Frank & Alvar, Little Margie Productions 20©05. Best viewed on full screen at Youtube.
Just when we began to think, what the hell happened to Frank Gehry, he pops up in the Financial Times with a boffo review* for LUMA Arles. Love it, hate it (some locals say it looks like a crushed tin can, the Shark thinks it wrecks the landscape), you can't ignore it. The front, covered in polished stainless steel siding, changes color as the day goes by. The tower sits on a giant glass drum, where most of the art is. Smaller exhibit rooms are in the tower, seen from miles away. Frank does not do subtle.
LUMA Arles, Google Maps Street View. Click and drag to look around.
From the start of his career, Frank has caused a stir. When he couldn't get work he did a crazy makeover of his own house. It got him a lot of press, but no big commissions. His career took off when the movers and shakers in Bilbao, Spain commissioned him to design the Guggenheim museum. It saved the town - a huge draw for tourists and new business. LUMA Arles had a different mission – to make a 21st-century statement in an ancient Romanesque city.
Vincent
Van Gogh lived in Arles. His work, at the time, was panned. He painted
the knock-out Starry Night in a mental hospital nearby. Vincent is
yelling from his grave, 'Fuck the naysayers FRANK!”
* Review by Edwin Heathcote
PS: Frank's favorite architect is Alvar Aalto. Little Margie produced "Frank & Alvar" in 2006 (Eki has updated the video - it looks brand new. He'll tell how he did it.)
Sources: Financial Times, personal experience
Next week: Gettin' OLD Ain't Easy
Note: As Maggy said, the above Youtube link is to a new version of our 15-year old documentary, Frank & Alvar. A lot has happened in digital imaging since then. In 2005 when we shot this documentary (released in 2006), the television cameras were big, extremely expensive, and looked rather impressive too. The image quality was great considering the period but in 2021, it looks like crap.
The Standard Definition (SD) of digital widescreen television in Europe, and most countries outside the USA, was anamorphic D1 PAL 50i, 720*576 pixels.
This standard squeezed the 16:9 widescreen image down horizontally by roughly a third to fit the old 4:3 television signal (that's the anamorphic part). The standard also called for 50 fields per second crammed in 25 interlaced (i) frames. This gave a smooth motion, the distinct "TV look" (as opposed to "film look").
For each frame, the image was divided into odd and even rows of "pixels", or TV lines. The odd row was captured and shown first, then the even one. This way the distinct motion steps could be doubled, at the expense of halving the vertical resolution of moving objects.
The "film look" was and still is something a lot of filmmakers wanted - we'd usually rather have our content look like a movie than like a soap opera. So, we would deliberately toss out one of these two TV fields, to get a more movie-like cadence: 25 frames per second instead of 50 fields per second (films are shot 24 frames per second).
This could be done somewhat intelligently, using the full image for static objects, but interpolating the image from just one field for moving objects. It looked pretty good but had the unfortunate side effect of further reducing the resolution of the image anywhere where there was motion.
So, in the end, the file that was broadcast was at a resolution that was 720*576 pixels for static portions of the image, but only 720*288 pixels for any areas that had motion. The motion detecting algorithm was not perfect, so some errors where both fields show in moving areas remained, but that was unnoticeable on a TV set. On a computer screen, which shows the whole image at once, those errors can be visible.
To sum it up, the resolution was abysmally bad compared to what we are used to nowadays: 1920*1080 full HD, or better*. Simply scaling the SD image up just enlarges all the errors, mushiness and softness of the image. That's also why old TV shows look so horrible on Youtube.
Next week: Gettin' OLD Ain't Easy
Note: As Maggy said, the above Youtube link is to a new version of our 15-year old documentary, Frank & Alvar. A lot has happened in digital imaging since then. In 2005 when we shot this documentary (released in 2006), the television cameras were big, extremely expensive, and looked rather impressive too. The image quality was great considering the period but in 2021, it looks like crap.
The Standard Definition (SD) of digital widescreen television in Europe, and most countries outside the USA, was anamorphic D1 PAL 50i, 720*576 pixels.
This standard squeezed the 16:9 widescreen image down horizontally by roughly a third to fit the old 4:3 television signal (that's the anamorphic part). The standard also called for 50 fields per second crammed in 25 interlaced (i) frames. This gave a smooth motion, the distinct "TV look" (as opposed to "film look").
