The sign is misspelled. Fix it without re-rolling
Do not re-roll a good image for one misspelled word. Quote the exact string, describe the font, set a mask margin, and know when to composite type instead.
The frame is the one you would have kept. Light, composition, the product, the face. Then you read the shop sign and it says GRAMD OPNEING. Re-rolling that image is the expensive mistake. You are throwing away a picture that is almost entirely correct for a fresh draw on the whole distribution, and the next one is not more likely to spell the word and keep the rest.
Text in a generated image is not spelled. It is rendered as a visual texture. Training images contain lettering as a tiny pixel share, so the model learns local letterforms and not word assembly. A wrong string is a small-region failure. Inpaint the word. Leave the picture.
The exception is anything that has to be pixel-perfect: a logo, a legal line, more than a few words. That job is a composite of real type, not a luckier sample.
Decide whether the word is even repairable
Correctness is discrete. A hand can be plausible. A word is either the right string of characters or it is a mistake. That is why a "close" inpaint is not close enough, and it is why you should not spend the repair budget on lettering that was never supposed to be read.
Three filters, in order:
Size. If the letters are a background shopfront forty pixels tall, you will get a different wrong word. Soften that sign into the depth of field instead of repairing it; blurring garbled signage is the right finish for incidental type. Hero type (a poster headline, a carton word, a storefront the shot is about) is the inpaint case.
Length. One to three words is the band where a quoted-string inpaint has a real chance. A sentence, an ingredients list, a paragraph of body copy is not. As of mid-2026 no model spells reliably across long copy, small type, and a busy scene at once. That is a workflow limit, not a checkpoint you are waiting on.
Stakes. Decorative type nobody can check can be "good enough." Your brand name, a price, a claim, a legal line cannot. If a wrong letter is a client escalation, do not ask the sampler to draw it. Composite real type.
If you are still generating rather than repairing, quote the string, make it large and high-contrast, and use a text-specialist for the first pass. Ideogram V4 is built around accurate text rendering. The Seedream and Ideogram comparison is the generation-side version of this problem. Once the rest of a good frame is already in hand, you are on the repair side.
The text-region inpaint, step by step
- Crop to the lettering plus a band of the surface it sits on. A word that was two percent of the frame needs to be most of the edit canvas. You are changing the pixel-share problem from the cause of the failure into something the model will actually spend budget on.
- Mask the glyphs plus a margin. Cover the wrong letters, a little of the surrounding plate (wood, paint, cardboard, glass), and any shadow the letters themselves are casting. A tight crop that stops on the ink leaves a halo, or a correct word sitting in the ghost of the wrong one. A working starting margin is in the 15 to 30 pixel band around the lettering, feathered. Go wider on a textured surface (brick, corrugated carton) so the join happens inside texture rather than on a flat colour.
- Quote the exact string. Not a paraphrase, not "fix the spelling," not the word in lowercase if the picture is in caps. The prompt contains the characters you want drawn, in the case you want them drawn.
- Describe the font as a physical object. Weight, serif or sans, colour, material (painted, vinyl, neon, embossed, printed). The model is matching a texture that already exists around the mask. Naming that texture is what stops you getting a clean Helvetica sticker on a hand-painted wall.
- Set denoising strength in the middle and adjust from what comes back. Too low and GRAMD survives under a slightly different texture. Too high and you get a new sign in a new typeface on a wall that no longer matches. Keep the seed fixed while you change wording.
A prompt that usually lands:
The painted wooden sign reads "GRAND OPENING" in condensed sans-serif capitals, cream paint on dark green timber, slight wear on the edges of the letters, matching the existing sign board. Exact spelling: GRAND OPENING.
What that prompt is doing: the quoted string, a type description, a material, and an instruction to match the board that is already in the picture. "Fix the text" is the weak version. It does not name the characters, so the model samples another plausible sign-shaped texture.
Qwen Image Edit is the editor to try first when the job is changing or correcting text inside an image, a task most editors smear into noise. Drive the same loop from the photo editor. The mask is the parameter that stops the rest of the frame moving.
Two or three attempts. If the third is still a different wrong word, the region is under-determined (too small, too low-contrast, or the surrounding plate does not give the model a type style to match). Stop inpainting and composite.
When to composite real type instead
Inpaint is for a short, already-styled string on a surface the model can see. Composite is for everything that has to survive a proofread.
| Job | Tool | Why |
|---|---|---|
| One to three hero words, large, high-contrast | Quoted-string inpaint | The model has enough pixels and a short string |
| Unknown decorative type in the background | Depth-of-field blur, not a repair | There is no correct answer worth hitting |
| Logo, wordmark, lockup | Real artwork, composited | A logo is a drawing, not a spelling |
| Price, claim, legal line, ingredients | Real type, composited | One wrong character is a fail |
| More than a few words, any size | Real type, composited | Failure rate compounds per glyph |
The composite sequence is the same one you would use for a label swap, because it is a label swap with a smaller mask:
- Inpaint the wrong lettering out, so you have a clean plate of the surface (the wall, the carton, the window) with correct light and texture.
- Set the real string in a design tool, in the typeface the brief actually uses.
- Place it, warp it to the surface if the board recedes, match the grade.
- If the join shows, a light inpaint over the edges only, not over the glyphs.
Yes, that is four steps. It is also the difference between a poster you can ship and a poster a client reads once. Do not run a fourth inpaint hoping the sampler will invent your brand typeface. It will invent a neighbour of it, confidently.
If you do not know the original font, do not guess a foundry name. Describe what you can see: "condensed sans, high contrast, slightly rounded terminals, painted cream on dark timber." If only one word on a longer sign is wrong, leave the neighbouring words unmasked so they become the style reference.
What not to do
Do not re-roll the image. The composition you liked is not conserved across samples. Treat the misspelling as a local sampling failure.
Do not write "no misspelling" or "correct English" as a negative. Naming the defect does not suppress it. Quote the string you want.
Do not inpaint a word that should have been an overlay. If the type is a headline for the audience, not a thing in the scene, generate a clean plate and set the type on a layer.
Do not proof at thumbnail size. Read the string at 100 percent, out loud, against the brief, before you keep the file.
FAQ
How much mask margin do the letters need?
Enough that the join happens in the surface, not on the ink. Start with 15 to 30 pixels around the lettering, feathered, and include any shadow the letters cast. Too tight and you get a correct word in a halo of the old one. Too loose and the model rebuilds half the board.
What if I do not know the font?
Describe the letters as objects: weight, case, serif or sans, colour, material, wear. If the type is a brand font you actually have, composite it. Inpaint will not reconstruct a licensed face from a sentence.
When is compositing faster than another inpaint?
As soon as you care about the exact letters. The second failed inpaint is the switch. Logos, prices, claims, and anything longer than a few words should skip inpaint entirely.
Can I just generate the whole image in a text-specialist model instead?
Yes, if you do not yet have a frame you like. That is the wrong move once the rest of the picture is the keeper. You would be re-rolling lighting and layout to chase one word. Inpaint the word.