Guide

GPT Image 2.5 Guide

GPT Image 2.5 is OpenAI's newest image model, in two editions: Flare generates fast, Sunburst holds its precision through detailed edits. Both read long, structured prompts, render exact text, keep the subjects in your reference photos recognisable and change only what you asked for. This guide covers the prompt formula, text and layout work, reference-based editing, transparent backgrounds and multi-turn refinement. Every example loads into the Molyin image generator with one click.

Updated 2026-09-15

01 Core Formula

GPT Image 2.5 follows instructions closely, which changes what a good prompt looks like: you are writing a brief, not casting a spell. Decide what the image is for before you type, name what is actually visible in the frame, and say what must not appear. Short prompts work well. Structured prompts are what give you control.

1.1 The Basic Prompt Formula

Three things carry an image: the subject, the medium it is made in, and the light falling on it. Get those down first, then add framing and constraints. Write down what is actually in the frame rather than how it should feel — materials, wear, colour, texture and camera cues all land, while adjectives like "beautiful" do nothing. If you want a photograph, ask for a photograph in those words.

[Subject] + [Medium] + [Lighting] + optional [Framing] [Composition] [Text] [Constraints] [Mood]

Required

Subject

Who or what, narrowed until it could only be this one: age, material, wear and tear, expression, what the hands are doing.

Required

Medium

Photograph, watercolour, flat vector, 3D render, slide, app screen. Name it in the first sentence and everything else falls in line behind it.

Required

Lighting

North light through a window, a studio softbox, dusk, a single lantern. Light decides how real the picture looks, and it does it faster than any adjective.

Optional

Framing

How much is in the frame and from where: medium close-up at eye level, 50mm, shallow depth of field, head to feet in shot.

Optional

Composition

Where things sit and how they relate: subject centred or off to one side, empty space kept clear for text, gaze leading out of frame.

Optional

Text

Wording that has to appear, in quotation marks, with its position, its typographic treatment and how many times it appears.

Optional

Constraints

The things you do not want: no extra lines of text, no watermarks, no logos, no advertising gloss.

Optional

Mood

One or two words, no more: unposed, grateful, tense. The visible details are what actually carry a mood.

Candid Studio Portrait

Prompt

Photorealistic candid photograph of a ceramicist at her wheel in a daylit studio. Clay-spattered apron, forearms streaked with slip, hands steady on a half-pulled cylinder, eyes down on the work. Grey north light from a tall window on the left, shallow depth of field, 50mm lens at eye level, fine film grain. Chipped shelves and worn tools behind her, slightly out of focus. Real skin texture, ordinary and unposed, no retouching, no studio gloss.

Subject, medium and light in the first two sentences; framing and lens next; a constraint to close. The last line matters more than it looks — without it you tend to get a polished advertising portrait instead of a working one. Loads on Flare at 1024x1536 / medium.

1.2 Structured Prompts: Scene, Mood, Style, Constraints

Once a brief has more than a handful of requirements, stop writing paragraphs and start writing labelled blocks: Scene, Mood, Style, Constraints. Any format you can re-read works — blocks, bullets, even short JSON. The reason to do it is the second round: you can swap one block and keep the rest of the brief intact instead of rewriting a wall of text and wondering what changed.

Anniversary Card in Labelled Blocks

Prompt

Create a greeting card illustration for a neighbourhood bakery's first anniversary. Scene: a flour-dusted marble counter at dawn, one seeded loaf cooling on a wire rack, a folded linen cloth beside it, steam still rising. Mood: warm, quiet, grateful. Style: premium card photography, soft window light, realistic crumb and fabric texture, shallow depth of field, generous empty space across the lower third. Constraints: original artwork, no logos, no watermarks. Include only this text, exactly once: "One year of early mornings. Thank you."

Four blocks, and the card text is quoted and counted so it appears once and nowhere else. Change the Scene block to a florist or a bike shop and the other three blocks still do their job. Loads on Flare at 1024x1536 / medium.