For each frame, the image was divided into odd and even rows of "pixels", or TV lines. The odd row was captured and shown first, then the even one. This way the distinct motion steps could be doubled, at the expense of halving the vertical resolution of moving objects.
The "film look" was and still is something a lot of filmmakers wanted - we'd usually rather have our content look like a movie than like a soap opera. So, we would deliberately toss out one of these two TV fields, to get a more movie-like cadence: 25 frames per second instead of 50 fields per second (films are shot 24 frames per second).
This could be done somewhat intelligently, using the full image for static objects, but interpolating the image from just one field for moving objects. It looked pretty good but had the unfortunate side effect of further reducing the resolution of the image anywhere where there was motion.
So, in the end, the file that was broadcast was at a resolution that was 720*576 pixels for static portions of the image, but only 720*288 pixels for any areas that had motion. The motion detecting algorithm was not perfect, so some errors where both fields show in moving areas remained, but that was unnoticeable on a TV set. On a computer screen, which shows the whole image at once, those errors can be visible.
To sum it up, the resolution was abysmally bad compared to what we are used to nowadays: 1920*1080 full HD, or better*. Simply scaling the SD image up just enlarges all the errors, mushiness and softness of the image. That's also why old TV shows look so horrible on Youtube.
The "zoom in and enhance" trope in movies has always been the laughing stock of anyone who works in the industry. You simply cannot do that - you cannot magically introduce new detail to where there is none.
Well, that used to be the case. But it's not anymore. There's a lot of talk about Artificial Intelligence (AI) currently, and this is one of the best practical use cases of AI in action**.
Zoom in and enhance actually works now. And that's pretty wild.
The way the AI is trained is rather ingenious. You take millions of high-resolution images and make low-resolution copies of them. Then the AI tries to figure out what kind of patterns (often just a few pixels) in the small image correspond to the actual details in the big image. The cool part is, that there's another AI that keeps a score of how well the scaling AI works, how successful it is in reproducing the original large image. Bad guesses get weeded out, good guesses are kept.
 |
| Right-click to open image in a new tab. Compare the AI-enhanced image on the top to the old-style upscaling on the bottom. Also notice how small the original image was. |
No human knows what exactly the AI does, it is way too complex for that. After enough training, you just have a "black box" that knows what needs to be done in order to make a decent big image out of a small image. And it works. Not perfectly, but surprisingly well, even in this first generation of AI image enhancers.
In the case of Frank & Alvar, i used Topaz Video Enhance AI, which is the first commercial software package to use this technology, as far as i know. I actually got better results after scaling the 720*576 image further down a little, to 720*408 16:9 aspect ratio before feeding it to the AI. This reduced the interlacing artifacts i mentioned earlier.
The resulting image quality is surprisingly good. There are places where the AI did some pretty weird choices, making for some surreal imagery when closely examined, but these usually go unnoticed in the moving video. And there also are some shots that could be easily mistaken for a modern HD camera.
Color me impressed.
CU
--
Eki
* Even most new phones have cameras that do 4K (Ultra HD, UHD) video, which is an even higher resolution. SD is about 0.5 megapixels, HD is about 2 megapixels, and UHD is about 8 megapixels. That's 16 times more resolution than SD television had.
** AI image processing can do some rather cool other stuff too - from imagining new frames between the actual ones for a pretty convincing slow motion to creating believable digital humans, landscapes, and other images to making your photos look like Van Gogh paintings. And we're still just starting, scratching the surface of what is possible. And then there are all the other use cases - AI has replaced journalists for some simple reporting, and it can even excel on comedy: https://arr.am/2020/07/17/jerry-seinfeld-and-eddie-murphy-talk-shit-about-san-francisco-by-gpt-3/
* Even most new phones have cameras that do 4K (Ultra HD, UHD) video, which is an even higher resolution. SD is about 0.5 megapixels, HD is about 2 megapixels, and UHD is about 8 megapixels. That's 16 times more resolution than SD television had.
** AI image processing can do some rather cool other stuff too - from imagining new frames between the actual ones for a pretty convincing slow motion to creating believable digital humans, landscapes, and other images to making your photos look like Van Gogh paintings. And we're still just starting, scratching the surface of what is possible. And then there are all the other use cases - AI has replaced journalists for some simple reporting, and it can even excel on comedy: https://arr.am/2020/07/17/jerry-seinfeld-and-eddie-murphy-talk-shit-about-san-francisco-by-gpt-3/
No comments:
Post a Comment