Thanksgiving AI Photo Prompts: 8 Looks Tested on 3 Models

Last November, right after Google released Nano Banana Pro, social feeds filled up with AI Thanksgiving photos: people dropping themselves into a golden dinner table, a 1970s family snapshot or a parade crowd. The prompts were easy to find. What the prompt lists didn't show is what happens when the photo has two people in it, which is what most Thanksgiving photos have.

So we wrote eight Thanksgiving AI photo prompts, from the classic dinner table to a Friendsgiving toast, and ran all eight with the same photo of a couple on three image models: Nano Banana Pro and Nano Banana 2.1 from Google (the Nano Banana family is what Gemini uses for images), and GPT Image 2.5 from OpenAI (the model behind ChatGPT Images 2.5).

Below are the eight prompts you can copy, the side-by-side results, which model keeps both faces best, and how to make a whole family's photos in one go.

How we tested#

• One couple photo, made for the test. The two people in every picture are fictional: we generated their selfie with an AI model rather than use real people's photo. It is a plain phone shot in a hallway, she in a mustard sweater, he in a navy one.
• Eight looks, eight prompts. Each prompt describes the scene, the clothes where it matters, the light and the camera style, and ends with the same identity line for both people: keep both people exactly as they are: same faces, same features, same skin tones, same ages, same hairstyles.
• Three models, as available in October 2026: Nano Banana Pro and Nano Banana 2.1 (Google), and GPT Image 2.5 (OpenAI).
• Same settings everywhere: the couple photo as the reference image, 1K resolution, 3:4 portrait, one attempt per prompt and no cherry-picking. All 24 images rendered on the first try.

One run per prompt is a small sample, so treat the results as tendencies, not rules. We ran the same kind of test with a single selfie for Halloween AI photo prompts, and several patterns repeated here.

Results at a glance#

The short version:

• To look like yourselves: Nano Banana Pro.
• For candid, lived-in party scenes: Nano Banana 2.1, which is also the cheapest of the three at 1K.
• For the most styled, magazine-like food and color: GPT Image 2.5, if you accept a few extra years and a familiar pose.

The 8 Thanksgiving AI photo prompts#

Each prompt below is exactly what the models received. Change the scene freely; keep the last two sentences.

1. Thanksgiving dinner table#

Thanksgiving dinner: turn the two people from the reference photo into a warm candid photo at a Thanksgiving dinner table at home. They sit side by side at a long table set with a roast turkey, mashed potatoes, cranberry sauce, pumpkin pie and small autumn leaves, candles glowing, other family members softly blurred around the table. Warm golden evening light, shallow depth of field, natural candid smiles, professional lifestyle photo. Keep both people exactly as they are: same faces, same features, same skin tones, same ages, same hairstyles. No text, no logos.

The photo everyone wants, and all three delivered a usable one. Two things to notice. All three kept the mustard and navy sweaters from the selfie, because the prompt didn't mention clothes. And GPT Image 2.5 copied the selfie pose: heads together, his arm reaching toward the camera, as if the couple had taken the original photo at the table. Nano Banana 2.1 gave the most natural "caught mid-conversation" shot.

2. 1970s family snapshot#

1970s family snapshot: turn the two people from the reference photo into an authentic 1970s Thanksgiving snapshot. They stand in a wood-paneled dining room with harvest-gold and burnt-orange decor, a roast turkey on a platter in front of them, wearing period clothes: a patterned turtleneck and a brown corduroy jacket. Direct on-camera flash, faded warm film colors, soft grain, slightly off-center amateur framing. Keep both people exactly as they are: same faces, same features, same skin tones, same ages, same hairstyles. No text, no logos.

The strongest look of the test on every model: wood paneling, shag carpet, gold curtains and a cornucopia. The prompt never said who wears what, and all three gave her the turtleneck and him the jacket. GPT Image 2.5 had the most convincing period colors (and a floral turtleneck); Nano Banana Pro and Nano Banana 2.1 were closer to the original faces. If you only try one prompt from this list, try this one.

