GPT Image 2.5 Prompt Guide – The Five-Part Formula

A great image prompt is not a magic spell — it is a specification. GPT Image 2.5 was trained to follow detailed instructions, so the more precisely you describe the deliverable, the closer the first result lands. This guide teaches the five-part formula behind every prompt on this site, using OpenAI's official image-prompting structure.

The five parts of a GPT Image 2.5 prompt

1. Deliverable — what is being made? Start by naming the artifact: "Product photo of…", "Poster design for…", "UI mockup of…". This single opening phrase tells the model which visual language to use before it reads anything else.
2. Subject — what exactly is in the frame? Describe the object, person, or scene with material and detail: "a matte ceramic mug in sand-beige glaze", not just "a mug". Concrete nouns beat adjectives.
3. Composition & setting — where is everything? Camera angle, framing, placement, background: "centered on a travertine pedestal, generous negative space above". This is where you decide what the viewer looks at first.
4. Visual direction — how should it feel? Lighting, color palette, and style: "soft morning window light from the left, palette of sand and cream". One light direction and one palette keep the image coherent.
5. Constraints & copy — what must not appear, and what text goes in? Close with exclusions and text rules: "No text, no logos, no watermark", or Text: "JAZZ NIGHT" in bold serif capitals. Quoting on-image text is essential for clean typography.

A worked example

Here is a raw idea turned into a five-part prompt:

Before (vague):

"a nice photo of my coffee mug for my store"

After (structured):

"Product photo of a handmade ceramic coffee mug in matte sand-beige glaze, centered on a smooth travertine pedestal. Straight-on eye-level composition with generous negative space above; soft warm gray seamless backdrop. Gentle morning window light from the left wraps the curved surface and drops a soft shadow to the right. Palette of sand, cream, and warm gray. Square 1:1 crop, medium telephoto compression. No text, no logos, no props."

Same mug, same idea — but the second prompt specifies the shot type, the light, the palette, the framing, and the exclusions. That is the difference between hoping and directing.

Six tips that make an immediate difference

Put on-image text in double quotes. Write Text: "OPEN 9 AM" so the model renders exactly those characters. Keep it under a handful of words — short text renders most reliably.
State the aspect ratio. Ask for "Square 1:1", "Vertical 4:5", or "Widescreen 16:9". Supported sizes include 1024×1024, 1536×1024, and 1024×1536, plus custom sizes within OpenAI's limits.
Use one light source direction. "Soft light from the left" beats "good lighting". Two competing light directions is the most common cause of muddy results.
Name the palette, not the mood. "Palette of sage green and cream" steers color better than "calm vibes". Three colors maximum.
Iterate one variable at a time. Change only the lighting, then only the angle. You learn what actually caused the improvement — and you can reproduce it.
Re-state your keepers when editing. On follow-up edits, repeat what must stay: "keep the same mug, same light, change only the background". Multi-turn edits can drift; restating anchors them.

Common mistakes to avoid

Stacking adjectives instead of facts. "Beautiful, stunning, amazing" adds nothing. "Brushed steel case, deep navy dial" changes the render.
Describing everything at once. One clear scene beats five half-described ones. If you need a different shot, write a different prompt.
Forgetting exclusions. If you do not want logos, watermarks, or extra props, say so — models otherwise fill empty space.
Ignoring the deliverable word. "Photo of a logo" and "Logo design of…" produce very different outputs. Name the artifact first.

Ready to practice? Grab a prompt from the library, or assemble your own with the free generator.