The short version: for a WhatsApp status or a Story, use the diya close portrait at 9:16. For the feed, use the saree or kurta portrait at 4:5. If you are sending wishes rather than posting yourself, the card prompt is the one you want, and it should go out on the 6th or 7th, not the 8th.
Diwali 2026 falls on Sunday 8 November, and the five days run from Dhanteras on Friday 6 November to Bhai Dooj on Tuesday 10 November. The search for a prompt always starts about three weeks before, and it always starts with the wrong question — people look for the Diwali prompt, when there are really five or six different photographs being asked for and they need different things.
A portrait lit by a single oil lamp is not the same job as a full-length shot of you in a kurta at the door, and neither is a card you send to forty people on WhatsApp. The six below are written for those different jobs. All of them keep your own face, all of them are free, and all of them run in Google Gemini, ChatGPT or Nano Banana on a phone.
Timing matters more than the prompt
A festival photo has a short shelf life. A Diwali portrait posted on the 6th reads as anticipation, on the 8th it reads as celebration, and by the 12th it reads as something you forgot to post. If you are sending a card rather than posting a portrait, go earlier still — a wishes card that lands on the morning of the 8th is competing with every other message that day.
So work backwards. Decide which two or three photos you actually need, generate them in the last week of October, and keep them ready. Generating on the night itself is how people end up posting something with six fingers on it.
The one rule all six share
Identity line
Preserve the exact face, skin tone, age, facial structure, eyes, eyebrows, nose, lips, jawline, cheeks, body proportions and natural recognisable appearance. Do not lighten, smooth or slim the face.
Every prompt below opens with that line, and it is doing two jobs. The first is obvious: it is what stops the model handing you a stranger. The second is that it explicitly bans the thing these tools do by default on Indian faces — lighten the skin, erase texture, narrow the jaw. If you drop that clause to shorten a prompt, that is the one you will miss.
The six Diwali prompts
Diya close portrait
The signature Diwali shot. Warm lamp light, 9:16, made for a Story.
Saree portrait
Standing, full ethnic wear, lamps behind. The feed photo.
Kurta portrait
The men's version: kurta, Nehru jacket, warm doorway light.
Couple portrait
Two faces to hold, which is twice as hard. Written to manage that.
Rangoli doorway
Full length at the threshold. The one that shows the house.
Wishes card
Not a portrait. A 4:5 card to send, with room for your own text.
1. Diya close portrait
This is the photograph most people mean when they go looking for a Diwali prompt: a close, warm, lamp-lit face with the rest of the frame falling into darkness. It is the strongest of the six because the lighting does all the work — a single warm source low in the frame, everything else unlit.
It is also the hardest for the model, for exactly the same reason. Light coming from below is rare in the training data, so this is the prompt most likely to drift off your face. Upload a bright, front-on reference photo rather than a moody one; you are asking the model to relight a face it can already see clearly, not to invent one in the dark.
The prompt puts the lamps on a surface in front of you rather than in your hands, which avoids the fingers-near-flame problem entirely. Shoot it at 9:16 for a Story or a status.
Best at 9:16 for a Story or WhatsApp status. For the feed, change the last line to 4:5.
2. Saree portrait
The most-searched variant by a distance, and the one where the identity clause matters most, because draped fabric and heavy jewellery give the model a lot of excuses to rebuild the person wearing them.
Two things this prompt does deliberately. It asks for the lamps behind the subject rather than in front, so the face is lit by a softer fill and the flames become background bokeh — much more forgiving than the close portrait. And it names the fabric and the drape rather than leaving it open, because "saree" alone will get you something that looks like a costume more often than not.
The last line matters. Models add religious and marital markers on their own, and they get them wrong.
3. Kurta portrait
The men's equivalent, and the one this site has the most to say about. The giveaway in a bad AI kurta photo is almost never the face — it is the clothes. Models produce a kurta that fits nobody: no shoulder seam anywhere near the shoulder, a hem that stops at a random point, sleeves that end mid-forearm as if by accident.
So this prompt specifies the fit rather than just the garment. If you want to understand why those three things are what give it away, the fit rules cover the same ground for real clothes, and the dress codes guide has what actually works for a festive evening.
