5 ms·
> Any detail you put in would be hallucinated and not representative of reality. A good algorithm will absolutely represent reality, otherwise you would notice
by nomel 5y ago
> Any detail you put in would be hallucinated and not representative of reality.
A good algorithm will absolutely represent reality, otherwise you would notice that it's not working. It may not reflect the actual reality behind the black hole, but it would remove the whole, making it indistinguishable from what you would think should be behind the hole.
With claims of "the hole is part of the history" aside, it would make these photos much easier to look at, allowing one to focus on the content of the photo, rather than the hole.
- VWWHFSfQ 5y ago> it would remove the whole, making it indistinguishable from what you would think should be behind the hole There is no algorithm in existence or will ever be discovered in our lifetimes that could cover those black holes with enough accuracy, or creativity, to make it indistinguishable from the real thing.
- random314 5y agoThe training data for such a neural network is extremely easy to generate. Combine it with a GAN, and Demis Hassabis will get it working in 5 years - if you can get him interested. https://www.whichfaceisreal.com/methods.html https://www.whichfaceisreal.com/methods.html
- Ma8ee 5y agoI'd say that Lightroom already today does quite impressively, and I'm certain that there are specialised software that does even better.
- oneeyedpigeon 5y agoNot even if it's supplied other images from the same location, potentially at the exact same time period?
- gruez 5y agoThat will get you a "best guess", not necessarily what was actually captured. It'll be the same if you told a historian + artist to draw in the blanks.
- GaggiX 5y agoLaMa give you already pretty robust result. https://github.com/saic-mdal/lama https://github.com/saic-mdal/lama Certainly neural networks capable enough to perform this task in our lifetime will be developed.
- Ansil849 5y ago> It may not reflect the actual reality behind the black hole And therein lies the problem. I'd rather see literally nothing, a black hole, than some AI-constructed representation of what an algorithm determines 'should' be there.
- pvaldes 5y agoA trade-off solution could be to turn the hole color from black to a gray that match the average tone from the perimeter area surrounding it. The hole would be less noticeable at the first sight and would allow people to focus in the rest of the photo.
- Ansil849 5y agoMy impression was that that is precisely the whole point of this project: to draw attention to the gaping hole in the photos, to highlight the crass censorship which led to these photos being suppressed for years. The hole _is_ the focus of the photo, hence why the title is 'The black hole photographs', to highlight that these photos were chucked into a black hole.
- pvaldes 5y agoThis is not censorship, is editorial choices
- Ansil849 5y agoTry reading the article. In this instance the editors were the censors, as they were government workers. And regardless of whether or not you want to call it censorship or some obfuscated term like "editorial choices", the point remains that the black holes draw attention to these "editorial choices", which seems to very clearly be the point of this project. Removing them, or making them unobtrusive is counter to the aims of the project.
- nomel 5y ago
- watwut 5y ago> A good algorithm will absolutely represent reality, otherwise you would notice that it's not working. It may not reflect the actual reality behind the black hole If you define "represent reality" as "lie in a convincing way", then maybe. But, it has zero with reality.
- ben_w 5y agoFilling the hole in a way that nobody can find where the hole was without the original is easy. But that is likely to mislead people into thinking that what filled the hole is what was really there at the time. Even small errors can radically change an image, as this (super-resolution rather than content-aware-fill) example shows: https://www.theverge.com/21298762/face-depixelizer-ai-machine-learning-tool-pulse-stylegan-obama-bias https://www.theverge.com/21298762/face-depixelizer-ai-machin...
- npteljes 5y ago>you would notice I don't think this is a good metric. For fun and entertainment, it works, but lots of contexts need more than that. Take an example of lossy audio, like MP3. People are generally fine with 128k mp3s, but I doubt audio engineers would want to store masters in that format and bitrate. In case of colorization, the issue is that the color really is lost on the film. Multiple shades of different colors map to the same shade of grey on black and white media - my favorite example is that B&W TV make-up was bright green. So them no matter what you do, it'll be just a conjecture of what the original color might have been.