How to write an AI image prompt for a product
Write it in the order you want the model to read it: the product with a pointer to its reference photo, the ground with a hex code, the light as something falling on a surface, the palette as color names with hex codes, one sentence of camera and lens, and a sentence about grain. The prompts behind our kept photos ran about 140 to 200 words on Nano Banana models and about 300 on GPT Image.
The six parts, in order
In our agency’s 2026 campaigns, the prompts behind photos an art director kept on Nano Banana models all had these six parts. Each settles one thing a photographer would settle on set. The subject and its reference always came first. Light and palette sometimes shared a sentence, and camera, grain and color grade came last.
| Part | What to write | From a kept prompt |
|---|---|---|
| 1. Subject and reference | Point at the reference photo, then say what the product is doing and where it sits in the frame. | "The product shown in the reference image sits perfectly centered" |
| 2. Ground | The surface or backdrop, with a hex code. | "on a white surface (#FFFFFF)" |
| 3. Light | The source, its direction and hardness, and what it does to a named surface. | "Sharp overhead flash renders every granule of cocoa powder in crystalline detail" |
| 4. Palette | Two to four color names, each with a hex code, tied to things in the frame. | "Rich cocoa tones (#3E2015) grade outward to lighter cocoa dust brown (#B07D46)" |
| 5. Optics | One sentence: camera angle, focal length, aperture, where focus sits. | "Framed directly overhead at 50mm, f/8, focus even across the entire flat plane" |
| 6. Grain and grade | Grain in its own sentence or phrase, then the color grade. | "Heavy visible film grain throughout." |
Google and Black Forest Labs both put the subject first in their own prompt structures, and Black Forest Labs says FLUX.2 "pays more attention to what comes first." OpenAI’s guide for GPT Image suggests the scene first, then the subject, key details and constraints, and naming what the image is for, such as an ad. Google also asks for the scene described in sentences, and says a list of keywords "won’t cut it". Whichever order you use, keep it for every prompt in a set, so a change in the picture traces back to a change in the words.
A real prompt, taken apart
This prompt made one of ten kept photos for a packaged food ad, rendered on Nano Banana 2. It is 145 words. A person on the job added the grain sentence by hand; nine of the ten kept prompts carry it, and the tenth asks for grain in its own words.
The product shown in the reference image sits perfectly centered on a white surface (#FFFFFF) surrounded by concentric rings of cocoa dust radiating outward like a bullseye target, created by a burst of powder frozen in perfect circular formation. Rich cocoa tones (#3E2015) grade outward to lighter cocoa dust brown (#B07D46) at the outermost ring. Sharp overhead flash renders every granule of cocoa powder in crystalline detail while casting a tight shadow directly beneath the brownie. The crumb texture of the brownie’s top surface echoes the dusty rings below. Framed directly overhead at 50mm, f/8, focus even across the entire flat plane, composed as a perfect circle radiating from center. Heavy visible film grain throughout. Color grading references medium format Hasselblad digital, clinical sharpness, neutral whites, controlled warmth in the brown spectrum, pronounced grain, high contrast at center, saturation naturalistic and grounded in earth tones.
| Sentence starts | Part | Why it is written that way |
|---|---|---|
| "The product shown in the reference image" | Subject, reference, ground, action | The model learns which picture is the product before anything else, and the ground gets an exact white. |
| "Rich cocoa tones (#3E2015)" | Palette | Each color is a name plus a hex, tied to a thing in the frame: the rings. |
| "Sharp overhead flash" | Light | The light acts on surfaces. It renders the granules and casts a shadow in a named place, directly beneath the product. |
| "The crumb texture" | Texture | Ties the product’s surface to the scene, so the two read as one photograph. |
| "Framed directly overhead at 50mm, f/8" | Optics | Angle, lens, aperture, focus and composition in one sentence. A top-down frame of a flat plane at f/8 keeps everything sharp. |
| "Heavy visible film grain throughout." | Grain | Added by hand. It appears in nine of the ten kept prompts on the job, and only 14 of the job’s 120 prompts. |
| "Color grading references medium format Hasselblad digital" | Finish | Names a camera look, then spells it out in plain words (neutral whites, high contrast at center) in case the name is read loosely. |
How long should a product prompt be?
