Is your image still real after AI edits?

Localized AI editing

What happens to an image that you upload to ChatGPT or Gemini and have it make small changes?

I took a trip this past year to Brooklyn Botanic Garden during cherry blossom season. After taking the photo below, I thought it would look much nicer with ducks in the pond at the bottom of the image. I asked ChatGPT to edit my photo to add ducks, and then asked Gemini to do the same.

Original Image

Image generated from ChatGPT asked to "Add ducks to the photo".

Image generated from Gemini asked to "Add ducks to the photo".

At a quick glance, both ChatGPT and Gemini did a great job adding ducks while not changing the rest of the image. All the trees look the same, the people are still there. Besides the localized edits, this is still the photo I took, right?

Unintended differences

Let's take a closer look to confirm. Areas that are not the pond should be exactly the same if this were a truly localized edit. Zooming in on parts of the image can let us compare the original photo to the edited ones.

Figure 1. The people at the left of the scene. As we zoom in on the man’s face, we see clear differences between the original and the edited images.

There are very clear differences between these photos! While this person's general hair, skin color, and pose stayed the same, the facial features have shifted dramatically. The person in the edited images does not actually exist.

Let's look at a few more examples:

Figure 2. Another group outside the requested edit. Compare the faces, sunglasses, and hair.

Figure 3. The man near the center-right. The collar on his shirt and facial features are inconsistent between photos.

Figure 4. People on the right. The sunglasses and even the tilt of her head have shifted. We only asked for edits to the pond area of the photo.

This does not just change the people but also the rest of the image as well. High texture areas can end up blurred or faded. Small details easily overlooked are different or missing. The whole image has been changed and most of the time the user is not aware of it.

Figure 5. The building in the background. Its dome remains recognizable, the branches crossing its facade differ.

Figure 6. The flowering trees. Fine branches and blossom textures provide another example of changes outside the requested edit.

Why can this happen?

Generative image models tend to make changes not directly to the pixels but in a compressed latent space. A very simplified way to show this is illustrated below. The full resolution image that you took on your phone is passed through an encoder that compresses the image and loses fine detail information in that process. It then is prompted to make changes in this compressed latent space. In the decoder phase it reconstructs the image with the changes. But because it lost some of the fine details it guesses at what those original details were, making small but noticeable differences.

Figure 7. A simplified latent-space editing workflow. Note: This is a representative workflow of how some image editing models work, not verified ChatGPT or Gemini internals.

These differences are also acknowledged from the companies themselves. OpenAI admits ChatGPT may edit beyond where instructed to, and its API documentation explains that an editing mask provides guidance without guaranteeing an exact boundary.[1][2] Gemini’s guide describes localized editing through prompts and recommends describing important details to help preserve them from changing.[3]

If you are editing with something like ChatGPT or Gemini remember to inspect the parts you did not ask it to change. Unwanted changes can sneak through a rough look over. Art you made by hand and then edited with these tools can change the very details you put hard work into creating in the first place. AI detection tools may flag your whole image as AI generated (as now it technically is!) even though you only intended for small changes.

If exact preservation matters, keep the original and use a workflow that explicitly copies untouched pixels. One option is to take the generated area, manually crop it out and paste it back onto the source image. Some software allows you to select only certain sections and will only touch the pixels in that area. Conversational editing is convenient, but it is important to be aware that it may make your image more AI than you originally intended.

Want to check to see where AI is detected in an image? Try out our tool:

Pangram Image Research Preview

---

How the comparisons were prepared

The original image was 2048 × 1542 and Gemini outputted the same resolution; ChatGPT output was 1445 × 1088. For a fair visual comparison, both the original and Gemini images were downsampled to ChatGPT’s resolution using Lanczos resampling. Gemini shifted the contents in the image so it was translated back to match the original to compare the zoomed in differences. Enlargements use nearest-neighbor scaling.

Sources

1. OpenAI, Images in ChatGPT
2. OpenAI, Image generation
3. Google, Gemini image generation