7
\$\begingroup\$

This shot was taken on a Samsung My-Cam f2.8 film camera on Kodak Gold 200, but for some reason, the winding mechanism failed and the shots which were back to back got double exposed. Luckily both the photos are 90 degrees to each other, so it's quite easy if you squint to see the other image.

I need help to find software or a person who can separate these images. They are quite precious to me.

double exposed photo of a group of people

Improve this question
\$\endgroup\$
15
  • 19
    \$\begingroup\$ You can not. Period. \$\endgroup\$
    – Rafael
    Apr 29 at 6:22
  • 17
    \$\begingroup\$ Learn to love this photo :) Print it twice. Display both side-by-side, one rotated 90 degrees from the other - maybe in a single frame with two apertures. \$\endgroup\$
    – osullic
    Apr 29 at 9:53
  • 7
    \$\begingroup\$ @Rafael Please expand that comment a bit and post that as an answer \$\endgroup\$
    – Philip Kendall ♦
    Apr 29 at 10:32
  • 8
    \$\begingroup\$ Theoretically, an AI algorithm could generate two images and then computationally double-expose them, back-optimizing the two inputs such that they generate the observed digitized image. Technically there are an infinite number of pairs which would result in the same double exposure image, but most are obvious Frankenstein's. An AI trained on what normal photos look like should be able to give you what it thinks is "most likely", though. -- I'm unaware if any such algorithms have been trained. You might have better success asking at an image-manipulation venue, rather than a photography venue. \$\endgroup\$
    – R.M.
    Apr 29 at 15:30
  • 4
    \$\begingroup\$ @ArnavBadhe There are online communities that accept requests to restore or color damaged photographs. Some of them are spookily good at it. Maybe one of them will accept your request as an unusual challenge? \$\endgroup\$
    – MackM
    Apr 29 at 18:08

4 Answers 4

Reset to default
9
\$\begingroup\$

It's possible to work at this slowly through an iterative process. It takes a lot of effort and will take some artistic skill to get good results.

The basic process is:

  1. Make a copy of the layer in photo editor.
  2. Use clone brush and smudge to remove all details that are not part of the image in this orientation. Fill in with the color you expect should be there. Don't worry too much about removing detail, because it will come back later.
  3. Set the mode of edited layer to Subtract (note: not the same as Difference).
  4. You should now get most of the other photo. Adjust Brightness/Contrast or Levels setting on the edited layer to get the best result you can at this point.
  5. Copy Merged to new image, rotate by 90 degrees and work on the opposite image.
  6. Repeat steps, always subtracting from the original and then working on the opposite image. Note that subtraction should always be between the original image and the edited image, never between two edited images.

Any detail you remove from one image will appear in the other one. You'll need to figure out somehow what detail belongs in which image.

Here is what I got with this method in 4 iterations. I admit it is not great, but might act as a starting point for sufficiently skilled artist to improve.

enter image description here

enter image description here

Improve this answer
\$\endgroup\$
1
  • \$\begingroup\$ That's so useful thank you so muchhhh! \$\endgroup\$
    – Arnav Badhe
    May 1 at 0:21
7
\$\begingroup\$

You cannot because there is no information in the image which would allow to separate them.
Closest you could get to having double exposure which you could separate is having two monochromatic objects which would expose different film crystals (channels) and then you could separate the photos somehow. Nothing like that is possible in general case and definitely not in your case.

Improve this answer
\$\endgroup\$
3
  • 10
    \$\begingroup\$ As a commenter noted, it cannot be done automatically. But an AI (or human) could theoretically deduce from the patterns what each image ought to have been, and emit what the two source images probably were. \$\endgroup\$
    – Mooing Duck
    Apr 29 at 16:07
  • \$\begingroup\$ Could it be done if I have reference images of the place and other photos of how everyone looked that day? \$\endgroup\$
    – Arnav Badhe
    Apr 30 at 4:11
  • 1
    \$\begingroup\$ @ArnavBadhe that would be more like a repainting job rather than separation that can be done in editor. It's really that difficult. "Similar" photos won't do much. \$\endgroup\$
    – Euri Pinhollow
    Apr 30 at 8:24
6
\$\begingroup\$

It is going to be a long task, and the results will never be "good" let alone perfect. I'm assuming a digital tool like photoshop or gimp etc.

