This reminds me of how a lot of older TV shows and movies had an 'enhance' button that could somehow magically take a tiny, blurred image and blow it up to a level of detail where even the most minor things could be seen again. Like a suspect's eyes in the reflection of a window that looked like it was maybe 4 pixels in size in the original image.
The AI examples and reasoning in the article are exactly why this can't work, and is so unrealistic. The info isn't just hidden away in the info in a way some tool can figure out what it is with another effort, it's simply not present entirely. Any tool that tries to enhance a low quality image or video is just going to be making up the missing data.
This is generally true. What I find fascinating is the application of information that is not present in the data but that can be confirmed by the data. I’m thinking specifically of deblurring algorithms people use to recover text. If you can guess what font is being used in a screengrab with, say, someone’s bank account blurred out, you can take the i formation contained in the downloaded font and try combinations until you match the blur algorithm that may also be information not contained in the image.
What’s also crazy, is that even if all you do is black out the text, kerning alone can still be used to guess the correct text with some degree of accuracy. I worked on redaction software at one point, and these were very much the types of things we thought about when dealing with text.
>What I find fascinating is the application of information that is not present in the data but that can be confirmed by the data. I’m thinking specifically of deblurring algorithms people use to recover text.
I would argue that high frequency information is not missing in a blurred image, rather it's spread out spatially. Every sample in the blurred image is a convolution of multiple samples (i.e. high frequency data) from the originating source. Theoretically, it should be invertible, provided the exact function is known that does the convolution, but there are practical limitations, such as knowing how to apply the inverse operations at the edges and dealing with the extreme sensitivity to errors.
> Any tool that tries to enhance a low quality image or video is just going to be making up the missing data.
So there is some things that instinctively don't sound correct here.
A frame may not contain the info to enhance, maybe it's too blurry... but a series of frames can (I'm pretty sure) between them have enough info to deblur an object in the background, it really helps to know the focal length, and distance maps...
My only point is information in a video is more than the sum of the information of it's frames, it also captures delta, not just from frame to frame, but relevant frame to relevant frame. That's a fair bit to work off of.
I do agree with our general point when information is 'absent'. I disagree how much absent information is in a video which can't be reconstructed by this delta. It certainly isn't zero, or 100%.
Inventing colors from a grayscale is a good example of absent information and of where this goes awry. Deblurring/blurring operations, less so.
> but a series of frames can (I'm pretty sure) between them have enough info to deblur an object in the background
You're probably thinking of geometrical super-resolution, such as gigapixel photography. Blurring is the discarding of high-frequency information, and only a very very small part of this information can be recovered using techniques like this (as in, so little that you wouldn't be able to notice it).
> My only point is information in a video is more than the sum of the information of it's frames,
It's actually less than the sum of the information of it's frames. An off-the-shelf lossless compression algorithm can give you an upper bound for the amount of information present in a video file.
> This reminds me of how a lot of older TV shows and movies had an 'enhance' button that could somehow magically take a tiny, blurred image and blow it up to a level of detail where even the most minor things could be seen again.
2 main problems with this article: one it's fundamentally asking for something impossible, and two for the part that _is_ possible it's using the wrong tool for the job.
It's not possible to magic details into existence that weren't there. So if you think a color was wrong, well, that is not something that can be solved other than by using a time machine or finding a separate reference picture. Any human restoration artist would be doing the exact same thing of guessing the color.
The second problem brought up is that ChatGPT is willing non-color details into existence that aren't in the original picture (eyebrows etc). For this it's simply that the public ChatGPT image generator wasn't made for this task. You'd have to build a custom harness but I'm sure it can be done.
I'm genuinely curious what you thought "AI restoration" was doing if the fact that it added details that it thought were reasonable was surprising to you
Yeah, this is fundamentally expected behavior but the results can vary wildly based on how much of a gap there is between available information and the 'resolution' of the desired output.
I feel like managing the level of information available and judging if its sufficient to make an extension of details lead to a reasonable outcome is the responsibility of the human/user more than the AI. If half the image is destroyed, it is less of a restoration and more of an imagination but the process is the same as if it just had some scuff marks to buff out. Its on the person using the tool to recognize the extent of impact that using the tool is having and to communicate that extent when discussing the results.
otherwise i could give the "restoration process" a torn corner of a photo from my great grandmas house and after it completely makes up a photo i call it restored
But this is exactly what you expect when you ask intelligence (machine or human) to fill in missing information. It picks some plausible, middle-of-the-road filler. If you were a restoration artist with an airbrush (or photoshop) with just the instruction to make the picture nicer, that's what you'd do. If you got told the right colour, you use it, and if not, you guess.
> But this is exactly what you expect when you ask intelligence (machine or human) to fill in missing information.
I haven’t read the blog post since it’s currently hugged to death (so I’m admittedly filling in missing information), but one could alternatively expect the intelligence to research the correct missing information, and to report back if it is unable to.
If it's possible, sure. But sometimes it's not possible, as with the brother's eyebrows or the horizontal slats on the gable in the first picture, both of which were invented by the AI.
