As a final test after part 1, I was curious if I could push any of the algorithms into giving me some real initiative on colour by placing them in a situation where the subject was recognisable as an object that typically has saturated colours in it. I was also curious whether the filename and metadata would be used as cues (spoiler: seems not).
Sample 8 (colourful balloons, minimalist)
My take: colorize-it balloons look like oranges and colourise.sg ones like eggs. At this point, I’m not sure these programs know what balloons are.
Sample 9 (hot air balloon)
My take: colorize-it’s ability to sometimes generate engaging colour where the others cannot, is what you’ll want to keep it around for.
Sample 10 (colourful balloons, abstract)
My take: By this point, it’s clear that colourise.sg thinks balloon-shaped objects are eggs. It may have never actually seen balloons.
Sample 11 (woman with balloons)
My take: The human element helps the AI understand that it’s looking at coloured balloons. Even colourise.sg caught on to this now and surprisingly gave the most colourful rendition, although it looks like that was partly because it conflated the colours with the sky. Still, much better than before. colorize-it.com is adventurous as ever. None could figure out that the road should be grey.
Sample 12 (berry kiwi salad)
My take: auto-colorizer recognises the raspberries, but is reticent to make strong colour choices. colorize-it is all over the shop again, hedging its bets by making some of the berries red and others green. colourise.sg doesn’t seem to know European berries.
I was also surprised that a simple texture like wood was unclear to all but auto-colorizer.
Sample 13 (furniture)
My take: colourise.sg works out a nice contrast between a reddish brown seat and the green plants by the window, while colorize-it and auto-colorizer return essentially monochrome images.
Sample 14 (boy with cat)
My take: Colorize-it wins this one, hands down. I tried another cat image, but cat faces are apparently unknown to these colourisers. The take-home message surely is that the algorithms we have available are fairly good, but the training sets are now the limiting factor. With good training come good results.
First, the good news. In part 1, we saw that these auto-colourisers can often be trusted to produce good skin tones, in the sense that if you use three of them, at least one will turn out good. This can be a real time-saver for doing colourisation work.
Beyond that, auto-colourisers are a bit hit and miss. In this testing, none stood out as being consistently good. In fact, the most noteworthy feature I found is how conservative they are. They mostly avoid bold colours, especially mixing them in the same image, and sometimes don’t colour things at all. I can only assume that this is because the learning process placed heavy penalties on getting bold colours wrong so that it was better for the algorithms to err on the side of caution. Perhaps the way that colour deviations are calculated needs to be reconsidered in light of this.
While careful hand-colourising can in many cases no doubt get much better results, depending on one’s level of skill, some of the above images – especially of vegetation outdoors – provide useful starting points. In the colourise.sg rainbow image, the only thing missing not passably colourised was the actual rainbow, so a fair amount of work could theoretically be saved using this auto-colouriser – so long as it actually gives a good result.
For an actual colourising task, it’s probably worth running an image through a few of these automatic programs just to see if any of them produce useful results.
If I had to pick only one of these three, I would choose colourise.sg – it processes the fastest, provides the highest resolution download image, and is generally a safe choice when you don’t require strong colour. However, it doesn’t know much other than vegetation, sky and people (but the latter quite well and reliably). In spite of this, the resulting image isn’t usually worse than what you started with. In that context, it should be noted that it states clearly that it is intended “specifically for old Singaporean photos”, so I was pushing it beyond the original intention of the creators, but I think it did okay, keeping in mind how the others performed.
The second best option in my opinion is colorize-it.com. I could see myself keeping it in my arsenal for situation where I want bolder colours. It does come with some colour bleeding, though, so will require manual repair where that happens.
auto-colorizer cannot be recommended as it failed twice, once when processing the clown and a second time when processing an image of a mango fruit (not shown in my final report). Also, the payment option is displayed whether or not a result was generated, which could lead some people to attempt to pay and then find out they get nothing for their money – and the “tiny PayPal donation” turns out to be $10. (And the download resolution offered after payment is smaller than the free one from colourise.sg.)
Other recent AI-themed content