If your third edit in ChatGPT keeps undoing your first one, you are not prompting badly. Until this week, the image model simply could not hold a composition steady across several rounds of changes. ChatGPT Images 2.5, released by OpenAI on 8 September 2026, fixes most of that and adds three tools most people have not opened yet: Sketch, on-image Comments and Templates. Here is how to use each one for real marketing work, with prompts you can paste today.
What is ChatGPT Images 2.5 and what actually changed?
ChatGPT Images 2.5 is OpenAI's image generation and editing model released on 8 September 2026. Compared with Images 2.0, it produces sharper detail, keeps reference photos more recognisable, follows edit instructions more reliably across multiple turns, and generates up to 50% faster. It ships with new in-app tools: Sketch, Comments, Templates and shareable prompts.
The definition worth remembering: Images 2.5 is an edit-first model. The headline improvement is not prettier pictures, it is that the model now understands what not to change when you ask for a small fix.
According to the official OpenAI announcement, people already create more than 3 billion images a week across ChatGPT and the GPT-Image API, and the new model is available to every ChatGPT tier, ChatGPT Work and Codex on web, desktop and mobile. Two API variants also launched: GPT-Image-2.5 Flare (fast, default) and GPT-Image-2.5 Sunburst (slower, more precise).
If your team lives inside Google Workspace instead, the closest comparison is Google's own image tool, covered in our guide to what Google Pics is and how it works.
One honest caveat up front: this is a rolling release. If you do not yet see Sketch or Templates in your account, wait a day or two rather than assuming you are on the wrong plan.
How does multi-turn editing work in Images 2.5?
Multi-turn editing means you generate an image, then refine it through a chain of follow-up instructions in the same chat. In Images 2.5, each edit builds on the previous result rather than regenerating from scratch, so earlier changes stay in place and quality does not degrade over five or six rounds the way it often did in Images 2.0.
The practical rule is to stop stacking instructions into one giant prompt. Give one change per turn, in the order a designer would work: layout first, subject second, text third, colour last.
A typical product-shot sequence for a Hong Kong skincare brand now looks like this:
--- Turn 1: "Studio product photo of this serum bottle on a pale marble surface, soft top light, 4:5 portrait."
--- Turn 2: "Keep everything. Add two water droplets on the bottle shoulder."
--- Turn 3: "Keep everything. Replace the marble with brushed dark slate."
--- Turn 4: "Keep everything. Add the text 'New Formula' in small white sans-serif at the top left."
The phrase "Keep everything" is doing real work. It tells the model the rest of the frame is locked, which is exactly the behaviour OpenAI says it trained for in this release.
How do you use Sketch in ChatGPT to control layout?
Sketch is a new drawing canvas inside ChatGPT. You type @Sketch (or pick the canvas from the menu), draw a rough layout with a mouse, finger or stylus, add a text description of style and details, and submit. The model treats your doodle as a composition guide and generates a finished image that follows where you placed each element.
This solves the single most annoying problem in AI image work: you can describe a layout in 80 words and still get the headline in the wrong corner. A 15-second scribble communicates position far better than prose.
Try this for an event banner:
--- Step 1: Type @Sketch and draw a large rectangle across the top third, a circle bottom-left, and three small boxes in a row bottom-right.
--- Step 2: Paste this prompt: "Use my sketch as the layout. Top rectangle is the headline area, leave it as clean dark navy. Bottom-left circle is a portrait of a female speaker in a blazer, Hong Kong office background. The three boxes are sponsor logo placeholders in light grey. Flat corporate style, 16:9."
--- Step 3: On the result, request text as a separate turn: "Keep everything. Put 'AI Workflow Summit 2026' in the headline area, bold white sans-serif."
Sketch is rough by design. It reads position and proportion, not fine detail, so do not try to draw the speaker's face. Draw where the face goes.
What do Comments on images do, and when should you use them instead of a prompt?
Comments let you tap or click a specific spot on a generated image and leave a note such as "make this cup blue." The model edits only that region and leaves the rest untouched. Use Comments for localised fixes to a single object; use a normal text prompt when the change affects the whole image, such as lighting, style or aspect ratio.
The difference matters because text prompts describe, while comments point. "The mug on the left" is ambiguous when there are two mugs; a comment pinned to the mug is not.
Good candidates for a comment rather than a prompt:
--- A misspelt word in generated signage (pin the word, write "spell this as 'Causeway Bay'")
--- A wrong colour on one product in a line-up (pin it, write "match the colour of the bottle next to it")
--- An extra finger or object artefact (pin it, write "remove this, fill with background")
--- A logo placeholder that should stay blank (pin it, write "keep this area empty grey")
A gotcha from early testing reported by TechRadar: comments work best one at a time. Pin three notes at once and the model sometimes blends them. Pin, apply, check, then pin the next.
How do Templates help you start faster, and what are their limits?