1.3 Real-World Context and People

A place and a date bring their own set dressing: name them and you inherit period-accurate clothing, vehicles and signage without listing any of it. People need more care than objects. Say how much of the body is in frame, how large they are relative to the scene, where they are looking and what their hands are doing, and the figure stops floating in the composition.

Period Scene with People

Prompt

A photorealistic wide photograph of an American drive-in cinema on a warm evening in August 1958, half an hour after sunset. Period-accurate cars parked in rows facing the screen, families on folding chairs beside their bumpers, a boy of about ten standing on a car roof with his hands in his pockets, looking up at the screen. Warm projector beam with insects crossing it, deep blue sky holding the last of the dusk. Natural colour, documentary framing, nothing modern in shot.

The date does the research for you, so the prompt spends its words on the one figure that carries the picture: how big he is, where he stands, where he looks. Check the cars and clothing afterwards — inferred period detail is usually right and occasionally invents something. Loads on Flare at 1024x1536 / medium.

02 Exact Text and Layout

Legible text and a grid that holds are what make this model useful beyond illustration: ads, diagrams, slides, interface mockups and comic panels are all in range. The method is the same for all of them. Quote the exact words, say how many times each appears, describe the layout the way you would brief a designer, and name what is not allowed on the page.

2.1 Render Exact Text

Put the copy in quotation marks, say where it sits and how it is set, and state that it appears once. Unusual names drift — spell them out letter by letter when they do. Finish with a ban on everything you did not ask for: extra lines, watermarks, stray logos. Then read the rendered text before you use the image, because legible is not the same as correct.

Campaign Ad with a Tagline

Prompt

A campaign photograph for a running brand called Ashfield. Four runners of mixed ages crossing an empty city intersection at first light, breath visible, relaxed mid-stride, understated kit in slate and rust. Cool morning light, long shadows, asymmetric composition with clear sky in the upper third. Render the tagline "Miles Make the Map." exactly once, upper left, set in a compact modern sans and clearly legible against the sky. Premium sportswear photography, no other text, no watermarks, no logos.

The tagline is quoted, counted, placed and typeset, which is four separate instructions doing four different jobs. Drop any one of them and you usually get the line twice or buried in a shadow. Loads on Flare at 1024x1536 / medium.

2.2 Diagrams and Infographics

Name the process, say who is reading it and state what they should understand by the end. List every label you need — anything you leave to the model may come back plausible and wrong. Ask for a flat visual system: consistent icons, arrows heavy enough to follow, white space, no decoration. Diagrams are the one output you must fact-check line by line.

Explain a Process

Prompt

Create a technical infographic explaining how an air-source heat pump warms a house in winter. Follow the loop in order: outdoor air across the evaporator, refrigerant compressed, heat handed to the water circuit, warm water out to radiators and the hot water tank, refrigerant expanded and back outside. Label every stage and show direction of travel with arrows. Flat illustration on a white background, one accent colour for the refrigerant loop and another for the water circuit, labels large enough to read on a phone, no decorative clutter.

Spelling the loop out in order is what keeps the arrows pointing the right way; a prompt that just names the parts tends to produce a diagram that looks right and flows backwards. Loads on Flare at 1024x1536 / medium.

Classroom Diagram with Fixed Labels

Prompt

Create a classroom diagram titled "The Rock Cycle at a Glance" for secondary school students. Show igneous, sedimentary and metamorphic rock as three labelled nodes in a cycle, connected by arrows for the processes between them: weathering and erosion, deposition and compaction, heat and pressure, melting, cooling and crystallisation. Mark magma and sediment as intermediate states. Flat handout style on white, simple consistent icons, one accent colour, arrows thick enough to follow at a glance. Large legible text, no tiny captions, no decoration.

Every term the diagram is allowed to use is listed, which is the only reliable way to keep a teaching diagram accurate. Dense labelling is also where the higher quality tier earns its cost. Loads on Sunburst at 1536x1024 / high.

