Something changed in our showroom over the last year. Couples used to arrive with a photo of a cousin's card, or a page torn from a catalogue. Now they arrive with a phone full of AI-generated designs — peacocks in gold on ivory, elephants flanking a temple arch, mandala borders that would have taken an illustrator a week.
And here is the honest part, from a press that has been printing kankotri in Raopura since 1995: the designs are good. Not "good for AI" — genuinely good. The 2026 generation of image models produces ornamental work at a standard that would have cost serious money to commission three years ago.
So this guide is not a warning about AI. It is the opposite. It is how to get a design out of AI that we can actually print, and what has to happen between the image on your phone and five hundred cards in your hands.
What AI is genuinely good at now
It is worth being precise, because the situation in 2026 is very different from the one most articles describe.
Text rendering is largely solved. OpenAI's GPT Image models reach roughly 99% accuracy on English text and above 90% on Hindi, Bengali and Arabic. Google's Nano Banana Pro, built on Gemini 3 Pro, is explicitly built for text-heavy layouts and infographics. The old joke about AI writing "MARRIAJE" is out of date.
Resolution is no longer the bottleneck. A standard 5×7 inch card at print quality needs about 1500 × 2100 pixels. Nano Banana Pro outputs up to 4K. Canva's AI tops out around 2048 × 2048 on Pro, which is adequate for a card-sized piece. All three clear the bar.
Ornament and symmetry have improved sharply. Mandala work, jaali screens, paisley borders and temple arches — the motifs that dominate Indian invitation design — are exactly the kind of repeating geometric detail these models now handle well.
So what is left? Quite a lot, but none of it is about the picture.
The gap nobody mentions: an image is not a print file
This is the whole point of the article, so it is worth stating plainly.
An AI model produces a flat RGB raster image designed to look correct on a glowing screen. A printing press needs a layered CMYK vector file with physical tolerances built in. Those are not the same object, and no amount of prompting closes the distance. Here is what actually has to change.
1. Bleed and trim
Paper shifts by a fraction of a millimetre as it moves through a press, and guillotines are not perfect. Print files carry roughly 3mm of extra artwork beyond the finished edge — the bleed — so a slight drift trims into spare background instead of leaving a white sliver. AI images end exactly at their edge. The background has to be extended outward before anything is cut.
2. RGB to CMYK
Your screen mixes light; a press mixes ink. Converting between them moves colours, and it moves the colours Indian wedding cards depend on most. Deep maroons flatten. Vivid magentas and oranges — the marigold and kumkum range — lose their glow, because those saturated tones sit outside what CMYK ink can reproduce. Every AI "gold" becomes a muddy yellow-brown, because gold is not a colour in CMYK at all.
That last point leads directly to the biggest one.
3. Foil, emboss and laser cut are not colours — they are processes
When an AI paints a gold border, it is painting a picture of light reflecting off gold. It has rendered a highlight. The press cannot print a highlight — it has to physically stamp metallic foil onto the paper.
To do that we need a separate layer showing exactly which shapes get foil, as solid vector paths. Same for embossing, which needs its own raised-area map, and for laser cutting, which needs a closed cut path with enough material left that the card does not fall apart. In an AI image all three are baked into the same flat picture, indistinguishable from the artwork around them.
This is the single most common surprise for couples, and it is worth understanding before you fall in love with a design: the shimmer in your AI image is free, and the real version of it is a physical process on a real machine.
4. Raster text versus live type
Even when a model spells everything correctly, it produces text made of pixels. Printed type made of pixels has soft, slightly fuzzy edges, especially at the small sizes used for venue addresses and muhurat lines. Properly typeset text is vector — mathematically sharp at any size.
So text gets rebuilt regardless of whether the AI spelled it right.
5. Gujarati, and why "usually correct" is not the standard
Indic scripts render far better than they did. But a kankotri is not a poster. It carries family names, a shubh muhurat, and the order in which elders are listed. A single misplaced matra is printed hundreds of times, handed to relatives, and cannot be recalled.
At 95% accuracy, roughly one line in twenty is wrong. For a name, that is not a risk worth taking for a convenience you do not need — because the text is being re-typeset as vector anyway.
Which brings us to the most useful thing in this article.
The technique that solves most of this: prompt for the art, not the words
Ask the model to generate the artwork only, and to leave a clean empty panel where the wording belongs.
You get better results in every direction. The model stops spending effort on typography and puts it into ornament. You avoid script errors entirely. And you hand us a file where the text area is already clear, which is exactly what we need to typeset into.
Add a line like this to any prompt below:
Leave the central panel completely empty and clean — no text, no lettering, no placeholder words. Only decorative border and background.