Templates are guided starting points for common image formats such as Poster, Merch, product photo, flyer or icon. Instead of a blank prompt box, you choose a format and fill in fields for the message, design elements and style. The model then assembles a first draft that already respects the conventions of that format, which cuts the blank-page problem for non-designers.
For a marketer, Templates are most useful as a way to lock format before you spend turns on creative details. Pick Poster, and the model already knows it needs a hierarchy of headline, subline and call to action.
The limit is equally clear. Templates encode generic conventions, not your brand. A template poster will be well structured and completely anonymous. Treat it as scaffolding: generate from the template, then run two or three "Keep everything" turns to swap in your palette, typography feel and photography style.
If your team already keeps a brand style block, paste it as the very first message in the chat before choosing the template. Images 2.5 is noticeably better than 2.0 at holding a stated visual style across the whole session.
What is a complete Images 2.5 workflow for a real marketing asset?
A reliable Images 2.5 workflow has five stages: set style once, fix layout with Sketch, generate a base image, refine with single-change turns and Comments, then export at the sizes you need. Done in this order, a social campaign visual takes roughly 8 to 12 minutes, most of it spent checking text and brand details rather than regenerating.
Here is the exact sequence for a WhatsApp and Instagram promo for a Mong Kok bubble tea shop:
Stage 1: Style anchor. First message: "For this whole chat, use this visual style: warm afternoon light, matte pastel palette (peach, sage, cream), clean Japanese minimalist layout, no gradients, no neon."
Stage 2: Layout. @Sketch a tall cup slightly left of centre, an empty band across the top, and a small rounded box bottom-right.
Stage 3: Base image. "Follow my sketch. The cup is an iced brown sugar milk tea with visible pearls and condensation. Top band stays clean for text. Bottom-right box is a price tag placeholder. 4:5 portrait."
Stage 4: Refinement. "Keep everything. Add the headline 'Buy 1 Get 1 Every Tuesday' in the top band, bold rounded sans-serif, dark brown." Then comment on the price box: "write HK$32 here in the same brown."
Stage 5: Export. "Keep everything. Give me a 1:1 version for the feed and a 9:16 version for Stories, keeping the cup fully visible in both."
Save the chat. Next week's promo is a Stage 4 change, not a restart.
What still goes wrong in Images 2.5, and how do you work around it?
Images 2.5 still struggles with dense Chinese typography, exact brand colour codes, and any request that mixes a global change with a local one in the same turn. Long Traditional Chinese headlines can render with wrong or invented characters, hex colours are approximated rather than matched, and "change the background and fix the logo" often does neither cleanly.
Work-arounds that hold up in practice:
--- Chinese text: keep on-image Chinese to four to eight characters, then verify every character. For anything longer, generate the visual with a clean empty band and add the copy in Canva or Figma.
--- Brand colours: describe colours in words ("deep forest green, close to Pantone 3435") and accept a near match, or recolour in a design tool afterwards. Do not trust a generated hex value.
--- Mixed instructions: split them. One global change per turn, then one comment per local fix.
--- Reference photos of people: fidelity is much improved, but always check hands, teeth and jewellery before anything goes to a client.
Also remember that OpenAI embeds C2PA metadata and an invisible watermark in every image. That is a feature for provenance, but it means you should disclose AI generation where your client or platform policy requires it.
Try it now: a 10-minute exercise
Open ChatGPT, type @Sketch, and draw three shapes: a wide band across the top, a large circle in the centre, a small square bottom-right. Then paste the prompt below and follow the four turns. You will have a usable LinkedIn visual and, more importantly, a feel for how the new tools divide the work between you and the model.
Paste this after your sketch:
"Use my sketch as the layout. The top band is a clean charcoal headline area, leave it empty. The centre circle contains a flat illustration of a laptop with a chat window on screen. The bottom-right square is an empty light-grey box. Modern flat corporate style, muted blue and charcoal palette, 1.91:1 landscape."
--- Turn 2: "Keep everything. Add the headline '3 Prompts That Saved Me 5 Hours' in the top band, bold white sans-serif."
--- Turn 3: Comment on the grey box: "write 'Read more' here in dark blue."
--- Turn 4: "Keep everything. Warm the lighting slightly and give me a 1:1 version too."
If any turn changes something you did not ask for, reply "Undo that. Keep everything and only [the change]." That single sentence is the most useful habit to build with this model.
The bigger lesson is that Images 2.5 rewards people who work like designers: fix structure before decoration, change one thing at a time, and point instead of describe whenever you can. Learn the tool's cold edges and it becomes a warm, fast collaborator. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Turn One Good Image Into a Repeatable Content Engine
Sketch, Comments and Templates get you a great visual in ten minutes. The next step is building that into a workflow an AI marketing employee runs for you every week. We'll walk you through every step, from style anchors and prompt libraries to review checkpoints and publishing.
