ChatGPT Images 2.5: Sketch & Edit Images in Singapore

OpenAI released ChatGPT Images 2.5 on 8 September 2026. If you have only ever typed questions into ChatGPT, here is the short version: ChatGPT can also make pictures, and Images 2.5 is the model that now draws them. Sharper detail, more natural lighting, and generation up to 50 per cent faster than Images 2.0, according to OpenAI's launch post. It went out to every ChatGPT tier, free plan included, on desktop, phone and web. No new app, no waitlist.

For context, OpenAI says people create more than 3 billion images a week across ChatGPT and its API. Most of that is throwaway fun. A decent slice of it, though, is the boring visual work small businesses in Singapore pay for every month: a product photo, a flyer, a menu board, a worksheet for tuition class. If you have been quoted S$80 to S$200 by a freelancer for one poster, this update is worth twenty minutes of your evening.

It also replaces something. Our earlier walkthrough of ChatGPT Image 2 still explains the basics, but the model, the editing and the interface have all moved since. Treat that piece as history.

The headline feature is Sketch. Instead of describing an image, you draw a rough one inside ChatGPT and let it finish the job. Type @Sketch in the chat box and a small drawing window opens. Draw with your finger or your mouse, tap the tick, then type a line about style.

This sounds gimmicky until you try to describe a layout in words. "Sofa on the left, TV console against the far wall, window behind the sofa" falls apart by the third sentence, and the image you get back is nothing like the room in your head. Sketch flips the job: the drawing carries the layout, and you only supply the look. Reviewers have already handed it a mouse-drawn doodle and gotten a realistic photo back, so the bar for drawing skill is about zero. One box for the sofa is enough.

Two other things shipped with it. Templates give you starting formats for flyers and product photos when you have no idea what to draw. And you can now comment on a specific part of an image. Point at the eyes and ask for green, instead of re-rolling the whole picture and losing the parts you liked. Every edit is saved as its own version, so an earlier image you preferred is one scroll away.

1. Open the ChatGPT app or chatgpt.com and sign in. A free account works. If you have never used it, the account is just an email address.
2. Type @Sketch in the message box, or open Images in the sidebar and pick Sketch at the top.
3. Draw rough shapes for the layout. Boxes and circles are fine. You are showing where things go, not making art.
4. Tap the blue tick, then write one or two lines about style: colours, mood, "clean photo" or "flat illustration".
5. Send it. If one detail is off, do not redraw. Comment on that part of the image and change only that.
6. Generating a fresh image costs more of your daily allowance than editing an existing one, so comment-edit before you re-roll.

One catch worth knowing up front: the free plan caps how many images you get per day. OpenAI does not publish the number and it moves with demand, so the limit that matters is the one the app shows you.

Each of these is a real request from a real small operator. Paste them as-is, but replace anything in [square brackets] with your own details. Never put NRIC numbers, full addresses or customer names into a prompt.

1. Home baker: Instagram flyer for the weekend orders

Turn my sketch into a finished Instagram flyer, square 1:1.
Keep the layout exactly as drawn: a photo of a [pineapple tart] in the top half, empty space in the bottom third for text.
Style: warm daylight, clean off-white background, minimal, no sparkles or glitter.
Do not add any text. I will place the price and order details myself.

2. Shopee or Carousell seller: a main image that passes the rules

Edit this photo of my product. Keep the product exactly as it is: same shape, colour, material, including the small scratches in the second-hand condition I am selling it in.
Put it on a pure white background, centred, filling about 70 per cent of the frame.
Square 1:1, output 1024 x 1024 pixels.
Add a soft natural shadow underneath so it does not look like it is floating.
No text, no logo, no watermark, no props.

3. HDB renovation: see the living room before you commit

My sketch is the layout of a 4-room HDB living room seen from the doorway. It is roughly 3.4m by 3.8m, with the built-in unit on the left where I drew it.
Turn it into a realistic render in three versions: warm wood Japanese minimal, light grey Scandinavian with a fabric sofa, and dark industrial.
Keep the built-in in the same position in all three. No text or pictures on the walls.

4. Parent or tutor: a science diagram without copyright trouble

From my sketch, draw a clean cross-section of water moving up a plant stem for a Primary 5 science worksheet.
Plain white background, thin black outlines, simple shapes a 11-year-old can copy by hand.
Leave it unlabelled. I will add the labels myself so they are spelt correctly.

5. Anyone with a shoebox of old photos

Clean up this old family photo. Remove the scratch marks and the crease across the top right corner.
Keep the faces exactly as they are. Do not change features, ages or skin tone.
Do not colourise it if the original is black and white.
Keep the same framing and aspect ratio.

For the renovation one, remember the render is a mood, not a quotation. Send the picture to your contractor for a price, and check HDB's rules before anything is built. Our HDB renovation guide covers the paperwork side.

Text inside images is the biggest trap. Short English words usually come out fine now. Long lines, small print and Chinese characters still break often enough that you should not trust a menu board, a price list or a name card to it. Sketch the layout in ChatGPT, then set the actual words in Canva, Google Slides or Word where you control the font and the spell check.

Second, accuracy. AI editing will happily "improve" your product until it no longer matches what arrives in the parcel. That is a refund waiting to happen, and platforms treat misleading listings as a seller problem, not a buyer problem. Check the final image against the physical item before you upload.

Third, privacy. Do not upload photos of colleagues, customers, students or patients. Under PDPA you are handing personal data of other people to a third party, and they have not agreed to it. Your own family photos are your call, but think first. The same applies to screenshots: bank statements, SingPass screens and medical letters should never go into the box.

Finally, provenance. OpenAI attaches C2PA metadata and an invisible watermark to what it makes, and the metadata travels with the file in most cases. Fine for a poster, worth knowing if you are running ads or claiming an image is a photograph you took.

• Free: S$0, with a daily cap on images and slower generation.
• Go: US$8 a month, about S$10. More images, same models.
• Plus: US$20 a month, about S$25 at the current rate of roughly S$1.27 to the US dollar. This is the tier for people who use it weekly.
• Pro: US$200 a month. Only makes sense if you generate images all day.

OpenAI bills in US dollars, so a Singapore card usually adds a foreign currency fee of one to three per cent on top. Compare that with local rates: a freelance designer here typically quotes S$80 to S$200 for a simple poster, and a product photo shoot starts north of S$300. Two months of Plus costs less than one poster. If you would rather learn it properly than muddle through, AI courses at local training providers can be paid with SkillsFuture Credit. Check your balance on the SkillsFuture portal before you pay cash. The free plan is genuinely enough to test all five jobs above.

ChatGPT sits in our AI tools directory alongside the alternatives worth comparing, including Gemini and Canva, which handle text-heavy designs more gracefully.