AI colorize of a Minox frame

[Minox]

AI for you these days, eh?

 

[Don_ih]

How accurate were its guesses? It kind of looks like a colour film photo but the development was somewhat off in temperature or something. Kinda cross-processed-looking.

 

[Helge]

It’s a huge misunderstanding that you can colourize photos and have them look natural.
The colours are already present in the photo as part of the tonality.
The colours are always going to look smudged and extraneous.

That’s why colourizers back in the day quickly arrived at the abstract techniques, that clearly shows off that the photo has been processed, and makes a point of it.

Seem self important, ahistorical whippersnappers have to learn that all over again.
Except I doubt they will ever develop the taste or introspection to get it.

 

[Minox]

Guys, my example was meant to show to what extent the software developed nowadays, versus say, 10 yrs ago. Never was said or affirmed that such modification of a b&w photograph should be considered for serious purposes, or to replace colour photography.

Yes, the colours are generally looking ok-ish, but one can always see that some sort of manipulation is at work here. And no, the colours do not seem natural, and my half-written article on this matter provides plenty of examples where this colorizing thing went wrong and in some particular instances, altogether and completely bad.

I get the strange feeling that my post came across as something entirely not meant to be.

 

[Helge]

Guys, my example was meant to show to what extent the software developed nowadays, versus say, 10 yrs ago. Never was said or affirmed that such modification of a b&w photograph should be considered for serious purposes, or to replace colour photography.

Yes, the colours are generally looking ok-ish, but one can always see that some sort of manipulation is at work here. And no, the colours do not seem natural, and my half-written article on this matter provides plenty of examples where this colorizing thing went wrong and in some particular instances, altogether and completely bad.

I get the strange feeling that my post came across as something entirely not meant to be.

Sorry, I got what you where trying.
And it’s fine.
I was talking in general about the recent craze and even insistence on colourizing anything B&W.

 

[Don_ih]

I was just asking how accurate the guesses were and saying it looks like cross-processed film.

Was the post supposed to sit there and receive no comment? I asked about the colours, Helge talked about AI photo colouring. Those comments seem on-topic.

 

[L Gebhardt]

The easy way to see how good it is would be to desaturate a color image and then have the app colorize it. Could also test other conversion to black and white methods to see how different they are. I wonder if using a red filter on the sky would produce a darker blue sky in the colorized output. Which app did you try?

 

[Minox]

I was just asking how accurate the guesses were and saying it looks like cross-processed film.

Was the post supposed to sit there and receive no comment? I asked about the colours, Helge talked about AI photo colouring. Those comments seem on-topic.

I did not say that, with all the due. It goes without saying that if one posts anything, the post should receive any number and types of comment. It's just that I had the feeling (now of course revised) that my post was misinterpreted. My bad, since obviously it wasn't.

Thank you !

 

[Don_ih]

I wonder if using a red filter on the sky would produce a darker blue sky in the colorized output.

At the very least, using a coloured filter changes the way different colours translate into tone and shade in b&w, so a coloured filter would likely mess up the way the AI colours certain things. However, this technology is probably at a state where it can "recognize" objects and make "educated guesses" as to what colour the things should be.

 

[Minox]

I have noticed two things: a) the software mainly recognizes the foliage and trees and such, probably (my guess) based on the shape of leaves or something, because the green is there (albeit not in natural shades, and many times hesitant and incomplete) and b) sepia or ortho negative frames cannot be read, so the frame is colorized in deeper shades of sepia or deep black and white.

Noticed also that the shadows are very little if, at all, changed after colorizing, the red is extremely dull, yellow or/and orange the same etc. Interesting though, the indoor frames are most of the time having a blue-ish tint.

I have tried to colorize a colour photography, and it lost its colours: red dulled quite strongly, green and yellow went away in a large proportion. Dark colours went almost completely purple. I have attached these two pics here, for comparison.

Original colour photograph

Colorized

 

[Minox]

The easy way to see how good it is would be to desaturate a color image and then have the app colorize it. Could also test other conversion to black and white methods to see how different they are. I wonder if using a red filter on the sky would produce a darker blue sky in the colorized output. Which app did you try?

Something online, a free version of IMG2GO. I haven't tried a desaturated image, only genuine bw negative frames. Interesting point, I shall have a go and see what happens. Thank you !

 

[foc]

AI for you these days, eh?

View attachment 337563

Thank you for posting this. I think the AI did a very good job colourising.
I have tried to colourise photos before in Photoshop (I am handy in Photoshop) but found it difficult to get a good result. This app appears to be much better than anything I could try.

 

[DREW WILEY]

Seem like the "collective memory" of the AI in this case involves a lot of old faded discolored snapshot. Not surprising if the term "colorize" was fed into it. Gives things a vintage look. Blue stripes on the kid's T-shirt seem like a sailor. Maybe it should have added a can of spinach too.

