GPT IMAGE 2.5 SKETCH TO IMAGE · STRUCTURE, NOT STYLE

GPT Image 2.5 Sketch to Image AI Generator
Draw it badly. That is the point.

A rough shape is enough to lock the composition, and the model fills in everything else — including the words you scribbled on it.Upload a phone photo of a napkin, a graph-paper layout, a wireframe. Sketch to image reads it as structure rather than as a style to imitate, so your proportions and the position of every element survive.

Worked sketch to image runs · drawing included

Drawing

The same river valley rendered photorealistically, standing exactly where the sketch put it

Result

OPENAI · PUBLISHED PAIR

Turn this drawing into a photorealistic image. Preserve the exact layout, proportions, and perspective. Choose realistic materials and lighting consistent with the sketch intent. Do not add new elements or text.

Think
HD
1K
3:2
PNG

Full prompt

The valley and the motorcycle are OpenAI's own published pairs, with their published prompt. The koi is from OpenAI's public prompt library; the croquis is reference material held for this page. Anything carrying Full prompt is written up in the prompt library.

First render, no login
Draft is a fifth of the price
Your layout, kept
Words on the sketch are read
Failed runs are free
Model ID on every result

1 free image, no card, no login · failed runs cost nothing · Draft costs 3 credits

Draw it here, or photograph the one on your desk

What GPT Image 2.5 sketch to image does with your drawing

Sketch to image turns a drawing into a finished picture: you hand over a sketch, GPT Image 2.5 reads it as the layout rather than as the artwork, and your prompt supplies the medium, the materials and the exact wording. Composition and proportions survive; the pencil lines do not. A phone photo of paper is a valid sketch.

But the model can read that drawing two completely different ways, and by default it does not know which one you meant.

What survives
Treated as structure
Composition, proportions, where each element sits
Treated as style
Only the hand-drawn feel
What you get
Treated as structure
Your layout, finished
Treated as style
A different picture that looks sketchy
When you want it
Treated as structure
Almost alwaysthe default here
Treated as style
Only when hand-drawn is the final look
How to ask for it
Treated as structure
Sketch is structure — on by default
Treated as style
Say use as style reference only

How a sketch is read: as structure or as style
	Treated as structure	Treated as style
What survives	Composition, proportions, where each element sits	Only the hand-drawn feel
What you get	Your layout, finished	A different picture that looks sketchy
When you want it	Almost alwaysthe default here	Only when hand-drawn is the final look
How to ask for it	`Sketch is structure` — on by default	Say `use as style reference only`

Say the sketch is a structure reference, not a style reference

This is the sentence that separates "my poster, finished" from "a different poster that looks like a pencil drawing". It is on by default here, and it is written out in full further down so you can take it to a tool that has no toggle.

Phone photo of paper is fine. So is a napkin. Grid lines, pencil smudge and wonky perspective are all fine — that is what the input usually looks like.

OpenAI's own published examples

Sketch to image, before and after: a drawing in, a photograph out

Two examples from the GPT Image 2.5 prompting guide. The first is a sketch to image run of exactly the kind this page is about; the second is what happens when a picture is used as a style reference instead of a structural one — the distinction the section above is about.

prompt Turn this drawing into a photorealistic image. Preserve the exact layout, proportions, and perspective. Choose realistic materials and lighting consistent with the sketch intent. Do not add new elements or text.

A river valley sketch in flat contour lines, the input to the sketch to image run — SOURCE · LINE DRAWING

The sketch rendered as a photorealistic river valley, layout and perspective unchanged — GPT IMAGE 2.5 · STRUCTURE REFERENCE

Three clauses do the work: preserve layout, choose materials consistent with intent, add nothing. That is the whole recipe on this page.

OpenAI's published examples, from the GPT Image 2.5 prompting guide · cited, not ours

Five of seven sketches have words on them

Sketch to image with text: the words you scribbled on it

Think about what people actually sketch. Logo drafts have a brand name on them. Poster layouts have a headline. Packaging roughs have a product name and an ingredient line. Wireframes have button labels. Five of the seven things people sketch most often have words in them — and that is the part most sketch to image tools quietly drop.

Why edge detection loses your handwriting
An edge-detection pipeline reads handwriting as lines, not as language. So it renders something that looks like writing from across the room and falls apart when you zoom in. It isn't a bug in those tools — it is what the technique does. Your letters were never words to it.
Looks like writing until you zoom in
Don't make the model read your handwriting. Type it.
The sketch holds the layout, the prompt holds the language. Write the exact words in quotes and say what may not appear:Text in the image: "SINGLE ORIGIN" as the headline, "Ethiopia · Yirgacheffe" underneath in small caps. No text other than what is specified.The box you drew is still doing the work — it tells the model where the headline goes. The prompt only has to tell it what the headline says.
Position from the sketch, spelling from the prompt
Spell unusual names letter by letter
Y-I-R-G-A-C-H-E-F-F-E. An invented or foreign brand name has no token the model has seen spelled correctly, and it will produce something plausible instead. Thirty seconds of hyphens is cheaper than finding the misspelling after it is printed.
The cheapest fix on this page
Small or dense labels? Run it at HD.
Standard is fine for shapes and rough for text you'll actually read. This is one of the few places where the quality tier reliably changes the outcome rather than just the price — so it is worth the four extra credits exactly here and not much else.
Draft for composition, HD for the keeper
Drafting a logo? Ask for a flat white field.
On Starter and above, set Background to Transparent with PNG and add "no checkerboard" to the prompt — without that clause you get the failure OpenAI documents in its own guide: a painted checkerboard, which is a picture of transparency rather than transparency. On the free tier, ask for the mark centred on a flat pure-white field with no shadow and key that colour out; on a hard-edged logo it usually comes away clean in one step.
Flat white beats a painted checkerboard

Frames on this band are OpenAI's own published GPT Image 2.5 examples, from the Images 2.5 announcement of September 8, 2026.

Getting words right in text to image

Take this anywhere, including to tools without the toggle

How to write a sketch to image prompt that doesn't wander off

Name the sketch as the structural reference
Use the uploaded sketch as the structural reference: keep its composition, proportions and the position of every element. This is the line Sketch is structure inserts for you. It is printed here in full because it works in any tool that takes a prompt.
The line
Say what to render it as
Render it as {style}. A medium, a material, a lighting condition — not an adjective. "Photorealistic product shot on a marble surface, soft window light from the left" is a brief; "beautiful" is not.
Then
Quote the text, and forbid the rest
Text in the image: "{exact words}". Do not add elements that are not in the sketch. The second clause is the one that stops invented captions, stray logos and fake watermarks arriving uninvited.
Then
Two images? Say which is which.
This is the part almost nothing else lets you do. Upload your sketch and a reference photo of the look you're after, then assign each one a job: Image 1 is the sketch — use it for structure only. Image 2 is the style reference — use it for palette, material and lighting only.It is straight out of OpenAI's prompting guide, and it is the difference between "a poster shaped like mine" and "my poster, in that studio's visual language". Unlabelled references get averaged instead of combined.
Two references

A sketch to image prompt is three or four clauses long, and every clause here is a sentence you can read back and check against the picture you got. That is the whole test of a prompt technique: if you cannot tell whether it worked, it did not do anything.

What arrives here on a Tuesday morning

Logo drafts, wireframes, posters: what sketch to image is used for

The sketches that turn up here are rarely art. They are the rough version of something that has to get made by Friday — and five of these six have a word written on them somewhere.

Napkin sketch to logo
A mark drawn in three strokes, plus the brand name typed into the prompt so it is spelled the way you registered it. Ask for it centred on a flat pure-white field with no shadow, run it at HD, and key the white out afterwards. Four Draft attempts to settle the shape, one HD for the file you hand over.
Type the name; never let it read your handwriting
Wireframe to UI mockup
Boxes and button labels in, a rendered screen out. The wireframe decides where the nav, the hero and the cards sit; the prompt decides the platform, the type and the palette. Quote every button label you actually want — an unlabelled box gets filled with plausible nonsense.
One screen per run reads better than a flow
Poster layout to finished poster
The thing posters need is the headline in the right place, at the right size, spelled right. Draw the block where it goes, quote the words in the prompt, and say what may not appear. This is the use case the whole page is built around, and the one edge-detection tools are worst at.
Draw the text block, type the text
Packaging rough to product shot
A bottle outline with a label panel scribbled on it becomes a lit product photo. Keep the panel proportions in the sketch, put the product name and the ingredient line in the prompt, and name the surface and the light — "matte ceramic, soft window light from the left" does more than any adjective.
Name the material and the light direction
Floor plan or room layout to a render
The one job here with no words in it, and the one where the competition does fine. A plan view holds the geometry; say the camera height and the direction of the light or you get a plan view back as a picture of a plan view. Worth knowing that a ControlNet tool will hold your lines more literally.
Say where the camera is, or it stays a plan
Fashion croquis to a photographed look
The drawing is the specification, not a suggestion — say that in those words. Name the two colours, the sleeve, the hem length and the closure, then change only the medium and the setting. Everything you do not name is a thing the model is allowed to reinvent.
Spell out every detail you refuse to lose

Every one of these runs on the same sketch to image AI generator at the top of the page. The sketch and the sentence are the whole difference.

Contrast and closed shapes, not drawing skill

Which sketches work, and which ones struggle

Sketch to image does not care how well you draw. It cares whether there is enough contrast and enough closed shape for the model to work out where one thing ends and the next begins — which is a property of the pen and the light, not of the hand.

Line quality
✅ Works
Closed shapes, clear zones
❌ Struggles
Broken, sketchy strokes
Why
It has to work out where one thing ends
Photographing paper
✅ Works
Phone photo, even lighting
❌ Struggles
Hard shadow, glare, your thumb in frame
Why
The pre-upload check flags it before you pay
Where the words go
✅ Works
A box where the text goes
❌ Struggles
Handwriting with no box around it
Why
A position is more useful to it than a word
Density
✅ Works
One subject, one background area
❌ Struggles
Dense overlapping scribble
Why
Nothing to parse
Contrast
✅ Works
Pencil smudge, grid paper
❌ Struggles
Very faint lines on white
Why
Too little contrast for the edges to read

Sketch inputs that work and inputs that struggle
	✅ Works	❌ Struggles	Why
Line quality	Closed shapes, clear zones	Broken, sketchy strokes	It has to work out where one thing ends
Photographing paper	Phone photo, even lighting	Hard shadow, glare, your thumb in frame	The pre-upload check flags it before you pay
Where the words go	A box where the text goes	Handwriting with no box around it	A position is more useful to it than a word
Density	One subject, one background area	Dense overlapping scribble	Nothing to parse
Contrast	Pencil smudge, grid paper	Very faint lines on white	Too little contrast for the edges to read

If yours is in the middle column, the fix is usually thirty seconds with a darker pen, not a different tool.

The pre-upload check runs in your browser. Nothing is uploaded and nothing is charged.

Attempt four is the one that matters

Iterating on a sketch render: it came back close, change one thing

Sketch to image renders almost never land on the first try, and that's fine. The useful question is not what attempt one looks like — it is whether attempt four is still made of your pixels.

Every edit re-renders from your original sketch
Not from the last result. Chained editing is where quality goes to die — each pass re-encodes the one before it — and nothing here chains. Edit number four is computed from the same sketch edit number one was.
Nothing degrades as you iterateEvery edit re-renders from your original sketch
Tick what must stay put before you ask
Composition, text, lighting, product geometry. Four seconds of checkboxes, and it is the difference between an edit and a re-imagining — which on a sketch render means the difference between fixing the colour and losing the layout you drew.
Five checkboxes, written into the prompt for you
Every version is kept
Nothing overwrites anything. Open the version strip, click any earlier render, branch from it. On a page whose whole workflow is iteration, an interface that shows one image and replaces it is the wrong shape.
Branch from any earlier attempt

How image to image editing works here

Where each of them beats us

GPT Image 2.5 sketch to image vs ChatGPT @Sketch, Krea and Free.ai

Three other places to run a sketch to image conversion, and two of them win outright on something. Worth knowing before you spend an afternoon here.

Account required
Here
No, for image one
ChatGPT @Sketch
Yes
Krea
Yes
Free.ai
No
Pick model, quality and size
Here
Yes
ChatGPT @Sketch
No
Krea
Partly
Free.ai
No
Tells you which model ran
Here
Receipt on every result
ChatGPT @Sketch
No
Krea
No
Free.ai
No
Live canvas — draw and watch it change
Here
No
ChatGPT @Sketch
No
Krea
Yesthey win, outright
Free.ai
No
Free renders per day
Here
1 anonymous
ChatGPT @Sketch
With a ChatGPT account
Krea
Limited
Free.ai
~20, commercial rights includedmore generous than us

Sketch-to-image alternatives compared
	Here	ChatGPT @Sketch	Krea	Free.ai
Account required	No, for image one	Yes	Yes	No
Pick model, quality and size	Yes	No	Partly	No
Tells you which model ran	Receipt on every result	No	No	No
Live canvas — draw and watch it change	No	No	Yesthey win, outright	No
Free renders per day	1 anonymous	With a ChatGPT account	Limited	~20, commercial rights includedmore generous than us

When to come here instead: the sketch has words on it

Krea built the real-time canvas — you draw and the picture changes under your hand. A GPT Image 2.5 render takes seven to ten seconds, which is not a live canvas and no amount of engineering makes it one; if drawing-and-watching is the workflow you want, go use Krea. Free.ai runs FLUX with ControlNet edge detection, which is purpose-built for holding your lines, and gives away far more than we do.Everything above is about shapes. If your sketch is a room layout or a character pose, any of these will serve you and two will do it cheaper. But the moment there's a brand name on the napkin, a headline on the poster or a label on the wireframe, that's a language problem rather than an edge-detection problem — and that is the one thing GPT Image 2.5 is unusually good at.Come here when the sketch has words on it.

Krea and Free.ai capabilities checked on their own pages, September 9, 2026.

Draft first. It's a fifth of the price.

What a sketch to image render costs

Sketch to image work is iteration — you'll run the same drawing four or five times before the composition is right. Do that at Draft, then spend HD once on the keeper. Same model, same prompt, roughly a fifth of the bill.

Per image1024×1024, credits off your balance

Draft — for finding the composition

3 credits

about half a cent upstream

Standard — fine for shapes, rough for small text

6 credits

HD — the one you keep, and anything with labels

24 credits

A five-attempt session: four Draft, one HD

36 credits

against 120 if every attempt were HD

What you are not charged for

Credits come off when an image comes back. A run that fails, times out or gets blocked by moderation costs nothing, and the pre-upload check is free forever because it runs in your browser and costs us nothing to give you.

Measured API cost per image

Ten answers, starting with the one everybody thinks first

Sketch to image: frequently asked questions

My drawing is terrible. Does that matter?

No. Closed shapes and clear zones matter; drawing skill doesn't. What breaks it is faint lines on white paper, dense overlapping scribble and hard shadows in the photo — all of which are fixable with a darker pen and better light rather than with a better hand.

What is GPT Image 2.5 sketch to image, exactly?

It is a sketch to image generator running OpenAI's GPT Image 2.5. You hand it a drawing — a phone photo of paper, or something you draw in the browser — and it treats that drawing as the layout rather than as the artwork. Your prompt supplies the medium, the materials and the exact wording; the composition, the proportions and the position of every element come from the sketch. The first image is free and needs no account.

Will it keep my composition, or redraw it?

It keeps it, as long as the sketch is treated as a structure reference — which is on by default here. The exact line is published in the prompt section above so you can use it in any other tool too.

Will the text I wrote on the sketch come out right?

