# How do I combine two photos into one with AI so it looks natural? Image 1 / image 2 prompts

Say which photo is the base and what goes into it, then match light, perspective, scale and shadows. Real Miku AI data, success rates and prompts.

Published 2026-11-12 · https://miku.app.itstudio.club/blog/ai-combine-two-photos-naturally

> **Short answer:** Refer to the photos as image 1 and image 2 in upload order, say which one is the base and what from the other one goes into it, then ask the model to "match image 1's light direction, colour temperature, perspective and scale, add contact shadows, and leave the rest of image 1 unchanged". On Miku AI, 236 people ran 1,769 two-image compositing tasks in the last 30 days. GPT Image 2.5 Flare succeeded about 89% of the time, the best of the models with a large sample.

## Why composites look fake

When two photos are put together and something feels off, it is usually one of these:

- **Light**: one photo has sun from the left, the other a front flash.
- **Perspective and camera height**: one is shot at eye level, the other from above, so the person looks pasted on.
- **Scale**: a person taller than the door, a cup bigger than a hand.
- **Shadows and contact**: no shadow under feet or a cup, so things float.
- **Colour and quality**: one warm and one cool, or one sharp and one noisy.

An AI model repaints the blend area rather than cutting and pasting, which helps. But it does not know which photo should set the rules, so the prompt has to say so.

## What Miku AI users did

We looked at tasks on Miku AI in the 30 days up to 12 November 2026 whose prompts asked to merge two images or put something from one image into another: **1,769 tasks from 236 users**. Every task used reference images, with a median of two. Tasks were grouped by prompt keywords, and models with fewer than 20 tasks are not listed. "Success" means an image was produced, not that the blend looked natural.

| Model | Tasks | Success | Median time | Credits per 1K image |
| --- | --- | --- | --- | --- |
| [GPT Image 2](https://miku.app.itstudio.club/models/gpt-image-2) | 831 | 72% | ~87 s | 0.3 |
| [GPT Image 2.5 Sunburst](https://miku.app.itstudio.club/models/gpt-image-2.5-sunburst) | 392 | 64% | ~90 s | 0.3 |
| [GPT Image 2.5 Flare](https://miku.app.itstudio.club/models/gpt-image-2.5-flare) | 374 | 89% | ~71 s | 0.3 |
| [GPT Image 2.6 Sunburst](https://miku.app.itstudio.club/models/gpt-image-2.6-sunburst) | 79 | 84% | ~80 s | 0.3 |
| [Nano Banana 2](https://miku.app.itstudio.club/models/gemini-3.1-flash-image-preview) | 24 | 79% | ~75 s | 0.25 |
| [GPT Image 2.6 Flare](https://miku.app.itstudio.club/models/gpt-image-2.6-flare) | 23 | 100% | ~52 s | 0.3 |

What stands out:

- Overall about 74% of tasks produced an image, and about 12% were refused by content moderation.
- Refusal rates differed a lot by model: about 18% for GPT Image 2.5 Sunburst, 14% for GPT Image 2 and 4% for GPT Image 2.5 Flare.
- GPT Image 2.5 Flare succeeded about 89% of the time over 374 tasks and was faster than GPT Image 2. GPT Image 2.6 Flare succeeded on all 23 of its tasks, a small sample.
- Aspect ratios were spread out: 16:9 about 35%, 9:16 about 26%, 3:4 about 19%, 1:1 about 10%. Match the frame to your base photo.
- About 55% of tasks were 1K, 23% 4K and 22% 2K.

## Step by step in the workspace

1. Upload the base photo (the scene or background) first, as image 1, then the person or object to add, as image 2. Upload order sets the numbers.
2. Pick a model. Based on the data, GPT Image 2.5 Flare is a good first try.
3. Choose a frame that matches the base photo: 9:16 or 3:4 for portrait, 16:9 for landscape.
4. To change only one part of the base photo, click "Mask edit" on image 1 and paint where the new element should go. Only that area is changed.
5. Check the blend at 1K, then generate at 2K or 4K.

## Prompt templates

Putting a person into a scene:

> Image 1 is a photo of a wooden pier by the sea. Image 2 is a half-body photo of a man in a white shirt. Place the man from image 2 into image 1, standing halfway along the pier, full body in frame, sized so the railing reaches his waist. Keep his face, hair and shirt as in image 2, and complete the lower body with light trousers and canvas shoes. Light him like image 1: warm evening sun from the right, a rim light on his right side, his shadow falling to the left across the boards. Match image 1's colour and grain, and leave everything else in image 1 unchanged.

Putting an object into a photo:

> Image 1 is an overhead photo of a desk. Image 2 is a small succulent on a white background. Place the succulent from image 2 to the left of the laptop in image 1, about a palm's width across, with the pot and plant shape and colours as in image 2. Keep the overhead angle of image 1, add a soft contact shadow matching the window light from the top of the frame, and do not move anything else in image 1.

## Common mistakes

- **Not saying which image is the base.** The model may put image 1 into image 2, or change both.
- **Using "it" or "that".** Write "the pier in image 1", "the man in image 2".
- **No scale reference.** Use something already in the scene as a ruler: a railing, a door, a cup.
- **No light direction.** Mismatched light and shadow is one of the quickest giveaways of a composite.
- **Very different resolutions.** A tiny source image blurs when enlarged; use originals.
- **Photos of people you have no permission to use.** Get consent before compositing someone's photo, and check rights for commercial use.

Putting a product into a scene is a common variant of this job; see [how to put your product into a scene photo](https://miku.app.itstudio.club/blog/ai-put-product-into-scene-photo).

## FAQ

### Which model is best for combining two photos?

In Miku AI's last 30 days, GPT Image 2.5 Flare succeeded about 89% of the time over 374 tasks with about 4% refused, so try it first. GPT Image 2 was used most, with about 72% success. These GPT models all cost 0.3 credits per 1K image.

### Can I combine three or more images?

Yes. Describe each one in order: image 1, image 2, image 3. GPT models accept up to 16 references and the Nano Banana models up to 14; see [how many reference images Nano Banana Pro can take](https://miku.app.itstudio.club/blog/nano-banana-pro-reference-image-limit).

### Do I get credits back if moderation refuses the task?

Ordinary failures are refunded automatically. On models with a content-policy rule, a content-violation failure may not be refunded. Reruns are charged normally. See [pricing](https://miku.app.itstudio.club/pricing).

### The person's face changed after compositing. What can I do?

Add "keep the face and hair of the person in image 2 unchanged", and use a sharp, front-facing photo where the person fills a good part of the frame.
