Wow, those examples are really incredible! Although, I actually liked most of the non-shaded images better than the shaded versions. The shading was a bit overdone and kind of robbed the colors of their richness.
colourise.sg is interesting, though the demo photos I've thrown at it haven't come out amazing. I like the fact they make it clear it's attempting to go for plausible rather than accurate (that seems to match my results).
Style 2 paints is pretty sophisticated. This thing not so much. I feel like they just took some model and trained it on a specific dataset and called it "new"
Tangentially related: if you have the chance to go see Peter Jackson’s They Shall Not Grow Old then do so. What he managed to do with that old footage is simply incredible. I was left in absolute awe at how he and his team were able to bring that black and white footage to life...not to mention the gravity of the subject matter.
When Ted Turned tried “democratizing” old classic B&W movies he was made fun of, a lot, not by the audience/viewers, but by the professional critics who claimed he was ruining the intrinsic art of those movies, the implication was he was degenerating them.
Sure, the colors were not quite right (I think they improved them after initial missteps), but he was a man ahead of his time with his vision in that regard. Few would bat an eye today.
Black and white movies should stay black and white. George Lucas shouldn’t have messed with Star Wars. Film movies should have film grain. You shouldn’t use a computer to remove grain and make Arnold look like a Madame Tussaud’s sculpture. You shouldn’t crop 4:3 sources to 16:9 because you think people prefer losing some image to black bars. Stop tinkering with shit to “improve” it.
I wouldn’t care, but I do care because it means I have to spend my time taking hours on the internet to find an obscure VHS rip of a version that has not been debauched by digital Dicks.
I mean, I agree with you, but I don't think GP overstated at all just how much the average consumer doesn't care about this stuff. Most of us know people who prefer to have their TV stretch (not just crop) the image from its original aspect ratio to 16:9. I know people who refuse to watch black and white films because they're "boring". Not because of the subject matter, but because they're black and white.
80s colorized movies looked pretty terrible. That was at least half the problem.
I’d say we’re actually more open to black and white media these days. It doesn’t have the ‘stink of the old’ that it had in the couple of decades after color TVs first became cheap. Now it’s viewed more as a stylistic choice and not simply an imperfection to be fixed.
They have their issues, for sure, but viewers, as I understand it, weren't the complainers, it was mostly the professional critics and artsy types who derided them. Joe and Jill Public cared little, as far as I understand it.
They complained about "artistic intent" and respecting it, etc., etc.
> none of these tools have been trained to colour images specific to the Singaporean historical context.
Interesting aspect of image colorization that I hadn't given too much thought to. Will this NN for example handle images taken in different historical contexts, and vice versa?
Cool technical achievement, however many pictures are B&W even in the age of colour for a reason, the meaning changes with colour, the sentiments that one as artistic wants to awake on the viewer aren't the same in the presence of colour.
This isnt cool. These are not accurate. To properly judge any colorization you have to compare against real color photos. Take a color photo, convert it to b+w. Then feed it to the colorization bot. Compare the original to the colorized. It never comes out very accurate. So you know that neither are the historical photos.
Converting b+w to color is like taking a photo and making it into a 3d model. No matter how hard you try, you are inventing data. It isnt a revitalized photo, it is a new piece, an interpretation of the original. Call it computer art, not history.
Style2Paints: https://github.com/lllyasviel/style2paints
It does increase memory consumption by a lot but idk, the improvement to performance is a little nuts