2.3 Slides, Charts and Page Images

Write a spec, not a wish. Name the exact artefact — one slide, a workflow diagram, a chart — then define the canvas, the hierarchy and the real numbers and captions that go on it. Made-up figures come out looking made up, so supply your own. Finish by banning the house style of bad slideware: clip art, stock photography, gradients, drop shadows.

Board Slide with Real Numbers

Prompt

Create one board-meeting slide titled "Retention by Cohort" in the visual language of a well-run software company. Clean white background, modern sans-serif type, generous margins. Include a cohort retention chart of six lines over months 0 to 12, all flattening between 60% and 72%; a left-hand column with three figures, "Net revenue retention 118%", "Logo churn 1.4% / month" and "Payback 9 months"; a single takeaway line under the chart reading "Cohorts flatten after month 4."; and a small footnote "Q3 2026, self-serve plans only". Muted blues and greys, clear data hierarchy, no clip art, no gradients, no drop shadows.

Supply the numbers and the slide becomes a layout job the model is good at. Leave them out and you get convincing-looking figures you cannot use. Loads on Sunburst at 1536x1024 / high, the landscape canvas closest to slide proportions.

2.4 Interface Mockups

Describe the product as something that already ships. List the sections in reading order, name the real interface elements, and keep the language practical — the moment a prompt starts talking about a concept or a vision, the output turns into a design-award poster instead of a screen.

Mobile App Screen

Prompt

Create a realistic mobile app screen for a public library. Show the reader's holds: a header with the branch name and opening hours, a list of three books ready for pickup with small covers, shelf codes and dates, a "Due soon" section with two more titles, and a search field pinned to the bottom. Calm and practical — white background, one muted green accent, clear type hierarchy, comfortable spacing, real interface elements rather than decoration. Present the screen in a plain phone frame.

Sections listed top to bottom, one accent colour, and a section heading in quotes so it renders as written. The portrait canvas matches the phone frame instead of fighting it. Loads on Flare at 1024x1536 / medium.

2.5 Comic Strips and Sequential Panels

A story only survives the jump to panels if you hand over the beats. Say how many panels there are and how they are arranged, then give each one a single visible action — not a feeling, not a plot summary. Keep the cast small and the camera roughly consistent and the strip reads in one pass.

Four-Panel Comic

Prompt

Create a vertical four-panel comic strip in clean line art with flat colour. Panel 1: a cat watches a robot vacuum sit in its dock for the first time, tail low, ears flat. Panel 2: the vacuum starts moving; the cat is already on the kitchen counter, staring down at it. Panel 3: hours later, the cat rides the vacuum across the living room, completely composed, one paw out for balance. Panel 4: the front door opens; the cat sits innocently on the floor beside the docked vacuum. No dialogue, no captions.

One visible action per panel and a running gag that pays off in the fourth. Banning dialogue keeps the model from filling speech bubbles with text nobody asked for. Loads on Flare at 1024x1536 / medium.

03 Editing with Reference Images

Switch to image-to-image, upload what you have and describe the change in ordinary language. This is where GPT Image 2.5 pulled furthest ahead of the previous generation: it touches the element you named and leaves the rest of the frame where it was, and faces and products survive the trip from your reference. Two habits make edits dependable — separate the change from the constraints, and give every uploaded image a job.

3.1 Separate the Change from the Constraints

Write the edit in two halves. First the single change, marked as the only one. Then the list of what must survive it: camera angle, lighting, geometry, layout, labels, everything in the background you were not thinking about. Close with what must not be added — text, logos, shadows, props. The preserve list looks like padding and is the reason the edit works.

Swap One Object, Keep the Room

Prompt

In this cafe photo, replace only the glass pendant lamps above the counter with woven rattan pendants of the same size, hanging at the same height. Keep the camera angle, the room lighting, the reflections on the counter and every other object exactly as they are. Match the warmth of the existing bulbs and cast realistic shadows from the new shades. Change nothing else.

