Matching Colors Between AI Images and Your Own Photos
Drop a generated element into a photograph and it will look wrong, even with a perfect cutout. The reason is almost always colour, and it is fixable in about ten minutes with a repeatable method.
- Match the generated image to the photograph, not the other way round.
- Fix the black and white points first with Levels. Everything else gets easier.
- Neutralise the colour cast with Curves on individual channels.
- Generated images are usually too saturated. Pull it down.
- Grain last. Apply it over both layers together so they share it.
Why They Never Match
Three consistent differences, and knowing them tells you what to correct.
| Difference | What it looks like | Fix with |
|---|---|---|
| Higher saturation | Colours slightly more vivid than reality | Hue-Saturation |
| No grain | Unnaturally clean surfaces | A noise layer on Overlay |
| One even colour temperature | Every part of the image lit the same way | Curves per channel |
| Lifted blacks | Nothing in the image is truly dark | Levels black point |
| Compressed highlights | Nothing is truly bright either | Levels white point |
Step 1: Read the Numbers
Guessing by eye is unreliable, because your eyes adapt within seconds. Measure instead.
- Open both images in GIMP, each as its own layer.
- Press O for the Color Picker.
- Click something that should be neutral grey or white in the photograph. Note the R, G and B values.
- Click the equivalent area in the generated image. Note those values too.
- The difference between the two sets is exactly what you need to correct.
A neutral area reads roughly equal on all three channels. If the photo reads 128, 130, 133 and the generated image reads 140, 136, 120, then the generated image is too bright and much too warm.
The colour converter has an eyedropper that reads any image and shows the values in every format, if you would rather do this outside GIMP.
Step 2: Match the Black and White Points
Do this before anything else. It fixes a surprising amount on its own.
- Select the generated layer and open Colors, Levels.
- Look at the histogram. Generated images usually have gaps at both ends, meaning nothing is truly black or truly white.
- Drag the black input slider right, to where the data actually begins.
- Drag the white input slider left, to where the data actually ends.
- Compare against the photograph histogram and match roughly.
Step 3: Kill the Colour Cast
This is the step that does the most work.
- Open Colors, Curves on the generated layer.
- Change the Channel dropdown from Value to Red.
- If the generated image is too red, pull the middle of the red curve down slightly. If not red enough, pull it up.
- Repeat for Green and Blue, using the numbers you measured.
- Go back to Value and add a gentle S curve for contrast.
Work in small moves. A curve adjustment of a few pixels on screen is a large change in the image. The colour correction guide covers the Curves dialog properly.
The shadows and highlights split
Real photographs almost never have neutral shadows. Cool shadows with warm highlights is the standard look, and copying it makes a generated image read as photographed.
On the blue channel in Curves, raise the shadow end very slightly and lower the highlight end very slightly. That single adjustment is most of what people mean by cinematic grading.
Step 4: Bring the Saturation Down
Almost every generated image is a little too vivid. Open Colors, Hue-Saturation and drop saturation by 5 to 12.
If only one colour is off, use the channel buttons in that dialog to target it. Generated greens and skies are the usual culprits.
Step 5: Share the Grain
The last step, and it must be last. Grain has to sit on top of both images so they share the same texture.
- Finish all colour work on both layers first.
- Add a new transparent layer above everything.
- Filters, Noise, RGB Noise at a low amount, around 0.05.
- Set the layer to Overlay at about 15 percent opacity.
- Adjust until it matches the grain already in the photograph.
The Whole Thing in Order
A Faster Check
Two ways to test whether you are actually there.
The squint test
Step back from the screen and half close your eyes. Detail disappears and only tone and colour remain. If one part of the image jumps out, it does not match yet.
The Difference mode test
Duplicate the composite, set the top copy to Difference blend mode, and flatten. Areas that are nearly black are matching. Bright areas are where the two images disagree most.
The blend mode visualizer lets you see what Difference does before you use it this way.
For Compositing Specifically
If you are placing a generated element into a real scene, three things beyond colour matter.
- Light direction. Shadows must point the same way. Paint a new shadow on a Multiply layer if the generated one is wrong.
- Edge quality. A hard cut against a soft photograph reads as fake. Feather by two or three pixels. See cleaning up edges.
- Perspective. The horizon lines must agree. Use Perspective transform on the pasted layer.
Layer masks explained covers the masking side, and removing backgrounds from AI images covers getting a clean cutout in the first place.
Common Questions
Why do generated images never quite match my photos?
Three things at once. Generated images run slightly more saturated, they have almost no grain, and they usually have a single even colour temperature where a real photo has mixed lighting.
What is the fastest way to match two images?
Sample a neutral area in each with the colour picker, compare the values, and correct the difference with Curves. Two minutes, and it gets you most of the way.
Should I match the photo to the AI image or the other way round?
Match the AI image to the photo. The photograph is the reference for what real looks like.
Does this work for compositing a person into a scene?
Yes, and it is the step that makes or breaks a composite. Colour matching does more for believability than the quality of your cutout.
Explore on GIMP.cc
Step-by-step GIMP guides for every skill level.
Resize, compress, convert, and check images online.
Answers to the most common GIMP questions.
Share this article: