AI Sign Mockup Test: GPT Image 2.5 Flare vs Sunburst

For an AI sign mockup on a photo of the client's building, use GPT Image 2.5 Flare. We gave Flare, its bigger sibling GPT Image 2.5 Sunburst and Nano Banana Pro the same five sign jobs on three storefront photos: channel letters by day and lit at night, a pylon sign, opening hours in window vinyl and a blade sign. Flare spelled every word and digit right in all five renders, kept the building and the camera, and took 26 to 34 seconds for 150 credits each. Sunburst was exactly as accurate for twice the credits and three times the wait, so the cheaper GPT model is enough. Nano Banana Pro is fine for big letters but made six errors in the small opening-hours text, one of them a wrong closing time.

Quick answer: Signage mockups: GPT Image 2.5 Flare (150 credits, 26-34 s), 4 of 5 jobs passed, 0 spelling errors in 5 renders. GPT Image 2.5 Sunburst (300 credits, 78-87 s): the same 4 of 5 and 0 errors, no gain for the price. Nano Banana Pro (150 credits, 18-26 s): 2 of 5, 6 spelling errors, all in small text. No model got a halo-lit night sign right with the camera kept.

All three ran in Visualizee (our tool), at 1K, once each, with the same words. This is a job test, not a launch note: for the same models on six general design jobs (a SketchUp model, staging, a kitchen, a car, a product, a sketch), see the best AI model for architectural rendering hub; for what AI rendering is and where it fits next to an engine, see AI rendering.

Why Sign Text Is the Hard Test for an AI Model

A sign mockup fails the moment one letter is wrong, so it tests the thing image models have been worst at: exact text. A client approving a fascia sign reads their own business name first, and a doubled letter or a dropped one ends the conversation, however good the brickwork looks.

Two things made GPT Image 2.5 the model to test. In our Oct 22 benchmark, both Flare and Sunburst printed a readable magazine title on a book that nobody asked for, so they clearly can write; the question was whether they write what you ask, letter for letter. And sign shops sell lettering by the inch: channel letters, a pylon panel, vinyl on glass, a projecting blade sign. Those four, plus the channel letters lit at night, are the five jobs here.

We judged each render on five things: spelling, counted letter by letter on a full-size crop; kerning and legibility at blog size; whether the sign sits on its surface in the right perspective; whether anything else on the building changed; and any brand text nobody asked for (a flag). The rules are the same as the hub's. Pass: the requested sign is right and nothing we listed to keep changed. Slip: a spelling error, a layout or lighting type other than the one we specified, or a kept element changed, such as the camera. A slightly different shade of the color we named, or a type style we left open, still counts as a pass.

The Inputs: Three Storefronts With Blank Sign Space

Each input is a street photo with the sign space empty: a dark green fascia on a brick high-street bakery, a light gray tyre and service unit with an empty pylon frame by the road, and a restaurant with a plain window and an empty wall bracket. All three are generated photos (Nano Banana 2.1, 18 to 21 seconds, 60 credits each), so there are no real addresses, no people and no real brands; we made them on a model outside the test so neither GPT model was re-rendering its own image.

The business names are invented and checked: a web search found no business called Wrenfield Bakehouse, NORDBY TYRES or Osteria Calma. The phone number is from the UK range reserved for fiction. With a real client you would use your own photo from the site visit: straight on or at a slight angle, the whole sign area in frame, no harsh shadows across it.

All 15 Renders at a Glance

Flare and Sunburst each passed four of five jobs with no spelling errors; both slipped on the same one, the night lighting. Nano Banana Pro passed two, and all six of its spelling errors are in the smallest text. No render added a brand nobody asked for, and none split into before-and-after panels, a habit Flare showed on renovation prompts in an earlier test.

Totals: Flare 750 credits, Sunburst 1,500, Nano Banana Pro 750, plus 180 for the three inputs, 3,180 credits in all. Times are measured on our server from request to finished image.

Channel Letters on a Fascia: All Three Spelled It Right

All three models put WRENFIELD BAKEHOUSE on the fascia as raised cream letters, letter for letter, centered and following the board's slight perspective. The shopfront, the sash windows and the neighboring blank fascias stayed as they were. This is the easy case: one line, big letters, a flat board.

What we typed: "Add a shop sign to the blank dark green fascia board above this shopfront: individual built-up channel letters (solid 3D letters about 10 cm deep, fixed to the board) in warm cream, in one line of capital letters in a clean serif typeface, centred on the fascia, filling about two thirds of its width and following the board's perspective. The letters spell exactly: WRENFIELD BAKEHOUSE. Letter for letter, nothing else written. Keep exactly: the camera and framing, the brick facade, both upper sash windows, the green shopfront with its display windows and door, the neighbouring shops with their blank fascias, the street and the daylight. Change only the fascia, by adding the letters. Don't add any other text, logo or sign anywhere."

Flare and Nano Banana Pro gave the cleanest letters; Sunburst's came out a shade more gold than cream. The image at the top of this post is Flare's. The two GPT images are 1360 pixels wide against the input's 1376, so they look a hair tighter, but the camera did not move: they line up with the input at the same position.

The Same Sign at Night: Nobody Got the Halo Right

The night job is where all three slipped, each in a different way, and none on spelling. We gave every model the same accepted day render (Nano Banana Pro's, which lined up with the photo most tightly) and asked for halo lighting: LEDs behind each letter throwing a glow onto the board while the letter faces stay solid. That is how many shops light raised letters, and the client is choosing it, so the render has to show it.

What we typed: "Show this same shop at night, about an hour after dusk. The channel letters WRENFIELD BAKEHOUSE on the fascia are halo-lit: LEDs behind each letter throw a soft warm white glow onto the green fascia around it, while the cream letter faces stay solid. Warm light inside the shop makes the display windows glow; the street is lit softly by street lamps out of frame; the sky above the roof is deep blue. Keep exactly: the camera and framing, every letter of WRENFIELD BAKEHOUSE as it is (same typeface, size, spacing and position, letter for letter), the fascia, the brick facade, both upper sash windows, the shopfront, display windows and door, the neighbouring shops with their blank fascias, and the street. Change only the time of day and the lighting. Don't add any other text, logo or sign anywhere."

Flare and Sunburst made the more convincing evening scenes, with warm display windows, a filled bakery behind the glass and lamp light on the road, but both lit the letter faces themselves, so the sign reads as front-lit letters, a different product. Nano Banana Pro drew the halo exactly as described, beige faces with a glow around each letter, but moved the camera down far enough to show the roof and chimney, and left the shop windows unlit. For a client, Flare's or Sunburst's image is the better picture; it just shows a different sign from the one you would quote for halo-lit letters.

Pylon Sign With a Phone Number: Right Words, One Layout Miss

All three wrote NORDBY TYRES, "Tyres - Brakes - Servicing" and 01632 960418 with no errors and filled only the pylon panel, leaving the building's cladding band blank as asked. The differences are layout.

What we typed: "Fill the empty white panel of the tall freestanding pylon sign on the right with printed sign graphics, flat on the panel and inside its frame: a dark navy background; at the top, the business name in large bold white capital letters on one line: NORDBY TYRES; below it, two smaller lines of white text, centred. Line 1: Tyres - Brakes - Servicing. Line 2: 01632 960418. Every word and digit exactly as written, letter for letter, nothing else written. Keep exactly: the camera and framing, the pylon's posts, frame and size, the metal-clad building with its roller-shutter door, glazed entrance and blank cladding band, the car park and its bay lines, the hedge, the trees, the sky and the light. Change only the pylon panel. Don't add any other text, logo or sign anywhere, and leave the building's band blank."

Flare gave the most even layout. Sunburst condensed the name so hard that the space between NORDBY and TYRES almost disappears at blog size; spelling is right, but a sign designer would open the gap. Nano Banana Pro broke the name onto two lines although we asked for one, which changes the panel's whole proportion, so it counts as a slip.

Opening Hours in Window Vinyl: Where Small Text Breaks

Small text is where the models separate. Five lines on the glass, the hours set at about 10 pixels tall in a 1K image: Flare and Sunburst got every letter, digit and colon right, left-aligned as asked. Nano Banana Pro made six errors: "Tos" for Tue (two letters), "Frl" for Fri, 14:00 instead of 16:00 as the Sunday closing time, and "Men clesed" for Mon closed (two letters). It also right-aligned the block, which is a layout miss, not a spelling error.

