ChatGPT Images 2.5: 11 Stunning Examples of What It Can Create

OpenAI released ChatGPT Images 2.5 on 8 September. It makes images up to 50% faster than the last version, and you can now draw straight onto a picture to tell it what you want changed.

People have been posting results since it landed. Here are 11 of them, including the ones that show where it still slips.

OpenAI Launches ChatGPT Images 2.5

This is the first big update to the image model since it launched. For developers it arrives as two separate API models.

Flare is the fast one. It aims to match the old model's quality at up to 50% lower latency, and it is the sensible default for everyday work. Sunburst takes longer and goes after precision, for premium creative jobs where the detail matters more than the wait.

The older versions gave themselves away on texture. Surfaces looked too clean, backgrounds went grainy, and objects had that plastic sheen. Images 2.5 puts the small imperfections back, which is most of what makes a rendered photo read as a real one.

The bigger change is editing. You used to describe a fix in words and hope. Now you point at it.

• Markup lets you circle or scribble on an area and say what should happen there.
• Point and comment works like sticking notes on a photo, and you can leave several before you submit.
• Smart eraser takes an object out and fills in what should have been behind it.
• Background remover cuts the subject out in one click, and it copes better with portraits than with busy scenes.

It is rolling out to ChatGPT and ChatGPT for Work, and to Codex, on desktop and mobile and web. OpenAI says every plan gets it.

Users Are Putting ChatGPT Images 2.5 to the Test

Here is what people built with it in the first two days.

1. A fake Apple event slide. Ben Geskin had it produce a presentation slide for the long-rumoured iPhone Duo, shown as a passport-style foldable with a cover display and an inner screen.

The slide carries a full set of invented specs in a clean layout with real white space. There is a lot of text on it, and the type and colour still do not shout AI at first glance, which older models never managed.

2. A hand-drawn fashion sketch turned into a runway shot. Yana Welinder, founder of Kraftful, compared a sketch of a blue top and a green pleated skirt against what Images 2 and Images 2.5 made of it.

Images 2 was faithful but flat. The 2.5 version came back with real lighting, fabric texture and depth, while keeping the design she actually drew.

3. The four editing tools, demonstrated. Paul J Lipsky walked through all four in canvas tools on YouTube.

He circled an empty patch of a living room floor and typed "Put a cat here." A cat appeared in that exact spot and nothing else in the room moved.

Then he pinned several comments at once, "Make this a dead plant" on a houseplant and "Put a rainbow outside" on a window. Both landed, though it sometimes caught the object next door by mistake.

The smart eraser was the interesting one. He highlighted a potted plant on a table, and the model worked out what the tabletop behind it should look like and filled it in.

4. A travel magazine page. Min Choi asked for a magazine-style travel guide, blending a personal photo into Mordor.

What came back is a dense 9:16 spread with landscape notes, key sights and access tips, plus some Lord of the Rings jokes. The typography holds together like a real photo book page.

5. Game and app UI you could actually ship. Anul Aggarwal generated pixel-art RPG buttons and icons themed around falcons, labelled with things like FLY and HUNT and NEST.

He also got a full set of pink screens for a women's health app called bloom, with progress trackers and navigation elements. The backgrounds came out properly transparent and the styling held across the whole set.

That last part matters more than it sounds. Batching dozens of assets that share a palette, line weight and style is the slow, expensive part of solo game and app work.

6. Head to head against Images 2. Brendan Jowett ran both new models against GPT Image 2 across 15 scenarios.

On a fruit realism test, Image 2 produced apples and pears that were too perfect, which is the usual giveaway. Flare and Sunburst both put in the dots and surface blemishes real fruit has.

The harder test stacked several instructions into one prompt. A giant teapot floating over a desert, a miniature pine forest, one red cabin, and a waterfall pouring from the spout into a circular pool. Image 2 poured the water into the desert. Both new models got every part right.

7. A mirror that actually reflects. One user generated a scrambled Rubik's Cube where the reflection matches the visible faces and their orientation.

That is a genuine spatial reasoning win. The replies were less kind, pointing out wrong sticker counts and colour pairings that break the rules of a real cube. So it handles complex geometry far better than before, and it is still not exact when an object has strict mathematical rules.

8. The noise problem is not fixed. Mark Kretschmann asked it for a detailed forest clearing with green foliage.

At normal size it is convincing. Zoom in and there is patterned noise in the leaves and grass, and other people in the thread saw the same thing. This is an old complaint and it has survived the update.

9. Sketch mode, then video. Jerrod Lew drew a rough tree as a guide and prompted alongside it.

He got a cinematic bonsai in a museum case, labelled "Resilience c. 200 BCE", then ran it through video tools to animate it.

10. Holding one style across different subjects. A creator built a reusable Art Deco poster prompt with tight rules. Square format, gold and navy linework on cream paper, a palette held to terracotta and teal and antique gold. Geometric forms and sunburst motifs, and nothing photorealistic or modern.

Every image kept the palette and the design language even though the subjects were completely different. Consistency like that is what turns a prompt into something you can reuse.

11. Adding an object to a real photo. Fernando del Moral gave it a photo of himself against a wooden background and asked it to put an unreleased iPhone in his hand.

His face, clothes, pose and the background all stayed put, and the phone sits naturally in his hand. This is the least flashy example here and probably the most useful, because taking one real photo and changing one thing about it is what most people actually want.

Takeaways

The speed and the realism are real, and the editing tools are the part worth trying first. Pointing at a spot on the picture beats writing another paragraph of prompt and hoping.

It is still not clean everywhere. The forest noise shows up when you zoom in. The Rubik's Cube breaks its own rules, and the point and comment tool sometimes edits the object next to the one you meant. None of that stops it being useful, but if you are producing something at full resolution, check it at full resolution.