Start by running an edge detection filter over the whole image, to draw a thin black line around every colour transition. You're basically "drawing lines in a cartoon" If this doesn't produce a nice outline, then you might have to do it by hand as below.

Make backup/duplicates of this layer.

Then comes the time consuming part - to carve away any edge-line that shouldn't be there for one image. Save often.

You have an advantage in that the two images are rotated with respect to each other, so the "wrong" image will be slightly harder to perceive.

The point here is to subtly tell viewers which bits are important and which are noise from the other image.

It may help to fade out the outlines by decreasing their opacity, so they act as guides without being too obvious.

I would suggest doing both photos, and always displaying both together. The third composite image makes it a set of three.

Also, enjoy the artistic statement of a blended family photo.

enter image description here

enter image description here enter image description here

The smaller the image, the better the effect. Good luck !

Improve this answer
\$\endgroup\$
2
  • 1
    \$\begingroup\$ thank you so much i will try that \$\endgroup\$
    – Arnav Badhe
    Apr 30 at 1:14
  • 1
    \$\begingroup\$ I like the way you're thinking about it. I feel like there is something clever to be done with average color values in each overlapping segment to tease them apart more fully. In your awesome GIF example, the man on the left has two other faces overlapping his. Can infer something about each face's contribution to the exposure based on its color outside of the overlapping region, and subtract? They'll be odd images, but more recognizable. A color-balancing jigsaw. \$\endgroup\$
    – MackM
    Apr 30 at 1:29
2
\$\begingroup\$

You get someone to paint the two original images, there is no other way.

I think, as said in the comments, you could get a machine learning algorithm to do the painting, but I don't think there is something ready made for this and paying a painter for, I don't know, a day of work per image, is going to be a lot cheaper than paying a software engineer for a few weeks of work to work out the machine learning model and train it.

(that being said, it seems like it would be easy to generate training data for this task, so maybe I'm overestimating the amount of work needed to train a ML model)

Improve this answer
\$\endgroup\$
7
  • \$\begingroup\$ I'm uncomfortable suggesting this, but I'm troubled that somebody will glibly say "AI can do that!". Basically, it might be possible to tell an "AI" to detangle the two images. But it would do that by finding two "good enough" photographs somewhere on the Internet or in some other accessible archive such that their sum was fairly close to the double-exposed one. So in other words, if you asked "an AI" to do it you'd almost certainly get two photos taken by other people, not a two photos extracted from your original. It might be possible to use the found images as masks to extract something fro \$\endgroup\$
    – Mark Morgan Lloyd
    Apr 30 at 16:53
  • \$\begingroup\$ @MarkMorganLloyd The probability that there exist two images on the internet that will sum up to a double-exposed input that you upload is basically 0. This prevents the machine learning model (note I never used the term "AI", that's you) from just giving you images from the training data as a response. But I did say "painting" and that's what I meant, you won't get data from nothing, someone or something will make up that data based on stuff they have seen before. \$\endgroup\$
    – Nobody
    May 9 at 5:31
  • \$\begingroup\$ I didn't say "an exact sum", and would suggest that with their current propensity for hallucinations and gross image mismatch everybody should realise that "AIs" are extremely slapdash at this sort of thing. And please note that I have carefully quoted "AI" in both my comments: it might be artificial but it is in no way intelligent in the way that an IKBS strove to be intelligent. \$\endgroup\$
    – Mark Morgan Lloyd
    May 9 at 6:58
  • \$\begingroup\$ @MarkMorganLloyd That you keep saying "AI" instead of talking about models makes me believe you don't know what you are talking about. Requiring the two images to "sum" (sum is a simplification) exactly to the input image is the basic and most obvious requirement for such a model and it ensures that the model can't simply output images from the internet. "hallucinations" on the other hand don't apply in the technical sense, but in a general, broad sense that's exactly what you want to the model to do: to hallucinate/guess about what the two images could have been. \$\endgroup\$
    – Nobody
    May 9 at 19:01
  • \$\begingroup\$ @MarkMorganLloyd The term "AI" has come to be used far too restrictively recently due to CHAT GPT and friends. AI has a range of meanings and need not mean "look for matching images". If you can conceive of a list of rules that could help a human artist to perform this task then a computer system is liable to be able to be trained similarly. This need not necesarily have access to ANY other images, although models of what human anatomy looks like may well help. \$\endgroup\$
    – Russell McMahon
    May 17 at 13:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.