What we typed: "Apply white cut-vinyl lettering to the large plain front window on the left, placed in the lower right part of the glass at about eye height, small as on a real restaurant window, flat on the glass and following its perspective. Five left-aligned lines. Line 1, slightly larger: Osteria Calma. Then the opening hours, four lines: Tue - Thu 12:00 - 22:00 / Fri - Sat 12:00 - 23:00 / Sun 12:00 - 16:00 / Mon closed. Every word and digit exactly as written, letter for letter, nothing else written. Keep exactly: the camera and framing, the cream building, the window frame and the interior visible through the glass, the wooden entrance door, the terracotta awning (blank), the empty black wall bracket, both potted olive trees, the upper floors, the street and the evening light. Change only the window, by adding this lettering. Don't add any other text, logo or sign anywhere."

This is the job that decides the model for signage. Big letters on a fascia and a phone number on a pylon were easy for all three; small lines like opening hours are where a client spots the mistake, and only the GPT Image 2.5 models got them right. A wrong closing time is worse than a typo: it reads fine, so nobody catches it until a customer arrives at a locked door.

Blade Sign: All Three Pass, With a Perspective Caveat

All three hung a dark olive board on two chains from the existing bracket and painted OSTERIA over Calma with no errors; the window stayed clear and the awning blank. Sunburst set "Calma" in italics, which we had not specified. Nano Banana Pro's board is the lightest green.

What we typed: "Hang a double-sided blade sign from the existing small black wall bracket beside the door: a rectangular painted wooden board, about 60 cm wide and 45 cm tall, dark olive green with a thin cream border, hanging below the bracket arm on two short chains, perpendicular to the wall so it is seen at an angle from the street, in the bracket's perspective. Painted cream text, centred on two lines. Line 1: OSTERIA. Line 2: Calma. Every letter exactly as written, nothing else written. Keep exactly: the camera and framing, the cream building, the bracket where it is, the big front window plain and clear with no lettering, the interior through the glass, the wooden entrance door, the terracotta awning (blank), both potted olive trees, the upper floors, the street and the evening light. Change only the bracket, by hanging this sign from it. Don't add any other text, logo or sign anywhere."

The caveat: all three drew the board almost square-on to the camera, with little foreshortening, although a board hanging from an arm that sticks straight out of the wall should be seen more from the edge from this angle. At blog size it reads naturally; a sign maker will notice. We counted it as a pass because the sign hangs from the right bracket and reads correctly, but it is not a reference for how the sign will look from that spot.

How Sign Shops Use a Mockup for Client Approval

A sign shop's mockup has two readers: the client, who approves the look, and the permit office, which approves the size and position. An AI render covers the first in under a minute; the second still needs dimensions.

For the client, the render on their own building replaces the flat artwork proof that most people cannot picture on the fascia. Show it with the exact wording you will fabricate, and ask them to check the spelling, the phone number and the hours line by line before they sign off.

For the permit, the requirements are explicit about measurements. Santa Barbara's sign permit application packet asks for scaled building elevations or, instead, "dimensioned photo-simulations showing the correct proportions" of the sign to the building, plus enlarged drawings with letter heights and type faces and a color rendering of every sign. Snohomish's design review application requirements (January 2026) ask for a scaled color rendering showing lighting and the style and spacing of letters, and scaled elevations or a photo-simulation with dimensions and ground clearance. A render made with AI is a good base for that photo-simulation, but the sizes, letter heights and clearances have to be added by you, from your own drawings.

What It Can't Do

The render shows the client a sign on their building; it does not specify, engineer or permit one. Plan around these limits.