Your last sentence is the important one, yeah? On a spectrum of intelligence or expertise, the higher end would be defined (ish) by knowing the right color or having better guesses anyway.
Like someone restoring the Sistine Chapel would ideally have some of: knows what it looked like before because they've been working on it for thirty years, because they've been researching and/or have access to photos or some other historical documents of it's priors, because they're educated on the types of materials and pigments from the time, maybe they know something about the daily use/tourism of the place and how/whether it should change some parts of the restoration. In the absence of those things, then infer that something should be purple-ish because it's blue to the left and red to the right. This is good, but perhaps below great. Maybe the gap should have a hard line between the two colors or maybe there was an object in there.
And of course plenty of people maybe just don't care, aren't getting paid enough, don't have enough discipline to get the right purple, or are playing to their supervisor's eye rather than their trained intuition. The last one having a lot to do with the notion of "plausible". This effort aspect is I guess different, but still related/adjacent to what comes out of or how that "intelligence" is deployed.
A restoration artist would not invent eyebrows because they would have worked under a theory of the sort "human faces are especially sensitized to human interpretation so if the noise image does not strongly support a feature then it is better to leave it vague and noisy than to try and pick one out of a large space of eyebrows belonging to other human faces". This is an entirely reasonable, intuitive theory.
I bet that ChatGPT actually would regurgitate exactly this notion but cannot effectuate it.
I think ChatGPT will just be calling out to a diffusion workflow to generate the new image, so this is using a general tool for a precise problem. It's not going to be great output.
I think, given you know some details about the film you're scanning and how it reacts to certain hues/brightnesses, you can remap colours from greyscale to a few possible colours right? Obviously it's not a 1:1 mapping since there's some data lost. (A dark blue is going to be a similar grey to a brown or something) but I'm almost certain there's old ML technology that essentially takes the most likely colour based on the context of the entire scene/some convolution matrix chunk of the scene, trained for specific film formulations.
This reminds me of an old problem I used to run into with duplicate data.
Duplicates are not always what they seem - duplicates have a relative quantity of how duplicate they are, for example - 100% duplicate identical is what it sounds like, but then what is a 50% duplicate? Well.. that could be a duplicate where semantic meaning that only a human is aware of (missing information) could be used to determine that these two datum are in fact duplicates in different forms, where 50% is identical, but the other 50% is semantically identical. Or 50% truly identical could, in some case, be "good enough" to be considered identical. These are just examples.
Why this example is relevant is because, when deduplicating data, we often have to consider which value(s) we keep as the true singular representation.
The problem is, long after the data is created, this may be impossible to do.
The solution? Often times it is no better than flipping a coin and guessing, i.e. randomly keeping one datum and discarding what we consider to be the duplicate. Or you can try to be fancy and merge the data or choose, based on some rule, what you assume is the data with the highest valid signal and discard or ignore the noise.
This happens every day in practical information systems with messy data.
I always come to the conclusion that in the end, one cannot truly replace missing signal precisely - one can only infer the correct signal. If the information was not recorded and or compressed or corrupted beyond recoverability, there is no way to get it back - only to infer or guess with some degree of confidence, and accept or reject the outcome.
This to me actually clarifies something deeply about the conversion between the real world to the digital one as data, and the conversion of data into information- it is a real physical process.
In my mind it makes clear that the recording of data (or generation of data) is a real physical process of transfer, from an instrument making observations to physical bits in a physical system, which become the raw data.
Without the raw, precise, complete data, your observation is lossy, and therefore everything above it is as well- e.g. information/interpretation, etc.
This isn't a negative. If stated correctly, the above just helps to explain how we imperfectly navigate through the real world with our tools.
All the same, we can do some pretty amazing things in the absence of data, even if some of it is "false" or "wrong".
I spent some time working in the data industry and was fascinated with this topic as well
dont forget the data itself is already a compression of reality. text data has lost tone, situational context, emphasis, etc.
then youve also got different data that yields the roughly same information that isnt exact but serves the same purpose. for example, birthday vs age; one can be used to infer the other although it is more accurate moving in one direction than the other.
youve got resolution of data. for example, A specific hexcode value vs "deep navy blue" vs "blue". sometimes the signal is low resolution but the data type is high resolution and you end up with some tells like when a lot of date data is the 1st of the month or 1st day of the year; implying the signal was at the month level or year level, no day signal available. which to call back an earlier example, can be caused by getting a signal for age but storing it in a birthday data type field
to your point, often the "data" signal is not a single field but set, and then youve got your source updating every month or year, and its not always the same format or exact set, maybe it was collected differently or two or more different sources contributed. its hard to apply any rules without assessing the type of data being considered. newer data that is missing older data, sometimes you carry the old pieces forward and sometimes you dont. sometimes you change the shape of the data by inferring events. job title changed? well the old job title isnt wrong data, its just not their current title and now can be moved to a past job titles" field. maybe past job titles arent something that add value to your use case and it can be dropped, etc.
Occasionally I've advocated to product-owners that their design needs a business/domain-concept for uncertainty or data-quality, rather than treating it as an implementation detail.
