Why a deck needs a set, not pictures
A single illustration is the easy case: describe one mood and you get one frame that looks right. A presentation is the hard case, because a deck is not one picture but ten or twenty that must read as one hand drew them all. The problem is consistency across slides, not the beauty of any single frame.
When each slide is generated from a fresh prompt, the palette drifts, the shape language shifts, and the line weight changes page to page. On screen the deck looks like different designers took turns.
GPT Image 2.5 helps: the launch page says it is “better at reflecting visual styles” and “follows specific editing instructions more reliably across multiple edits.” That is what lets a set stay a set.
This article covers atmosphere and concept illustrations: dividers, section marks, soft visuals that carry tone rather than data. Data-rich diagrams belong to a different job, in the infographic workflow. Keep the two separate.
Lock one style anchor
Before you generate anything, write one small block describing how every illustration should look. Call it a style anchor: it names the palette, the shape language, the lighting, and the level of detail. Once it exists you paste the same block into every prompt and change only the subject line.
Style anchor — paste this block unchanged into every prompt:
Palette: warm off-white #F4EEE5 base, clay #C97B5A, sand #D8C3A5, ink #2B2622.
Shape language: flat, rounded rectangles and discs, no gradients, soft thin outlines.
Lighting: flat even light, no cast shadows, no highlights.
Level of detail: minimal, two or three simple forms per image, no small text.
Reuse: keep this block identical in every prompt; change only the subject line.The anchor does the quiet work. Slides agree about the things the reader is not studying — the clay tone, the rounding of a corner, the outline weight — because every prompt carries the same words. You are pasting the same string.
If the deck needs a recurring subject, a mascot, a product, or a person, the anchor covers the look while a second technique holds the subject steady. That job is in character consistency.
Match the ratio to the picture area
The slide has one ratio, but the picture area usually has another. Generate at the ratio of the picture area, not of the slide, or the image is cropped or stretched in its slot.
The help centre is explicit: “You can also generate images in any aspect ratio. Use the aspect ratio picker to choose a different aspect ratio, or include your desired aspect ratio in your prompt.” State the ratio in words and the picture is composed for that shape from the start.
| Slide role | What the illustration must do | Ratio |
|---|---|---|
| Section divider | Fill the whole slide as a calm backdrop | Slide ratio (landscape) |
| Concept beside text | Sit on one side, leave room for bullets | Picture-area ratio (portrait) |
| Footer strip | Repeat small marks across a thin band | Strip ratio (wide) |
| Opening hero | Carry the first impression on its own | Slide ratio (landscape) |
The table never prints a pixel number, because the official pages publish none. Describe the shape in words and let the picker or prompt set it. The trade-offs are in ratios and sizes.
Two examples: divider and diagram
The picture above is the set this workflow produces: four panels, one anchor, no stray text. Both prompts below start from that same anchor.
One: the section divider
A full-bleed section divider for a slide. Flat illustration filling the whole
frame, no margins and no border. One large calm form low and centred: a pale
disc rising behind two low mounds, in the clay and sand palette. Flat even
light, no shadows, no gradient. No text, no letters, no logo, no watermark.
Wide landscape framing matching the slide. Leave the upper area empty for a
title.Check the image is truly full bleed, that the palette matches your anchor, and that the upper band is empty for the title.
Two: the inline diagram
A half-page concept illustration for the left half of a slide; the right half
stays completely empty for bullet text. Flat illustration: three stacked bars
of different lengths in the clay and sand palette, rounded ends, on the
off-white base. Flat even light, no shadows. No text, no letters. Portrait
framing sized to the picture area. Keep the left side detailed, the right side
blank.Check the picture reads at half width, that it shares the divider’s anchor, and that the empty right side stays clear for your bullets.
Keep the quiet side quiet
Every illustration should leave one side free of detail, so the body text has a calm place to land. Add this one sentence to the end of every prompt:
Leave one side of the image free of detail, forms and texture so body text
can sit there without a clash.This is not a style preference you can skip. It is a requirement: the overlaid text must stay legible, and the empty side is the shelf the sentence rests on, not a decorative choice.
When the text is part of the design rather than an overlay, the rules change: letterforms, hierarchy and placement become their own job. That is covered in text rendering.
Build the set in one conversation
Do not open a new chat for each slide. Keep the entire deck in one conversation from the first anchor to the last edit. The launch page explains why this holds the set together: “During longer ChatGPT conversations, Images 2.5 follows specific editing instructions more reliably across multiple edits. Earlier changes are more likely to stay consistent, and each new edit builds on the work you’ve already done without degrading image quality over time.”
A single conversation is itself a consistency mechanism: the model carries earlier choices forward, so the set does not depend only on copying a style string.
The model is also “better at editing only what you’ve asked for, while keeping the rest of the details the same — even with more complex subjects and backgrounds.” That lets you fix one slide without disturbing the others. The mechanics of targeted edits are in image editing.
Frequently asked questions
Can I use one chat per slide?
You can, but you lose the build-up that keeps a set coherent. The documented multi-turn behaviour depends on one conversation, so separate chats fall back to copying the anchor by hand and drift faster.
Should the prompt name pixel sizes?
No. The official pages publish no pixel dimensions and no slide sizes, so there is nothing to quote. Describe the ratio in words or use the picker, and avoid inventing numbers the vendor has not published.
Will it add text on its own?
It can when you ask, but unintended text can still slip in, so check every image before it enters the deck.
Check the set on the projector
Before you present, run the same short check on the whole set. View every illustration as a thumbnail to spot any panel off the anchor. Check contrast at presentation size, not on your laptop. Confirm no unintended text appeared and that each picture still reads when its slide is dense.