What is an AI image prompt?
An AI image prompt is the text you give an image model to describe the picture you want it to make. The model reads the words and generates an image to match. A prompt is the whole brief, written for a machine instead of a designer.
Hugging Face's diffusers documentation puts it in one line: text-to-image generates an image from a text description, "which is also known as a prompt". Every image model in common use starts from one, from Stable Diffusion to Google's Gemini image models.
What goes in a prompt
A choice the prompt leaves open goes to the model, and the model fills it with its safest guess. These are the parts worth deciding yourself:
- Subject: who or what the picture is about, and what they are doing.
- Setting: where it happens.
- Light and colour: one desk lamp, a cold blue night, late sun from the side.
- Style: photo, 3D render, flat illustration, pixel art.
- Framing: close-up or wide, the camera angle, where the subject sits in the frame.
- Words: the exact text in quotes, if the image should carry lettering.
Google's prompting guide for its image models builds its templates from the same parts. For a photorealistic scene, that means shot type, subject, setting, lighting, camera angle and lens. The rule behind them is plain: "Describe a scene in rich detail. The more specific you are, the more control you have over the results."
How models read a prompt
Stable Diffusion 1.5 and its peers were usually prompted with comma-separated tags. The example in Hugging Face's own docs is "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k". Newer models read full sentences. A sentence carries what a tag list can't: who stands in front, what they hold, where they look.
A few settings sit beside the prompt rather than inside it:
- Aspect ratio. Gemini's image API takes it as a parameter, and lists 14 ratios for its newest model as of November 2026, 16:9 and 9:16 among them.
- Guidance scale, in Stable Diffusion. A higher value makes the image follow the prompt more closely; too high, and the docs warn of artifacts.
- Seed. The random starting point. Fix it and the same prompt gives the same image.
Prompt vs negative prompt
| Prompt | Negative prompt | |
|---|---|---|
| Says | What the image should contain | What the image should avoid |
| Supported by | Every text-to-image model | Stable Diffusion and similar pipelines, as a separate field |
| Typical content | Subject, setting, light, style, words | "blurry, poor details, bad anatomy" |
Where you see it
In a chat-style image tool, the prompt is the message you type. In Stable Diffusion interfaces such as AUTOMATIC1111's web UI, it is the main text box, with the negative prompt as an extra field under it. In a thumbnail tool it can be out of sight: the tool turns your description, or your video, into a longer prompt before the model sees it.
Why the prompt matters for a thumbnail
A thumbnail prompt has more jobs than an art prompt. The image will be judged at about 168 pixels wide in a phone feed, with YouTube's duration badge over its bottom-right corner, next to a title it shouldn't repeat. "A gamer reacting to a win" leaves the size of the face, the contrast with the background and the empty corner to chance, and those are the choices that decide whether the thumbnail reads.
Left to its defaults, a model drifts toward airbrushed skin and even, waxy light: the look people now call AI slop. That is a prompting failure, not a limit of the model. Write the prompt for the size the thumbnail will be seen at, not the size it is made at.