If the jacket comes out looking like a waistcoat, remove the Nehru jacket line and run it again. One garment is easier than two.
4. Couple portrait
Two faces is more than twice as hard as one. Every model has a tendency to average two uploaded people towards one another, and the usual result is a couple who look oddly like siblings.
The prompt below manages that in the only way that reliably works: it tells the model which reference is which, and it separates the two people in the frame so there is less blending. Upload both photos, and say which is which in the chat before you paste the prompt. Expect to run it more than once — this is the one where a second and third attempt genuinely pay.
If the two faces still drift together, generate each person separately at 4:5 and post them as a two-image carousel instead.
5. Rangoli doorway
The wide one. Full length at the threshold, rangoli on the floor, the house behind you — the photograph that is as much about the home as the person.
Rangoli is where this prompt earns its keep. Left to itself, a model will produce a rangoli that is subtly wrong in a way everyone from India notices immediately: lopsided, asymmetric, a mess of petals with no repeating structure. The prompt asks for radial symmetry and a limited palette, which is what a real rangoli has, and it asks for it to be partly out of frame so there is less of it to get wrong.
Four colours is the instruction doing the work. Ask for more and the pattern stops reading as a rangoli.
6. Diwali wishes card
Not a portrait. This is the thing you send to forty people on WhatsApp, and it is a different design problem: it has to survive being viewed at thumbnail size, and it has to leave room for words.
The important instruction is the one telling the model not to write anything. Image models still render text badly, and a card with a misspelt "Happy Diwali" on it is worse than no card. Generate the artwork with an empty band, then type your own greeting over it in any phone photo editor. That also lets you send the same card with different names on it.
Add your greeting afterwards in any phone editor. Asking the model for the words is how you end up with "Happy Diwail".
The four things that go wrong
Hands near flames. The single most common failure. A hand holding a diya is two hard problems touching, and the result is usually six fingers or a thumb fused to the lamp. Every prompt here puts the lamps on a surface for that reason. If you want a hand in the frame, crop at the wrist.
Rangoli that is subtly wrong. A real rangoli repeats around a centre. AI rangoli tends to be a pile of pretty shapes with no structure, and anyone Indian clocks it in under a second, even if they could not say why. Asking for radial symmetry and four colours fixes most of it.
Skin that comes back lighter. These tools lighten and smooth Indian skin by default, and they do it quietly — a shade or two, texture gone. That is why the identity line bans it explicitly rather than just asking to preserve the face. Compare the output against your reference photo side by side before you post, not on the small preview.
Sparklers and fireworks. They look like the obvious Diwali prop and they are the worst thing you can put in an AI image. Fast-moving light, motion blur and hands all at once. Every prompt here excludes them on purpose. If you want a sparkler photo, take a real one.
How to run any of these
Open Google Gemini, ChatGPT or Nano Banana on your phone. Upload one clear, front-on, well-lit photo of yourself — not a group shot, not a dim one, not one where you are wearing sunglasses. Paste the prompt. Generate.
If the face comes back wrong, the fix is almost always a better reference photo rather than a better prompt. If it comes back right but dull, ask for "stronger warm lamp light, deeper shadow in the background" and run it again.
Gemini tends to hold a face closer on a single upload; ChatGPT tends to give warmer and softer light. If one flattens your features, paste the same prompt into the other before you start rewriting it.
What to wear in the photo
The one thing an AI photo cannot fix is a colour that fights your skin. Deep jewel colours work on almost everyone, which is why all six prompts specify them — but if you want to know which of them suits you specifically, the colour matching guide is built around Indian skin tones.
If the photo you are making is a stand-in for an outfit you actually plan to wear, two more: the dress codes guide covers what works for a festive evening, and dress for your build decides the cut before the colour does. And a kurta you own will photograph better than one the model invented, every time.
One honest note
Making a Diwali picture with AI is fine. Posting it as though it were a photograph taken at your own puja is not, and the line is simple enough: say it is AI when you post it.
Two things to leave alone. Do not place a deity or an idol into a scene with yourself — the card prompt above excludes idols for that reason. And do not generate a family member who could not be there, which is a use of these tools that goes somewhere unkind faster than people expect.
If any of this resonates, the daily Instagram is where I post the actual examples — @dressingschool.