Image model by OpenAI

GPT Image 2.5

OpenAI's image model comes in two versions on Predis.ai: Flare for fast everyday images, and Sunburst for careful edits where the details have to hold. Both take the same prompt, up to 5 reference photos, and sizes up to 4K.

A smoked-glass NOCTURNE perfume bottle on wet black marble beside a white orchid, made with GPT Image 2.5 Flare The same NOCTURNE perfume prompt made with GPT Image 2.5 Sunburst
One prompt, one size, two versions. Flip between them to compare the light, the glass and the lettering.
The prompt

Luxury fragrance ad photograph. A faceted smoked-glass perfume bottle with a heavy brushed-brass cap stands on a slab of wet black marble. The word NOCTURNE is etched into the glass in thin, widely spaced capitals, and the liquid inside glows deep amber where a single beam of light passes through it. One white moth orchid lies beside the base, and a few drops of water bead on the marble. Low-key studio lighting: a narrow rim light from behind, soft fill from the left, deep shadows, a faint reflection of the bottle in the stone. Shot on a 100mm macro lens, shallow depth of field, rich blacks, warm highlights. No other text.

1536 × 1024, medium quality. Made on Predis.ai on October 26, 2026. Nothing retouched.

Edits that hold

One photo, four edits

Each step started from the image before it, with one plain sentence about what to change. The machine, the cup and the counter stay put while everything around them moves.

The original

Product photograph of a compact retro espresso machine in glossy cherry-red enamel with polished chrome trim, a round pressure gauge on the front, two chrome toggle switches and a walnut-handled portafilter locked in place. The name FERRO is embossed in small chrome capitals above the gauge. It sits on a pale terrazzo countertop against a warm off-white plaster wall, with morning light from a window on the left casting a soft, long shadow. Three-quarter front view, 50mm lens, crisp detail, editorial interior styling, generous empty wall space in the upper left. No other text.

Made with GPT Image 2.5 Sunburst on October 26, 2026, 1024 × 1024. Each edit used the previous image as its reference.

The FERRO espresso machine, step 1 of 5
Start from your own images

A sketch becomes a product shot. A photo gets a new jacket.

A pencil design sheet of an arc desk lamp with dimensions
The sketch
The same arc desk lamp rendered as a brushed-steel product with the same dimension drawings
The finished lamp

From a design sheet

The sketch went back in as the reference with a seven-word prompt, and the lamp came out in brushed steel with the same arc, shade and measurements.

The prompts, as typed

Sketch: "pencil drawing or deisgn of an arc desk lamp" Finished lamp, with the sketch as reference: "the arc lamp as a finished product"

Made with GPT Image 2.5 Sunburst on October 26, 2026.

A skateboarder in a sage-green work jacket holding a board
Your photo
The same skateboarder, pose and board, now in a navy bomber jacket
One change

From a photo

Only the jacket changed. Face, pose, board and light came through as they were, which is what makes a reference photo worth adding.

The prompt

Full-body studio photo of an invented skateboarder: use the same as the reference image. same person, same posture, same angle. Only thing to change is that he wears a navy blue bomber jacket.

The original was made with Nano Banana 2; the edit with GPT Image 2.5 Sunburst on October 9, 2026.

Edit your photo with GPT Image 2.5
Sizes and quality

From a square post to a 4K banner

Pick the size before you make the image. Square, landscape and portrait each go up to 4K, and OpenAI marks anything above 2560 × 1440 as experimental.

lowmediumhighxhighmax

Five quality settings. The top two, xhigh and max, are new in 2.5.

Flare or Sunburst?

Flare

About 25 seconds

OpenAI's default for everyday images. Reach for it for social posts, quick variations and testing an idea before you commit to it.

Sunburst

About 80 seconds

OpenAI's most capable version, built for editing. Reach for it when a photo has to come back changed in one place and identical everywhere else.

Both take the same prompt, up to 5 reference images, every size and every quality setting, and both make up to 4 images at a time. We'd draft on Flare and finish on Sunburst, but that's your call.

Try Flare and Sunburst
Fine print

Where it still slips

Long text can still go wrong

OpenAI lists text as a known weak spot. Keep headlines short, and read every word before you post.

Edits follow words, not shapes

There's no brush. Say where the change goes, such as "the empty wall at the top left", and the model finds it.

Very large sizes are experimental

Above 2560 × 1440, OpenAI marks output as experimental. Make the image at a smaller size first, then go big on the one you keep.

Busy prompts take longer

A complex prompt can take up to 2 minutes, by OpenAI's own figure. Draft on Flare while you work out the wording.

GPT Image 2.5 questions

What is GPT Image 2.5?

It's OpenAI's image model, announced on September 8, 2026, and the one behind ChatGPT Images 2.5. It makes images from text or from up to 5 reference photos, and it edits an image you give it while keeping the rest as it was.

They're two versions OpenAI released together. Flare is the fast one for everyday images, and Sunburst is the more capable one for editing and campaign work, where details have to hold. On Predis.ai, Flare usually takes about 25 seconds and Sunburst about 80. Both take the same prompts, sizes and quality settings.

Yes. Add your photo as a reference and say what to change, such as a color, a background or a line of text. Everything you don't mention should stay the same, and the edits on this page each started from the one before.

On Predis.ai you can pick from 1024 × 1024 up to 3840 × 2160 (4K), in square, landscape and portrait. OpenAI marks sizes above 2560 × 1440 as experimental, so check those closely.

OpenAI says 2.5 is faster and keeps subjects from reference photos more faithfully, and it adds two higher quality settings, xhigh and max. GPT Image 2 is still in the Predis.ai Model Studio if a workflow depends on it.

Yes, as the headline on the FERRO ad shows. OpenAI still lists text as a known weak spot, so keep headlines short and read every word before you post.

Start from one photo

Upload a product shot and say what should change.

Make an image with GPT Image 2.5
Make an image with GPT Image 2.5

GPT Image is a trademark of OpenAI. Predis.ai is not affiliated with or endorsed by OpenAI.