 

[Bronson Dugnutt]

 

[reddesert]

View attachment 337656

Here, you can see that the AI has rendered the skin tones unnaturally yellow.

 

[Bronson Dugnutt]

Here, you can see that the AI has rendered the skin tones unnaturally yellow.

I LoLed. There's a lot of fun and rumination to be had with AI colorization.

 

[Helge]

View attachment 337656

Whoa there Poochie!

 

[warden]

The easy way to see how good it is would be to desaturate a color image and then have the app colorize it.

I tried that just now with an iPhone 7 snapshot using Photoshop's neural colorize filter with the results below. It struggled, especially with the yellows which are absent in the colorized image:

Edited to add:
It would be interesting to see what could happen if the software engineers could tailor the colorize filter based on the film used. After all Ferrania's P30 sees color differently than FP4 for instance.

 

[Minox]

Looks like this one was easy, but even so...

 

[Don_ih]

Frankly, I don't think there is much effort being put into developing an AI to colourize photos. Most of the AI effort seems dedicated to auto-writing webpages and creating believable but completely fictitious images. Not much profit in adding colour to b&w photos.

 

[pentaxuser]

Blue stripes on the kid's T-shirt seem like a sailor. Maybe it should have added a can of spinach too.

Drew, was it these blue stripes that indicate an AI colourisation or was this the lead into a Popeye joke?

Looks like this one was easy, but even so...

View attachment 337679

View attachment 337680

Drew, from what I see the blue stripes look perfectly normal in the first pic
Julian, I can see a lot of people who have bought "new" colour films saying these are great colours and how pleased they are that they are like this.

AI colourisation is in its infancy. Remember early digital cameras when 2-3 MPs was being heralded as the future not that long ago? How many pixels are possible now in say 20 short years?

pentaxuser

 

[reddesert]

AI colorization, text generation, etc, of the type that is all the rage now (machine learning), works by training a model on a very large database of existing images or texts. So what the model produces is strongly dependent on the training set, and on the scoring metrics that are used to train successive generations of the model.

A model can learn some generic features, for example, large blank areas in the top of the image are "sky" and are typically blue, clouds are white, grass is green, water is blue, etc. People are skin-tone (the question of which skin-tone is much more fraught; if the training set contains mostly people with pale skin, the output will as well). However, the model is going to have a harder time learning specifics - it might learn that flowers have color, but it won't necessarily associate plausible colors to particular kind of flowers. In warden's example, it essentially treated the flower bunch as "plant" and a lot of the flowers came out green.

I wonder if the reason that the colorized images look desaturated, like a somewhat faded print or a 70s postcard, is that the training process disfavored strong colors because they were too often "wrong." If it uses desaturated colors, the images come out plausible, avoiding obvious errors, but also undistinguished. Sort of like the way that ChatGPT can write mediocre text on any subject. It's remarkable that it works at all, but it's still mediocre.

 

[DREW WILEY]

Colorization technique began being important to Hollywood quite a while back, whether for better or worse to the old b&w films, you decide. In terms of propaganda purposes, and outright fraud, these more modern means will no doubt get exploited visually, and not just in script, for various nefarious purposes. It's already begun. But in terms of an "artsy" novelty, we'll get sick of it soon enough, just like all exploding fads. But that discussion, with its inevitable stack of barf bags, properly belongs to the digital or hybrid section, and not here.

 

[Minox]

Frankly, I don't think there is much effort being put into developing an AI to colourize photos. Most of the AI effort seems dedicated to auto-writing webpages and creating believable but completely fictitious images. Not much profit in adding colour to b&w photos.

That is true. If I wanted colour photography, I would put a colour film in my camera. But I am curious to understand how are the tones in bw photo "perceived" by the AI to create colour in those areas. It doesn't have much to do with photograph per se, tbh..

 

[Minox]

Drew, was it these blue stripes that indicate an AI colourisation or was this the lead into a Popeye joke?



Drew, from what I see the blue stripes look perfectly normal in the first pic
Julian, I can see a lot of people who have bought "new" colour films saying these are great colours and how pleased they are that they are like this.

AI colourisation is in its infancy. Remember early digital cameras when 2-3 MPs was being heralded as the future not that long ago? How many pixels are possible now in say 20 short years?

pentaxuser

Yes, I have read some articles on the new colour films, where things are coming out blue, purple, yellow or pink. They are probably liked by some, because I see a lot of those online. Can't say I am a fan, though; my idea of colour film has been and still is the classic ones. If I want to ruin a good Portra 400 and have all the images in say, purple, I may just use a filter to accomplish that .

So, yes, AI colorizing is something most people play with as a novelty I suppose. Using it for say, colorizing an entire bw negative? Hell, no. Why would one want to do that? You cannot present the results as neither bw nor colour.

That said, I guess this could be the next thing, up there with the lomo stuff.