Long enough to settle every decision you care about. In our work, length followed the model.
| Model | Length | Shape |
|---|---|---|
| Nano Banana 2 and Nano Banana Pro (Google) | About 140 to 200 words in our kept prompts | Subject and reference first, then ground, light and palette, one sentence of optics, then grain and grade. The ones we quote end there, with no list of rules. |
| GPT Image 2 (OpenAI) | About 300 words in our kept prompts | Opens "This is a photograph of", names each input image and its job, lists what must not change, and closes with rules such as "No text, no graphics". |
| FLUX.2 (Black Forest Labs) | We did not use it on these jobs | Black Forest Labs calls 30 to 80 words "usually ideal for most projects" and 80 or more for complex scenes. It takes no negative prompts. |
In the GPT Image prompt quoted in full in how to art direct AI images, most of the extra length is rules: which input image is the person and which is the garment, every construction detail of the trousers, and "nothing added and nothing removed." OpenAI’s guide says long prompts can work well, and that problems are easier to trace if you start from a clean base prompt and make one change per follow-up.
What to leave out
- Praise words. "Stunning", "premium" and "high quality" name no light, surface or color, so the model picks for you. Write the decision behind the word.
- Stacks of camera specs. One sentence of optics is enough. OpenAI warns that detailed camera specs "may be interpreted loosely". The words worth using are in camera and lens words.
- Two lights that disagree. "Golden hour" plus "hard overhead flash" asks for two key lights from different directions. Pick one key and say where it is. See lighting words.
- Words you do not want printed. A headline in the prompt invites the model to set type. Render the photo clean and add type in a second pass: how to get text right in AI images.
- "No" lists on models that ignore them. FLUX.2 has no negative prompt, so write "sharp focus" instead of "no blur". GPT Image prompts can end with a short list of exclusions, which OpenAI’s guide recommends stating outright.
- Objects and brands from the mood board. Take the light and the feeling, and leave the things behind.
A fill-in template
The product shown in [the reference image / INPUT IMAGE 1] [sits, stands, leans, falls, is held] [where in the frame]. [Surface or backdrop] ([hex]), [how it falls off: seamless, grading darker toward the edges, out of focus]. [Hard or soft] [light source] from [direction relative to the camera], [what it does to a named surface], [where the shadow falls]. [Color name] ([hex]) on [object], [color name] ([hex]) on [object], [accent name] ([hex]) kept small. Shot at [focal length], [aperture], [camera height and angle], focus on [the part that must be sharpest], [crop]. [Heavy visible / fine / near-zero] film grain throughout. Color grade references [film stock or camera look], [contrast], [saturation], [white balance].
- The prompt says which reference photo is the product, and only the photos this shot needs are attached.
- The ground has a hex code.
- There is one key light, with a direction and a hardness.
- Every brand color is a name plus a hex code, tied to an object.
- There is one sentence of camera and lens.
- Grain is named, or left out on purpose.
- Nothing in the prompt is a word you do not want printed on the image.
- The aspect ratio is set in the tool’s size setting.
Questions people also ask
- Do hex codes in a prompt work?
- They help, as a hint. Black Forest Labs advises marking the code with the word "color" or "hex" and tying it to an object. Every hex in our kept prompts sat next to a color name and a thing in the frame, as in "lighter cocoa dust brown (#B07D46) at the outermost ring". Check the result with an eyedropper, and see how to keep brand colors accurate for fixes.
- Can I reuse one prompt for every product?
- Reuse the parts that make a set look like one shoot: the ground, light and camera sentences, word for word. Rewrite the subject and action for each product and shot. On our trouser job, the backdrop and lighting sentences stayed identical across the store photos, and the framing and pose sentences changed from photo to photo.
- Should I let a chatbot write the prompt?
- It can draft one, but check the draft against the six parts. When you generate images through OpenAI’s Responses API, the main model rewrites your prompt for the image model and returns the revised version, so read what was sent before you judge the result.
- Should I write prompts in JSON?
- Black Forest Labs says FLUX.2 reads structured JSON prompts for complex scenes. Google’s guide asks for the scene in sentences. If your pipeline writes JSON, keep the six parts as fields in the same order.
Where Overs fits
Overs writes a prompt for every photo as one of its ten steps, after the shot plan and before a quality check, and you can watch each step. The export keeps the prompt, model and reference pictures behind each photo, so the prompts behind your keepers are there to reuse.
Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.