Upload the cafe photo. "Only" marks the change, the preserve list pins the room, and asking for shadows is what stops the new lamps from looking pasted in. Loads on Sunburst at 1536x1024 / medium.

Remove an Object

Prompt

Remove the parked scooter on the left side of the alley. Rebuild the wall and paving behind it so they match the surrounding texture and light. Change nothing else.

Name the object and say what should be there instead, or the gap gets filled with something invented. Small removals do not need the slower edition. Loads on Flare at 1024x1536 / medium.

3.2 Give Each Reference a Role

With more than one upload, stop calling them "the images". Number them and give each a role — image 1 is the scene, image 2 is the subject, image 3 is the garment — then say what moves where and at what scale. Up to 16 references are accepted, but an unlabelled pile of them is worse than two with clear jobs.

Combine Two References

Prompt

Image 1 is the scene: a stone courtyard at midday. Image 2 is the subject: a red cargo bike. Put the cargo bike from image 2 into the courtyard from image 1, leaning against the wall on the right, at a believable scale next to the doorway. Match the direction and hardness of the midday light, give it a correct contact shadow, and leave the rest of image 1 untouched.

Upload the scene first and the subject second, in the order the prompt names them. Scale and contact shadow are the two things that give a composite away, so both are specified. Loads on Sunburst at 1024x1536 / medium.

3.3 Preserve Identity, Change the Rest

When someone has to stay recognisably themselves, spell out what is fixed: face, features, skin tone, hair, expression, body shape, proportions, pose. Then describe the one thing that changes and how it should sit in the picture — matched light, matched shadow direction, matched colour temperature. Most of an identity prompt is the preserve list, and that is the correct shape for it.

Change the Clothes, Keep the Person

Prompt

Image 1 is the portrait. Images 2 and 3 are garments. Dress the person from image 1 in the jacket from image 2 and the trousers from image 3. Do not change the face, features, skin tone, hair, expression, body shape, pose or proportions in any way. Replace the clothing only, fitting the garments to the existing pose with realistic fabric weight and folds. Match the lighting, shadow direction and colour temperature of the portrait so the clothes belong in the photograph. Keep the background, framing and camera angle unchanged, and add no accessories, text or logos.

Upload the portrait first and the garments after it, matching the numbering in the prompt. Loads on Sunburst at 1024x1536 / medium — the precise edition is worth it whenever a face has to come back unchanged.

Put a Person in a New Scene

Prompt

Put this person in a busy night market, walking between food stalls with a paper cup in one hand, looking off to the right at something outside the frame. Keep the face, features and proportions exactly as in the reference. Photographic and grounded: mixed light from stall bulbs and a hand-painted sign, light rain on the pavement, people behind blurred by a slow shutter. Waist-up framing, natural colour. Not cinematic, no colour grading, no poster styling.

Upload one clear photo of the person. Framing, gaze and what the hands are doing are all specified, and the closing line keeps the result a photograph rather than a film poster. Loads on Sunburst at 1024x1536 / medium.

3.4 Style Transfer and Sketch to Photo

A reference does not have to supply the subject. Say that you want its palette, line weight and surface — and nothing else — then describe a completely different subject in the same sentence. Sketches work the other way round: keep the drawing's geometry and hand the model the job of inventing plausible materials and light.

Same Style, New Subject

Prompt

Use the illustration style of the input image — its palette, line weight and paper texture — and draw a new subject with it: a harbour crane at dusk seen from below, on a plain background. Keep the style, change the subject completely.

Upload the picture whose look you want. Naming the three things you are borrowing beats "same style as this", which often drags the original subject along with it. Loads on Flare at 1024x1536 / medium.

Sketch to Photograph

Prompt

Turn this hand-drawn sketch of a cargo bike into a photorealistic product photo. Preserve the exact proportions, frame geometry, perspective and the position of every part as drawn. Choose materials that suit the design: powder-coated steel frame, birch ply box, leather saddle. Studio lighting on a neutral background. Add no parts, no branding and no text that is not in the drawing.

