Old Shatin

Don't think so

I think this is why we need to see the original image next to it for direct comparison. The AI seems to have taken liberties with the architecture of Ng Yuen (the white building) which don't match with its shape. Also, the terrain seems off for it to be taken in 1965.

Also strictly speaking the body of water back then was still known as Shatin Hoi/Tide Cove. It had yet to be "nullah-fied" into the Shing Mun Rover channel.

Any chance of changing the image to show both versions (pre/post AI) like you have with the others?

thanks Aldi, this makes a bit more sense. It looks as though the blurred part of the original image is a building - possibly either the Babies Home itself (but then where was the image taken from?), or perhaps the Ecclesia Bible Institute - which is still there in the same buildings today. The fact that the AI has decided to just fill it with trees is why the image looks so off. If this was taken from the Babies Home/High Rock, there would be lower-lying land there and not a tree filled hill/slope.

Edit: Any idea why the left has been obscured? 

Hi there,

I believe this aerial photo should clear things up a bit:

https://www.hkmapservice.gov.hk/OneStopSystem/map-search?itemId=DAP&shtName=1963-5452

It was taken back in 1963, thus may be in one of the Mapping books published by the government.

There were fish ponds (probably some form of Kei Wai) and fields of unknown crops adjacent to Ng Yuen.

Current AI results likely add more anomolies than actually enhancing the photos as the enhancement would highly depends on what you feed into the program.  Grainy images usually means more anomolies.  Not to mention over-exposed ones.  Like handling over-exposed JPEG files of modern day digital cameras, close to nothing could be done of these files.

T

I think AI has its place, but it has its drawbacks as illustrated. This is also why I feel it's important to post the original image alongside so we can figure out the discrepancies and why they're there. It can bring out some useful discourse.

I’ve been thinking about AI-generated images a lot in recent days, and come to the conclusion that they are best left off the Gwulo website. Please don’t upload more AI-generated images to Gwulo.

As we’ve seen with recent uploads, we can submit a low-resolution, fuzzy black & white image to an AI service, which will generate details and colours to create a new image that is sharp and in colour. These new AI-generated images are attractive scenes, and watching the before-and-after is quite magical, so people clearly enjoy seeing them. They could attract more visitors to the Gwulo website.

But an unwanted result of the AI generation is that it introduces errors. It is not finding details and colours that were hidden in the original image, but were invisible to the human eye. Instead it is making a series of guesses about what those colours and details might look like, based on similarities with a huge library of other images. Some guesses will be correct and some will be wrong, introducing errors. Over time I worry that these errors will break down trust in the accuracy of the images on the Gwulo website.

One long term goal for the Gwulo website is to attract more visitors and contributors. Another is to make it a trusted resource for people researching Hong Kong’s history.

If I have to choose between the two, I’ll choose less visitors and more trusted accuracy, so that’s the first vote against the AI-generated images.

Next, apart from comments that the AI-generated images look good, I don’t see evidence of them contributing to our understanding of Hong Kong’s history. When there have been several comments on a photo, they are discussing possible mistakes in the AI-generated image. I don’t see them adding to our historical knowledge, but I do see them distracting us to spend time wondering how accurate they are, and discussing errors instead of talking about the original scene.

Finally there’s the simple issue that we end up with two copies of an image, the original and the AI-generated. That’s a problem because we do our best to avoid duplicates, to keep attention focused instead of spread over multiple copies of an image.

In summary, although using AI-generation to add colour and detail (AI services may advertise it as ‘enhancing’, ‘repairing’, ‘upscaling’, or similar) is technically impressive and produces good-looking results, I don’t believe it is an asset to a history website like Gwulo.

Some examples of errors in AI generation:

This is an excellent example of how attractive the result can look. But in its goal to produce an attractive image, the AI tends towards blue skies, lush greens, and whitewashed buildings, with the risk of introducing subtle errors. Was it really a sunny day? Was it really the wet time of year when everything is green, or one of the dry spells when the slopes look more dried out and brown? Were these buildings really whitewashed, or just made from light-coloured granite blocks?

Here the AI has incorrectly guessed that there was a hillside in the background, making the top-left corner of the AI photo inaccurate.

In this case the AI version has lost several chimneys and other details on the roof. I think it may also have incorrectly converted a pitched roof into a flat roof.

Hi David,

I've just seen your post about AI generated images. Totally agree with your decision, but I've got a question.

At https://gwulo.com/media/52429 and https://gwulo.com/media/52430 I uploaded photos that I had enhanced to make clearer using the "crop", "brightness" and "sharpness" functions of the standard photo editing technology on my PC. I'm fairly sure that "crop" and "brightness" are ok, but what about "sharpness"? Does it use AI? Sorry, no idea of what editing system my computer uses. I'll delete these images if there's a problem with "sharpness".

I'm still curious as to why the left hand side has been scribbled out. I'm wondering if there was some inter-missionary rivalry going on? 

If u zoom in, you can see that it's part of the chain-link fence, likely with a group of people posing in front of it, and it was a non-AI way of hiding the people from view.

the fence looks like it's a waist-high one to me, although it's hard to know. The scale doesn't feel right for people to be the reason. It's a shame because I would like to have seen the Ecclesia Bible Institute; there aren't many older images of it available online.

AI is not all inaccurate. There are various types of AI. The one in vogue nowadays is Generative AI. It tends to be more creative, and may add or subtract from the real scenery, depending on what it learns from a humongous amount of past data. The output will depend on what data it is trained on.

For example, a b&w photo coloring AI based on USA photos would color a white/gray baseball cap as alternating white and red, whereas one based on Singapore photos would color it as alternating white and blue, mainly because red is not a color most Singapore males would prefer to wear and much more baseball caps in Singapore photos would be grey rather than red.

Predictive/analytical AI etc. do not add or subtract an object from a photo. It improves the resolution by interpolating extra pixels, by analyzing the surrounding pixels in the photo and predicting what the extra pixels would likely be. This would not be more "inaccurate" than the original photo, which may not be well focused, have proper exposure, or too small in size.