Sketch lets you draw directly inside ChatGPT and use that drawing as the layout reference for the finished image, instead of describing the composition in words.
It shipped with GPT Image 2.5 on 8 September 2026 and solves a specific problem: telling a model where things go in a paragraph is slow, and the model still guesses.
What GPT Image 2.5 Sketch does
Sketch turns a rough drawing into visual guidance. You draw boxes, circles and arrows, then add written instructions, and the model builds a finished image that follows your layout.
OpenAI describes it as drawing directly in ChatGPT “as a reference for your final image”, and it is aimed at room layouts, clothing outlines or quick scribbles rather than finished artwork.
What matters is what the drawing controls. It carries position and proportion, not detail, so a five-second sketch of two rectangles already tells the model more than three sentences of description.
How to open Sketch in ChatGPT
The entry point differs between clients, and the official documentation describes it per surface.
- On the ChatGPT mobile app, type @ in the message box and select Sketch
- The official launch notes describe the same feature as typing @Sketch in ChatGPT
- The canvas offers colour, erase and undo controls for the drawing itself
- You confirm the drawing with a checkmark before the message is sent
- Then you add instructions covering subject, colours and style, and send the message
- On the API side this is image editing with a reference image, not the @ menu
Both descriptions point at one feature. The part worth remembering is that the sketch is a step you confirm before the written prompt goes out, not something you attach mid-sentence.
The sketch sets layout, the prompt sets content
Treat the two inputs as separate jobs rather than two ways of saying the same thing.
- The sketch answers where: position, count, relative size, empty space
- The prompt answers what: subject, material, palette, style, language of any text
- Exact wording that should appear in the image belongs in the prompt, quoted
- Elements you do not draw are not automatically excluded, so name anything forbidden
- Aspect ratio belongs to the prompt or the ratio selector, not to the drawing
A good default for the written half is one sentence naming format, subject, style and exact text, which is how a house sketch becomes a watercolour of a red house surrounded by trees.
What to draw for posters, product shots and infographics
Because the sketch only carries layout, the drawing is mostly about reserving zones.
- Poster: logo block at the top, headline in the middle, date and venue at the bottom, hero image to one side
- Product shot: product anchored on one side, copy column on the other, even margin around both
- Infographic: stacked panels, one per step, with a reserved strip for labels
- Social images: leave one clear area for text instead of filling the frame
These formats overlap with the rest of our GPT Image 2.5 coverage, including the GPT Image 2.5 overview for what the model can do, and the GPT Image 2.5 API overview if you need to batch the same layout.
Why a rough sketch beats a long prompt
Describing a layout in prose forces the model to interpret spatial words, and interpretation is where results drift.
- Spatial words such as left, behind and upper right stay ambiguous to a model
- Sketching removes that ambiguity in seconds
- Drawing quality is irrelevant, because the model reads structure rather than strokes
- Rounds get shorter, since you correct one zone instead of rewriting the whole prompt
- The same sketch stays reusable across stylistic variations of one layout
When Sketch results miss the mark
Sketch guides composition, and it does not guarantee the rest.
- Edits can bleed beyond the area you marked, because selections are not pixel-exact
- Text can come back wrong when the wording was vague rather than quoted exactly
- Dense layouts still lose small type, so keep the number of text zones low
- Style instructions can override the sketch when the two conflict
- Complex layouts take longer, since generation time scales with the request
When something lands wrong, change one variable at a time: the sketch zone first, then the wording, then the style, so you know which input caused the shift.
A worked example: sketching a slide layout
A sketch is easiest to justify when the layout is the thing you are unsure about. Take an infographic slide that has to explain a four-step process.
- Draw the frame the way you want the eye to travel: a headline band across the top, then four boxes in a single row, each with a small icon above the label
- Keep the drawing coarse. Blocks and arrows are enough; the model is reading the arrangement, not your line work
- Write the content separately and quote it: “A slide titled THE SOLAR FLARE. Four steps labelled Build-up, Release, Acceleration, Impact. No body copy beyond those four labels.”
- Send the sketch and the content together, then check whether the reading order survived
- To change the arrangement, redraw the sketch instead of adding more words
The failure mode is treating the sketch as decoration: a rough drawing plus a vague prompt gives the model two weak signals instead of one strong one. The drawing should carry the structure and the text should carry the content.
The control that shipped beside Sketch is covered in the GPT Image 2.5 templates guide.
Frequently asked questions
Do I need drawing skills to use Sketch?
No. The sketch carries position and proportion, so boxes, circles and arrows are enough. Rough is expected, and a messy drawing often works better than a tidy one because it leaves less for the model to misread.
Where is Sketch available?
It ships with GPT Image 2.5 across ChatGPT on desktop, mobile and web. The official step-by-step instructions describe the @ menu on the mobile app, while the launch notes describe typing @Sketch, so confirm the exact entry in your own client.
Can I use a sketch to edit an existing image?
Yes, and that is where it earns its place. Upload the image, mark the zone you want changed, and describe the change, leaving the rest of the composition untouched.
Does Sketch work for non-Latin text in the image?
It behaves the same way as ordinary generation. Put the exact characters in the written prompt and use the sketch only to reserve space for them, which is what keeps multi-language posters readable.
Sketch is not a replacement for prompting, it is a second channel for the part of a prompt that is hardest to write. Draw the layout, quote the text, and change one variable at a time.