For small businesses and designers making a flyer, menu, poster or ad with Depicta, through Claude, the CLI or the API.
Print needs an exact page size, enough pixels, safe margins, readable text and CMYK colour. Image models know none of this. You tell them some of it in the prompt, and handle the rest in your layout or print tool.
Start with the look: layout, typography, icons and photo, all in one image.
Use a model that takes reference images: gemini, or, named only, gpt-image for text-heavy layouts or gpt-image-pro for photo-led designs (OpenAI keeps inputs for up to 30 days, or longer where the law requires; avoid for photos of people). The example, a photo-led ad, uses gpt-image-pro.
Choose the aspect ratio closest to your page. You crop to the exact size in step 3.
Give each reference image one job, and name it: “layout and style”, “the scene photo”, “the exact logo” (a clean PNG).
Write out every word exactly, in quotes, and end with “No other text”.
Decide the text sizes now. Later passes copy the design as it is, small text included.
Ask for a safe area: “keep all text and logos at least 8% away from every edge”.
Terminal window
npxdepicta@latestimage-mgpt-image-pro-s4:3\
-rlayout.png:"layout, typography and icons"\
-rscene.jpg:"the scene and light"\
-rlogo.png:"the exact logo"\
'Print advertisement for my shop … headline "…"; contacts "…". No other text.'
The OpenAI models don’t offer a size choice, so check the output size. It’s usually too small for print, and step 2 fixes that.
For print you need about 300 pixels per inch at the final size. An A5 page in landscape (210 × 148 mm), for example, needs at least 2480 × 1748 pixels. More helps small text.
gemini is the model with 4K output. Pass the design from step 1 as a reference and ask for a faithful high-resolution copy. Quote all the text again.
-rdesign.png:"the design to recreate exactly: layout, typography, text, icons, photo"\
'Recreate this design faithfully as a high-resolution print master. Keep all text at least 8% from every edge. Text, exactly as written: "…". No other text.'
Enlarging a design you already have. An edit with gemini at 4K can make a larger copy of a finished design. Ask for the exact image at a higher resolution, with identical layout and colours. List every piece of text word for word, and end with “No other text”.
Crop to the exact page ratio, taking the cut from an edge that holds only background or decoration.
process convert takes -f jpeg, not jpg. For print, set --quality 95 (80 to 100, JPEG and WebP; needs CLI 1.4.0 or later). Without it, the server’s default applies.
Keep PNG if your layout tool accepts it. It is lossless, so nothing is lost before the printer’s own conversion.
Don’t sharpen or retouch with filters. They add halos that show in print.
AI-rendered text is pixels. Small, light text on a dark background suffers most on press: the inks never line up perfectly, and thin letters fill in.
Measure it. At final size, capital letters under about 2 mm tall (roughly 7–8 pt) are risky, especially light on dark or over a photo.
Set the text a reader must get right (phone, email, address, prices) as real text in your layout tool: 8–9 pt or more, regular weight, on a plain band or a calm area. Take the band’s colour from the image so the two match after conversion.
Do these steps in your layout or print tool, or ask your printer.
Page size. Make the document exactly the size the printer asked for.
Bleed. A design that runs to the trimmed edge may need extra image past it, often 3 mm per side. A placed box ad usually doesn’t. Ask if unsure.
Safe margins. Keep text at least 5 mm inside the trim.
Minimum type size. About 7 pt for dark text on white. About 8–9 pt, not a light weight, for light text on dark.
CMYK. Presses use four inks, not screen RGB. Convert with your printer’s profile, often FOGRA39 for coated paper in Europe. Give brand colours exact CMYK values. Metallic colours like gold can’t be printed in CMYK, so pick the nearest flat colour.
Total ink. Dark areas shouldn’t go over the printer’s ink limit, often around 300–330% on coated paper.
PDF/X. A print-safe PDF standard. Many printers accept PDF/X-1a (CMYK only, no transparency, fonts embedded or outlined). Some ask for PDF/X-4.
Soft-proof. Check the final PDF itself, not your layout preview. Export tools sometimes drop effects without warning.
Asking a model to enlarge small text later. Once a finished design is the input or the reference, the model copies the text sizes as they are, and a re-run may add a typo. Set the sizes in step 1, or set small text as real type in step 4.
Full-scene references for one detail. A complete scene given as a reference tends to be copied as a template, unwanted details included. Crop each reference to its job (the product, the background, the logo) and name that job.
Every word, number, email and web address checked at full size
Page size exact; bleed added if the printer asked for it
All text at least 5 mm inside the trim
Small important text at 8–9 pt or more, regular weight, or set as real type
Images at 300 ppi or more at final size
CMYK with the printer’s profile; brand colours checked
Total ink within the printer’s limit
The PDF/X version your printer accepts; the final PDF soft-proofed
If an image could pass as a real photo of real people, places, things or events, say it’s AI-generated (EU AI Act, Article 50(4)). Content Credentials don’t survive print, so add a short visible note, for example “Image created with AI”. See Section 8.2 of the Terms.