Upload the sketch. Written as a spec: what is fixed, what the model may decide, what is forbidden. Without the last sentence, sketches tend to come back with invented logos. Loads on Flare at 1024x1536 / medium.

3.5 Translate Text Without Touching the Design

Localising an asset is an edit like any other: replace the words, keep the design. Ask for the text and only the text to change, and tell it to fit the new lines into the existing blocks. Then proofread — a line or two sometimes stays in the original language, and that is easy to miss when the layout looks perfect.

Translate a Poster in Place

Prompt

Translate every piece of text in this poster into Spanish. Keep the layout, typeface, colours, spacing and imagery exactly as they are, and fit the translated lines inside the existing text blocks without resizing anything else.

Upload the poster and swap the target language for the one you need. Text-heavy edits are worth the high tier, since that is what keeps small print legible. Loads on Sunburst at 1024x1536 / high.

04 Transparent Backgrounds

Both editions return a real alpha channel. Set Background to transparent in the parameter panel, keep the output format on PNG or WebP, and ask for the isolation in the prompt as well so the two agree. Then look at the edges — hair, glass, thin shadows and fine outlines are where transparency quietly fails.

4.1 Logos on a Transparent Background

Describe the business and the shapes the mark is built from, then say where it has to work: a favicon and a shop sign are a stricter brief than "a logo". Ban the things models like to add around a cutout — a backdrop, a checkerboard pattern, a drop shadow — and generate several at once while you are still choosing a direction.

Logo on a Transparent Background

Prompt

Design an original logo for a bicycle repair shop called Freewheel: a single mark above the name. The mark is a geometric wheel whose spokes resolve into an open hand, with balanced negative space. Flat vector shapes, one line weight, no gradients. It has to stay readable at favicon size and on a workshop sign. Centre it with generous padding on a fully transparent background, clean alpha edges, no backdrop, no checkerboard pattern, no shadow, no watermark.

Loads on Flare at 1024x1024 / medium with Background set to transparent and PNG output. The prompt repeats the transparency requirement on purpose, so the setting and the words do not contradict each other.

4.2 Product Cutouts from a Photo

For a cutout from a photograph, upload the shot, ask for the subject isolated on transparency, and lock the two things that matter commercially: the shape of the product and the legibility of its label. Allow light retouching and nothing more. If you edit the cutout again later, say "transparent background" again or it quietly comes back opaque.

Product Cutout from a Photo

Prompt

Cut the vacuum flask out of this photograph and place it on a fully transparent background. Centred, upright, crisp silhouette, no halo or colour fringing along the edges. Preserve the exact shape, the brushed metal finish and the legibility of the label. Light retouching only — do not restyle the product, do not invent a shadow, and do not add any backdrop, pattern or scenery.

Upload the product shot. Loads on Sunburst at 1024x1536 / medium with Background set to transparent and PNG output; the precise edition is what protects label text and clean edges.

05 Iterate, Stay Consistent, Pick a Tier

The strongest way to work with this model is a chain: generate, feed the result back in as the reference, ask for one change, repeat. Quality holds across a long chain here, so the tenth step is as clean as the first. This chapter covers that loop, keeping a character recognisable across a series, and how to choose between the two editions and the five quality tiers.

5.1 One Change per Turn

Upload the previous output, ask for one change, and repeat what has to survive. Referring back to "the same style" carries some context, but when results start drifting the fix is to restate the constraints rather than add more instructions. Look at each result before you write the next prompt — two changes in one turn and you cannot tell which instruction caused what.

Change One Condition

Prompt

Same scene, late afternoon instead of midday: low warm sun from the left, longer shadows, slightly hazy air. Keep the composition, the subject, the camera angle and every object exactly where they are.

Upload your previous output as the reference. One change, one preserve line. If the composition still shifts, name the elements that moved and run it again. Loads on Flare at 1024x1536 / medium.