Working prompts
These are written as full briefs. The models reward detail, so do not trim them down. Replace the bracketed parts.
The reliable base prompt
A vertical Indian wedding invitation card design, 5:7 aspect ratio. Ornate [MOTIF] border framing a large empty centre panel. Colour palette: [COLOUR] with antique gold line work. Background: subtle textured ivory handmade paper. Fine, symmetrical, hand-illustrated linework in a traditional Indian miniature style. Leave the central panel completely empty and clean — no text, no lettering, no placeholder words. Flat front-on view, even lighting, no perspective, no shadows, no mockup, no hands, no table.
Why each part matters: 5:7 matches standard card proportions, so you are not cropping later. Flat front-on, no mockup is the most important instruction in the whole prompt — models love to render your card lying on silk next to flowers, which is useless as a print file. No shadows stops the model baking in lighting that will print as grey smudges.
For motifs, the styles moving fastest right now are peacock, elephant pairs, mandala, temple arch, lotus and marigold — but the prompt works with anything.
Adapting it to your tradition
Add one of these lines. The models know these references well.
- Gujarati kankotri — Traditional Gujarati kankotri style, Ganesh motif at the top centre, bandhani-inspired pattern in the border, red and gold palette.
- Navratri or festival season — Garba silhouettes and dandiya motifs worked into the border, nine-colour accent palette, festive marigold garlands along the top edge.
- South Indian — Temple gopuram architecture, banana leaf and jasmine motifs, kanjeevaram silk border pattern, deep maroon and gold.
- Rajasthani or Marwari — Jharokha window arch, mirror-work inspired detailing, royal blue and gold, haveli fresco style.
- Minimal and modern — Single-line botanical illustration, generous white space, one thin gold rule, muted sage and cream, contemporary editorial style.
Tool-specific notes
ChatGPT handles the longest and most nuanced briefs. Give it the full paragraph, then refine conversationally — "make the border thinner", "warmer gold" — and it holds the rest of the design steady while it changes one thing.
Gemini (Nano Banana Pro) is the one to use when you have a reference. It accepts up to 14 images at once, so you can supply a card you liked, a colour swatch and a motif, and ask it to combine them. Its 4K output also gives us the most headroom. Ask explicitly for 4K and 2:3 or 5:7 aspect.
Canva AI is the practical choice if you want to keep editing. Magic Layers splits generated artwork into separate editable layers, so you can move the border away from the text panel yourself. Canva can also export a PDF with CMYK and crop marks — the closest any consumer tool gets to a real print file, though foil and emboss separations still are not possible.
Two prompts worth avoiding
Do not ask for a mockup. "Wedding card on a silk cloth with rose petals and soft window light" produces a beautiful photograph of a card that does not exist. It is unusable for printing.
Do not name a living designer or studio. Beyond the ethics, it tends to produce a diluted imitation rather than a good original, and it introduces a copyright question you do not need on your own wedding invitation.
Bring us the image — that is genuinely the easiest path
Send the design on WhatsApp or bring it to the Raopura showroom. What happens then:
- An honest read. We will tell you which parts are printable as-is, which are lighting effects the model invented, and what the real equivalent costs in foil, emboss or laser cut.
- A material conversation. AI has no opinion about paper. Whether your design belongs on textured handmade stock, matte art card or a pearlised finish changes the result more than any prompt does — worth reading our guide to kankotri printing in Vadodara before you decide.
- Rebuild as a press file. Motifs redrawn as vector where needed, bleed added, CMYK conversion checked against real ink, foil and die separated onto their own layers.
- Gujarati typeset properly. Names, muhurat and family order set as live text, in the right order, proofed by someone who reads the language.
- A physical proof. You see it on the actual paper before we print the run.
If you are still choosing a direction, our guides to fonts and colour combinations and building a colour palette pair well with a prompting session — they will make your prompts noticeably better. If you are new to the format itself, start with what a kankotri actually is.
The short version
AI has become a genuinely good design partner. It has not become a printing press, and the distance between the two is not going to close — bleed, ink, foil and paper are physical problems, not software ones.
Use AI for what it is now excellent at: exploring directions quickly, and arriving at a shop knowing exactly what you want. That is a real advantage over flipping through catalogues, and we would rather you came in with a clear picture than a vague one.
Then let a press do the part that happens on paper.
Bring your AI design to Jalaram Cards, Raopura, Vadodara — or send it on WhatsApp. Tell us your guest count and the finishes you are drawn to, and we will tell you exactly what it takes to put it in your hands. Get in touch or browse our custom wedding invitations.