3. Carving the turkey#

Carving the turkey: turn the two people from the reference photo into a candid kitchen photo on Thanksgiving. The man carves a golden roast turkey on a wooden board while the woman laughs beside him holding a pumpkin pie in oven mitts, a busy home kitchen with steam, dishes and herbs behind them, warm window light, natural documentary style photo. Keep both people exactly as they are: same faces, same features, same skin tones, same ages, same hairstyles. No text, no logos.

With two people, the risk is the model swapping who does what. It didn't happen: 3 of 3 put the knife in his hands and the pie in hers. "The man" and "the woman" were enough to tell them apart. GPT Image 2.5 was the only one that actually sliced the turkey; the Nano Banana models posed with the knife on a whole bird. Both Nano Banana 2.1 and GPT Image 2.5 added aprons.

4. Autumn leaves portrait#

Autumn leaves portrait: turn the two people from the reference photo into a cozy fall portrait in a park with orange and red leaves falling around them. They wear flannel shirts and chunky knit scarves and hold paper cups of hot apple cider, golden hour backlight, soft warm haze, shallow depth of field, professional couple photo. Keep both people exactly as they are: same faces, same features, same skin tones, same ages, same hairstyles. No text, no logos.

The safe choice and close to a tie: a simple, front-facing, well-lit pose is where every model keeps faces best. It is also the look that works as a holiday card. Nano Banana 2.1 put the steam from his cup right in front of his mouth, so it reads a little like breath.

5. Backyard football#

Backyard football: turn the two people from the reference photo into a candid action photo of a Thanksgiving touch football game in a backyard covered in fallen leaves. The woman throws an American football and the man runs to catch it, both in hoodies and jeans, laughing, leaves kicked up, late afternoon sun, motion and energy, sports photo. Keep both people exactly as they are: same faces, same features, same skin tones, same ages, same hairstyles. No text, no logos, no numbers.

Action is harder than posing. GPT Image 2.5 got the moment right: ball in the air, her arm following through, his hands out to catch. Nano Banana Pro froze her before the throw with him running at her rather than away. Nano Banana 2.1 shot it wide and added other players in the background, so the faces are small. Nano Banana Pro also turned the sweaters into hoodies of the same colors.

6. Friendsgiving#

Friendsgiving: turn the two people from the reference photo into a fun Friendsgiving party photo in a small city apartment with warm string lights. They sit on the floor at a low coffee table crowded with mismatched plates of food, raising glasses of wine in a toast with a few friends around them, phone flash, candid and lively. Keep both people exactly as they are: same faces, same features, same skin tones, same ages, same hairstyles. No text, no logos.

The scene with the most extra people, and none of the friends came out as a copy of the couple, a common failure when a model invents new faces next to a reference. Nano Banana 2.1 was the only one that put them on the floor at a low table, with coats on a hook and a rug: it looks like a real apartment party. Nano Banana Pro sat them on a sofa. GPT Image 2.5 ignored the floor and the coffee table and built a full dining table, beautifully lit, with the couple once again in the selfie pose.

7. Parade morning#

Parade morning: turn the two people from the reference photo into a candid photo at a Thanksgiving parade on a cold city morning. They stand in a cheering crowd bundled in winter coats, wool hats and scarves, holding paper cups of hot cocoa, while a giant balloon shaped like a cartoon turkey floats above the street behind them. Crisp morning light, breath visible in the cold air, street photo. Keep both people exactly as they are: same faces, same features, same skin tones, same ages, same hairstyles. No text, no logos, no brand names.

We described the balloon instead of naming a real parade or a character, and all three drew a friendly turkey. GPT Image 2.5 turned it into a selfie again and added things nobody asked for: a pilgrim hat on the turkey and a Christmas wreath on a lamppost. The two Nano Banana models stayed closer to the prompt, with visible breath in the cold.

8. Vintage painted illustration#

Vintage illustration: turn the two people from the reference photo into a 1950s magazine-style painted illustration of a Thanksgiving dinner. They sit at a family table and smile as a grandmother sets down a huge roast turkey, warm gouache painting with visible brushstrokes, cheerful nostalgic mood, soft warm palette. Keep both people recognizable as themselves: same faces, same features, same skin tones, same ages, same hairstyles. No text, no lettering, no signature.

