to be perfect, add it afterwards in a design app

The full picture, area by area, is in the guide Hebrew in AI models.

Resolution and price

In most image models you pick the resolution before you generate, and it changes the price. Nano Banana Pro, Nano Banana 2 and Seedream 4.5 go up to 4K, and in GPT Image 2 you can pick 4K too, but according to OpenAI sizes above 2560x1440 are still experimental. Nano Banana 2 Lite makes 1K only, and according to Google it is the fastest and cheapest model in the Gemini family.

The prices are computed as the page loads, with the same calculation the studio shows before every generation. More ways to save are in the price guide, and a finished image can also become a video, as the guide how to turn an image into a video explains.

Every model works from the same balance and in the same studio, so you pick the model that fits each image and pay only for what you create.

Questions and answers