If you type the words into the prompt in quotes rather than relying on it reading your handwriting, yes. That is the thing this page exists for. The box you drew still tells the model where the headline goes; the prompt tells it what the headline says.

Can I photograph a paper sketch?

Yes, and most sketch to image runs here start exactly that way. Even lighting helps; hard shadows and glare don't. The pre-upload check runs in your browser and flags format, size, dimensions and contrast before you spend anything.

What can I upload, and how many sketches at once?

Up to four sketches in one run — useful when the layout is on one page and a detail is on another. PNG, JPG and WebP all work, and a straight-from-the-phone photo is fine. The browser-side check reads format, file size, dimensions and contrast before anything is uploaded or charged, so a photo that is too dark or too soft gets flagged rather than billed.

Can I draw here instead of uploading?

Yes. The canvas is the first thing on the page, and on a phone it's the main way in — pen, eraser, undo, clear, and a thickness slider. Nothing else.

Which GPT Image 2.5 model should I use for a sketch?

Fast while you are still finding the composition, Think for the one you keep — and Think plus HD for anything with small labels on it, which is the one place the tier reliably changes the result rather than only the price. Every result prints the model id it actually ran, and the difference between the two is laid out in Flare or Sunburst.

How is this different from ChatGPT's @Sketch?

@Sketch needs an account, and it doesn't let you set the model, the quality tier or the output size, and it doesn't tell you which model ran. This does all three. If you already have a ChatGPT account and none of those matter to you, @Sketch is good and it costs you nothing.

Isn't Krea better at this?

For drawing and watching the picture change live, yes — nothing here does that, and no amount of engineering on our side would, because a render takes seven to ten seconds. Come here when there are words in the sketch.

Is sketch to image free?

The first image is free, with no card and no account, and failed or moderation-blocked runs never cost anything. After that a Draft render is 3 credits and an HD one is 24. The daily limits, the resolution you get for free and what you are allowed to do with the file are all set out in what the free tier actually covers.

Can I get a transparent PNG for a logo draft?

Yes, on Starter and above: set Background to Transparent and keep the format on PNG or WebP — JPEG has no alpha channel, so the panel refuses that pair rather than handing you a flat file. Add "no checkerboard" to the prompt, because a model asked for transparency will sometimes paint the grey-and-white grid that represents it, which is a picture of transparency rather than transparency.On the free tier, ask for the mark on a flat pure-white field with no shadow and key that colour out; on a hard-edged logo the result is usually clean in one step.

Draw something bad and see what happens

Thirty seconds with a pen, or a photo of whatever's already on your desk. That is the whole of sketch to image, and you'll learn more from one render than from the rest of this page.

Draw in the browser or upload a photo
Free pre-upload check
Nothing overwritten
Failed runs cost nothing

gptimage25.top is an independent product with no affiliation with OpenAI. This is not an official OpenAI website. GPT Image, ChatGPT and OpenAI are trademarks of OpenAI. Alternatives on this page checked September 9, 2026.

Written and maintained by Andy SwiftPublished 9 September 2026Last updated 10 September 2026Tool version 2026.09.1

GPT Image 2.5 Sketch to Image AI Generator Draw it badly. That is the point.

What GPT Image 2.5 sketch to image does with your drawing

Say the sketch is a structure reference, not a style reference

Sketch to image, before and after: a drawing in, a photograph out

Sketch to image with text: the words you scribbled on it

Why edge detection loses your handwriting

Don't make the model read your handwriting. Type it.

Spell unusual names letter by letter

Small or dense labels? Run it at HD.

Drafting a logo? Ask for a flat white field.

How to write a sketch to image prompt that doesn't wander off

Name the sketch as the structural reference

Say what to render it as

Quote the text, and forbid the rest

Two images? Say which is which.

Logo drafts, wireframes, posters: what sketch to image is used for

Napkin sketch to logo

Wireframe to UI mockup

Poster layout to finished poster

Packaging rough to product shot

Floor plan or room layout to a render

Fashion croquis to a photographed look