The only non-photo look, and the one that split the models most. Nano Banana 2.1 painted the most authentic mid-century illustration, but in doing so it turned her into a generic 1950s woman: the period style beat the likeness. Nano Banana Pro kept both faces clearly recognizable, at the cost of a painting that looks closer to a modern portrait. GPT Image 2.5 sat in between, with a rich oil-like finish. We wrote "recognizable as themselves" rather than "exactly as they are" for this one, since a painting can't be exact.

Faces up close#

At thumbnail size all 24 look like the same couple. Up close:

• Nano Banana Pro is the most faithful for both people, in photos and in paint: same eyes, same smile lines, same beard shape.
• Nano Banana 2.1 is about as close in the photo looks. Where it drifts is style-heavy prompts: the painted version of her has a different nose and jaw.
• GPT Image 2.5 keeps the likeness but adds age to both, with deeper lines and heavier skin texture. We saw the same in the Halloween test, so it looks like a trait of the model rather than bad luck.

Two faces in one photo did not make things worse than one face. The bigger risks were the ones below: clothes and poses carried over from the reference photo.

Tips for writing your own Thanksgiving prompts#

The first two tips are what we saw in the test; the rest are fixes for the problems above that we did not re-run, so try one image before a batch.

• End every prompt with an identity line for everyone in the photo. "Keep both people exactly as they are" worked for all eight looks.
• Name people by something visible when they do different things. "The man carves, the woman holds the pie" was enough; with two women or two men, use "the person on the left" or hair color.
• Name the clothes, or you get the clothes from the photo. Wherever the prompt didn't mention outfits, the selfie sweaters came along. Write "wearing a cream cable-knit sweater instead of their own clothes".
• Ask for a new pose with GPT Image 2.5. It reused the selfie pose in three of eight looks. "Standing apart, not taking a selfie, both hands visible" leaves less room.
• Say what not to add. GPT Image 2.5 added a pilgrim hat and a Christmas wreath. "No hats on the balloon, no Christmas decorations" is cheap insurance.
• "No text" is a request, not a guarantee. Signs, banners and posters in the background are where stray lettering usually shows up. Zoom in and check before you post.

Make the whole family's photos at once#

In a chat app every photo and every look is its own round: upload, paste, wait, save. Three family photos and four looks is twelve rounds. With a bulk image generator it is one batch:

1. Start a New batch and paste the prompts you want into the prompt box, one prompt per line.
2. Upload each photo as a reference image: the couple, the kids, the grandparents, the dog. Choose Per image under How to use the references, and every photo is paired with every prompt.
3. Pick the model. Going by this test, Nano Banana Pro is the safest for faces. Use Nano Banana 2.1 for candid party scenes and the lowest price, or GPT Image 2.5 for the most styled food and color.
4. Check the total and start. The form shows the image count and credits before anything is spent. Three photos × four prompts = 12 images; on Nano Banana Pro at 1K that is 12 × 5 = 60 credits. Download everything as one ZIP.

Start with one photo and your two favorite looks, then run the rest. If one person's face drifts, try a sharper, front-facing photo of them. For more on keeping people consistent across many images, see consistent character AI images; for December, we have Christmas AI photo prompts tested the same way.

FAQ#

Which AI model is best for Thanksgiving photos of a couple or family? In our test Nano Banana Pro kept both faces most faithfully, Nano Banana 2.1 made the most natural candid scenes and is cheapest, and GPT Image 2.5 had the richest styling but made both people look older. With one run per prompt, test your own photo before running a batch.

Can I use these prompts in Gemini or ChatGPT? Yes. They are plain language and work with any model that accepts a reference photo. Gemini's image generation runs on the Nano Banana family; ChatGPT's runs on GPT Image.

Can I put two people from one photo into the same AI scene? Yes. All 24 of our images kept both people from a single couple photo. Write "the two people from the reference photo" in the prompt and end with an identity line that covers both.

Why did the AI keep the clothes from my photo? If the prompt doesn't describe outfits, models tend to reuse what the people wear in the reference photo. Name the clothes you want and add "instead of their own clothes".

Can I make Thanksgiving photos for the whole family at once? Yes. Upload everyone's photos as reference images, use Per image mode, and each photo gets every look in one batch.