Text to image · Image editing · Transparent backgrounds
Generate and edit images with OpenAI GPT Image 2.5. It is the model to reach for when the render has to carry a headline, a price, a label or a UI — and its edit route changes what you asked for while the subject, composition and background hold. Sign in and generate on this page, then take the result into Soku AI to resize, iterate and ship it as ad creative.Generate and edit images with OpenAI GPT Image 2.5.
Describe the image you want, including any words that have to appear in it. Choose an aspect ratio; the exact credit cost appears before you generate.
Sample result gallery
Nineteen runs from the generator on this page, at the same flare tier and quality setting you get. Nothing was retouched, no text was composited on afterwards, and on the three edit pairs the source is itself a GPT Image 2.5 render.
Two lines of copy at poster scale, set across a high-contrast four-colour illustration. Nothing here is composited — the letterforms, their spacing and the ink texture behind them all came out of one render.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
The offer is inside the image rather than in the ad caption: headline, product and a legible price badge in one frame, at the contrast a feed placement actually needs.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Continuous glass tubing is a hard case — every character has to be one unbroken run and still read as a word. The bloom, the wet brick and the coloured reflections come with it.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Dialogue has to sit inside its balloon, stay in reading order and survive heavy inking. The sound effect is lettered as artwork rather than typeset.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
The type runs from a flat colour field onto a photograph and has to stay legible over both. Paper grain, print colour and the raking light are one render.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
A menu is the honest test of an image model: every row pairs a name with a number, and one wrong digit makes the whole render useless. Chalk texture and smudging survive alongside the type.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Painted lettering is irregular by definition, which is exactly where invented characters usually creep in. Three price signs and a banner, all correct, over produce in full sun.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Foil is a material as much as a typeface: the letterforms have to catch the light along their edges and stay readable in the shadow side of the same word.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
The hardest kind of text: small, dense, and meaningless if one field drifts. Each label sits above its own value, the barcode stays on the stub, and the tear line separates them.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Twelve labelled bars, twelve value labels and a real axis. The values rise and fall in the shape the labels claim, which is what separates a chart from a picture of one.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Labelled stat tiles, a legend and a plotted series on an 8px grid — the layout logic holds instead of dissolving into decorative UI-shaped noise.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Packaging has to be repeatable: the same label system in three colourways, each with its own product name and weight, lit as one commercial shot.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
The wordmark, the sub-line and the small-caps footer are all rendered by the model, correctly spelled and correctly kerned, over a product it also had to light and place.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Four labels have to land on the four details they name. The style, the aged paper and the plate caption come along with it.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
Generated with a transparent background rather than cut out afterwards, so the enamel edge and the brass rim stay clean when it is composited over anything.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
The counterweight to the rest of this sheet: no text at all. One window as the only light source, glaze pooling at the feet, and a long shadow with the right falloff.
Source: Our GPT Image 2.5 flare run · quality high · no retouching
The instruction was time of day and weather only. The building, the window display, the stacked books and the camera position are the ones from the source frame.
Source: Our GPT Image 2.5 flare edit · the source is the render beside it
The product is the part that must not move: same shape, same matte terracotta glass, same unlabelled front. Everything around it — surface, prop, light direction, shadow — is new.
Source: Our GPT Image 2.5 flare edit · the source is the render beside it
Headline, time and door line all become Japanese at the same weight and optical size, while the arc, the paper, the palette and every type position stay where they were.
Source: Our GPT Image 2.5 flare edit · the source is the render beside it
Features
Quote the exact words in the prompt and they come back spelled, kerned and placed — headlines, price rows, badges, UI labels and callouts. This is the reason to pick this model over a cheaper one.
The edit route takes your image plus an instruction and holds the subject, composition and background around it. Name what must not move and it stays put, across several rounds of revision.
Ask for a transparent background and you get an asset with a genuine alpha channel — no key, no halo, no manual cut-out before it can be composited.
Square, 16:9, 9:16, 4:3 and 3:4 are generated at that shape rather than cropped into it, so a vertical placement is composed vertically.
Describe the scene, then state any text that must appear using the words "reads exactly" followed by the copy. Vague direction produces vague lettering; a quoted string produces the string.
Choose one of the five aspect ratios. The exact credit charge for the run is shown before you submit, and signing in happens in a dialog on this page.
Switch to the Edit tab to revise the result with a second instruction, or open it in Creatives to resize it for each placement and build variants.
Why Soku
A model playground gives you a file. Soku gives you the next step: sign-in, credits, generation, revision and the handoff into ad creative all happen without leaving the workflow.
Sign-in, credit check, payment and generation happen here. There is no separate playground to move to and no API key to provision.
Feed a result straight back through the Edit tab. The subject and layout hold, so a revision is a revision rather than a fresh roll of the dice.
Open a finished image in Creatives to produce each placement, then run the variants as ads and let performance decide which one earns the next round.
Use cases
Performance creative teams
Price badges, promo copy and CTA text render legibly, so the hook is in the creative itself instead of relying on the ad caption to carry it.
Brand and design teams
Hand the edit route a finished piece and a target language. The type changes; the arc, palette, spacing and product do not.
Product and content teams
Labelled diagrams, field-guide style plates and product screenshots hold their layout logic, which is what makes them usable in a post rather than decorative.
Specifications
These are the controls the live endpoint accepts today; the studio above is generated from the same server-owned specification.
Alternatives
Both are current image models and both are one click apart in Soku. They are different answers to different jobs.
Best for
In-image text
Transparent output
Edit references
Relative cost per image
Reach for GPT Image 2.5 when a wrong character makes the render useless. Reach for Nano Banana 2 when you need many images and the words are not the point.
Pricing
New users can create an account without leaving the page. Signed-in users see their live credit balance and the exact cost of the run before it is submitted.
Billed per generated image; the edit route costs marginally more
$0/ month
No card required
$19/ month
7,000 credits / month
$49/ month
20,000 credits / month
If the balance is too low, the page opens the plan and top-up flow. A provider-side generation failure refunds the committed credits. See all plans and the credit calculator.
Write the prompt, quote the copy that has to appear, and pick a shape. Sign-in, credits and the handoff into Creatives all happen here.
Start with my asset
Soku AI