• It won't match an exact font. You can ask for "a clean serif" or "bold condensed capitals", and the model draws something close. It cannot set your client's brand typeface or logo. For a brand mark, composite the real artwork onto the render instead of asking the model to redraw it.
• No real fixings, sizes or permit drawings. The models have no idea of letter returns, stand-off fixings, cable runs, wind load or how deep a pylon foundation goes. Nothing in the render is to scale; the permit office needs your dimensioned drawing.
• Small text has a floor. At 1K, text about 10 pixels tall held on both GPT Image 2.5 models and broke on Nano Banana Pro. Anything smaller, such as a web address under the hours or legal text on a pylon, is not safe on any model; try 2K (untested here) or add it afterwards.
• Lighting types get swapped. Asked for halo-lit letters, two models lit the letter faces instead. If the client is choosing between face-lit, halo-lit and unlit letters, check that the render shows the one you are quoting.
• Perspective on projecting signs is loose. Blade signs came out nearly square-on to the camera. Fine for "this is what it looks like"; not for "this is what you will see from the bus stop".
• Spelling still needs a check. Ten of ten GPT renders were letter-perfect here, but that is one test. Read every letter and digit against the approved wording before the client sees it.

How to Make a Sign Mockup on a Client's Photo

Start from one photo, one sign per request, and the exact words in the request. On Flare it costs 150 credits and about half a minute a render.

1. Take the photo at the site visit, the whole sign area in frame, straight on or at a slight angle, no strong shadow across the fascia.
2. Write the wording exactly in the request, line by line, and add "letter for letter, nothing else written".
3. Name the sign type and finish: built-up channel letters 10 cm deep, flat printed panel, cut vinyl on glass, painted hanging board.
4. List what stays: camera and framing, facade, windows, door, neighbors, other sign spaces left blank.
5. Check the render: every letter and digit, the alignment you asked for, the lighting type, and that the camera did not move.
6. Make the night version from the approved day render, not from the bare photo, so the lettering is the one the client already approved.

The window, door and garage door visualizer test uses the same method for the rest of the facade, and found that counts (panes, panels) drift back to the usual pattern; spelling, here, did not.

Frequently Asked Questions

Which AI model is best for storefront sign mockups?

GPT Image 2.5 Flare. In our test of five sign jobs it spelled every word and number letter for letter (0 errors in 5 renders), kept the building, camera and windows, and took 26 to 34 seconds at 150 credits. GPT Image 2.5 Sunburst was just as accurate but cost 300 credits and took 78 to 87 seconds. Nano Banana Pro, at the same price as Flare, made 6 spelling errors in the small opening-hours text.

Can GPT Image 2.5 spell sign text correctly?

Yes, in our test. Both GPT Image 2.5 Flare and Sunburst wrote WRENFIELD BAKEHOUSE, NORDBY TYRES with a sub-line and a phone number, five lines of opening hours on a window and a two-line blade sign with no spelling errors across ten renders. The smallest text was about 10 pixels tall in a 1K render and still read correctly. Check every letter anyway: one test is not a guarantee.

Is the more expensive GPT Image 2.5 Sunburst worth it for signs?

Not for checking a sign layout. Sunburst passed the same four of five jobs as Flare with the same zero spelling errors, at twice the credits (300 vs 150) and about three times the wait (78 to 87 s vs 26 to 34 s). It also packed the words NORDBY TYRES so close the gap nearly closed. Use Flare: in this test Sunburst added nothing you could see on the sign.

Can AI show an illuminated sign at night?

Partly. All three made a night version with the lettering spelled right, and Flare's and Sunburst's are convincing evening scenes, but none got halo lighting right while keeping the camera. Flare and Sunburst lit the letter faces instead of the glow behind them, which is a different product to quote; Nano Banana Pro drew a true halo glow but moved the camera and left the shop windows dark. Name the lighting type and check it before the client sees it.

Can I use an AI sign mockup for a planning or sign permit application?

Only as a starting picture. Permit offices ask for scaled or dimensioned drawings: Santa Barbara accepts dimensioned photo-simulations in place of elevations and asks for letter heights and a color rendering, and Snohomish asks for a scaled color rendering showing lighting and letter spacing. An AI render has no dimensions, so add sizes, letter heights and fixings in your drawing software before you submit.

The Short Version

For sign mockups, GPT Image 2.5 Flare is the model: zero spelling errors in five jobs, the building kept, half a minute and 150 credits a render. Sunburst matched it at twice the price, and Nano Banana Pro is fine for big letters but not for small text such as opening hours. Check every letter, check the lighting type at night, and add the dimensions yourself before anything goes to the permit office.