5.2 Keep a Character Consistent

Introduce the character on a plain background first — that image becomes your reference sheet, and clutter in it will follow you through the whole series. From then on, upload it with every scene and repeat a short consistency block: outfit, features, proportions, palette. Add one instruction not to redesign the character, because that is the failure mode.

Establish the Character

Prompt

Create a children's book illustration introducing a main character. Character: a nine-year-old lighthouse keeper's apprentice in an oversized oilskin coat and a red wool hat, carrying a small brass lantern. Round face, freckles, calm and determined. Theme: she guides seabirds that lose their way in the fog. Style: hand-painted watercolour, soft outlines, warm earth tones against cool sea greys, picture-book proportions with a slightly large head and an expressive face. Constraints: original character, no text, no watermark, plain background so the character reads clearly.

The plain background is the whole point: it gives later scenes a clean reference of the character and nothing else. Loads on Flare at 1024x1536 / medium.

Continue the Series

Prompt

Continue the story with the same character. Scene: she crouches on wet rocks at the foot of the lighthouse, holding the lantern up so a young gannet can find its way out of the fog. Character consistency: same oversized oilskin coat, same red wool hat and brass lantern, same face, proportions and palette as the reference image. Style: the same hand-painted watercolour, now with fog, sea spray and lantern glow. Constraints: do not redesign the character, no text, no watermark.

Upload the first illustration as the reference. The consistency block repeats what the reference already shows, and that redundancy is exactly what stops the drift. Loads on Flare at 1024x1536 / medium.

5.3 Flare or Sunburst, and Which Quality Tier

Choose the edition first, then the tier. Start on Flare: it is the faster of the two and already ahead of the previous generation, which makes it the right default for volume work. Move to Sunburst when control decides the job — identity edits, product retouching, dense diagrams, long edit chains. For quality, explore at low on a small canvas, produce at medium or high, and keep xhigh and max for a final frame that a lower tier visibly could not deliver. The tier alone sets the price, on both editions and at every size.

Explore Cheap: Four Drafts at Low

Prompt

A modern app icon for a habit-tracking app: a friendly paper aeroplane leaving a dotted loop trail on a rounded-square tile, soft mint-to-cream gradient behind it. Flat vector style, clean shapes, generous padding, no text.

Loads on Flare at 1024x1024 / low, four images per run. Four cheap variants tell you more about a direction than one expensive render of a prompt you have not tested yet.

Finish Sharp: One Frame at Max

Prompt

Premium product hero of a matte black ceramic pour-over coffee set on a slab of raw travertine: a conical dripper, a tall server with a walnut collar and a small matching cup, arranged in a loose triangle. Overhead-left studio softbox with soft fill, faint steam rising from the server, a few scattered light-roast beans. Photorealistic, crisp detail in the ceramic glaze and the stone pores, shallow depth of field, neutral warm palette. No text, no logos, no props beyond the set.

Loads on Sunburst at 2048x2048 / max. Run this preset only once the prompt is proven at a lower tier — max is for the frame you are going to ship, not for exploring.

GPT Image 2.5 Parameter Cheat Sheet

Everything the Molyin image generator exposes for GPT Image 2.5 Flare and GPT Image 2.5 Sunburst. The two editions share one capability set; the difference is speed versus precision.

GPT Image 2.5 (Flare)GPT Image 2.5 (Sunburst)
ModesText to image / Image to imageText to image / Image to image
Pixel size1024x1024 / 1536x1024 / 1024x1536 / 1536x1152 / 1152x1536 / 2048x2048 / 2048x1152 / 1152x2048 / 3840x2160 / 2160x38401024x1024 / 1536x1024 / 1024x1536 / 1536x1152 / 1152x1536 / 2048x2048 / 2048x1152 / 1152x2048 / 3840x2160 / 2160x3840
Qualitylow / medium / high / xhigh / maxlow / medium / high / xhigh / max
Output formatpng / jpeg / webppng / jpeg / webp
Batch size1010
Reference imagesUp to 16Up to 16

FAQ

The most frequent questions, with answers we verified ourselves.