This is especially true when you've got data where even a human specialist would throw up their hands and say: "It could be this, or that, I can't tell."
The human brain does the same thing. It fills in details that aren't really there. The eye can only resolve detail in a fairly small spot at the center of vision. But we normally don't perceive that. Our entire field of view seems equally sharp. The brain is filling in the detail.
>I went back and explicitly told ChatGPT not to resolve details that didn’t exist in the file.
I don't know what the author thinks this could have meant, other than "give me the original". Making up details is what he is asking it to do.
True, you can't see the eyebrows in the original, but you also can't see (IMO) an absence of eyebrows and since far more people have eyebrows than not, I would be extremely surprised if it had come up with a version without them.
I feel like the basics of information theory are a valuable thing for people to learn, especially as AI that makes up information starts to become more prevalent.
This reminds me of how a lot of older TV shows and movies had an 'enhance' button that could somehow magically take a tiny, blurred image and blow it up to a level of detail where even the most minor things could be seen again. Like a suspect's eyes in the reflection of a window that looked like it was maybe 4 pixels in size in the original image.
The AI examples and reasoning in the article are exactly why this can't work, and is so unrealistic. The info isn't just hidden away in the info in a way some tool can figure out what it is with another effort, it's simply not present entirely. Any tool that tries to enhance a low quality image or video is just going to be making up the missing data.
> is just going to be making up the missing data.
This is generally true. What I find fascinating is the application of information that is not present in the data but that can be confirmed by the data. I’m thinking specifically of deblurring algorithms people use to recover text. If you can guess what font is being used in a screengrab with, say, someone’s bank account blurred out, you can take the i formation contained in the downloaded font and try combinations until you match the blur algorithm that may also be information not contained in the image.
What’s also crazy, is that even if all you do is black out the text, kerning alone can still be used to guess the correct text with some degree of accuracy. I worked on redaction software at one point, and these were very much the types of things we thought about when dealing with text.
>What I find fascinating is the application of information that is not present in the data but that can be confirmed by the data. I’m thinking specifically of deblurring algorithms people use to recover text.
I would argue that high frequency information is not missing in a blurred image, rather it's spread out spatially. Every sample in the blurred image is a convolution of multiple samples (i.e. high frequency data) from the originating source. Theoretically, it should be invertible, provided the exact function is known that does the convolution, but there are practical limitations, such as knowing how to apply the inverse operations at the edges and dealing with the extreme sensitivity to errors.
> Any tool that tries to enhance a low quality image or video is just going to be making up the missing data.
So there is some things that instinctively don't sound correct here.
A frame may not contain the info to enhance, maybe it's too blurry... but a series of frames can (I'm pretty sure) between them have enough info to deblur an object in the background, it really helps to know the focal length, and distance maps...
My only point is information in a video is more than the sum of the information of it's frames, it also captures delta, not just from frame to frame, but relevant frame to relevant frame. That's a fair bit to work off of.
I do agree with our general point when information is 'absent'. I disagree how much absent information is in a video which can't be reconstructed by this delta. It certainly isn't zero, or 100%.
Inventing colors from a grayscale is a good example of absent information and of where this goes awry. Deblurring/blurring operations, less so.
> but a series of frames can (I'm pretty sure) between them have enough info to deblur an object in the background
You're probably thinking of geometrical super-resolution, such as gigapixel photography. Blurring is the discarding of high-frequency information, and only a very very small part of this information can be recovered using techniques like this (as in, so little that you wouldn't be able to notice it).
> My only point is information in a video is more than the sum of the information of it's frames,
It's actually less than the sum of the information of it's frames. An off-the-shelf lossless compression algorithm can give you an upper bound for the amount of information present in a video file.
coughRedDwarfcough
https://www.youtube.com/watch?v=JMIHNiR3CP8
https://youtu.be/FVOydVwOO4M
or samsung and it's "space zoom" AI feature that helps you take pictures of the moon:
https://www.androidauthority.com/samsung-fake-space-zoom-moo...
now that astronauts can take their phones into space, wonder what a shot of the earth would look like on a "space zoom" phone?
> This reminds me of how a lot of older TV shows and movies had an 'enhance' button that could somehow magically take a tiny, blurred image and blow it up to a level of detail where even the most minor things could be seen again.
Mandatory BladeRunner scene:
https://youtu.be/8-Rw-XAsBhM
2 main problems with this article: one it's fundamentally asking for something impossible, and two for the part that _is_ possible it's using the wrong tool for the job.
It's not possible to magic details into existence that weren't there. So if you think a color was wrong, well, that is not something that can be solved other than by using a time machine or finding a separate reference picture. Any human restoration artist would be doing the exact same thing of guessing the color.
The second problem brought up is that ChatGPT is willing non-color details into existence that aren't in the original picture (eyebrows etc). For this it's simply that the public ChatGPT image generator wasn't made for this task. You'd have to build a custom harness but I'm sure it can be done.
I'm genuinely curious what you thought "AI restoration" was doing if the fact that it added details that it thought were reasonable was surprising to you
Yeah, this is fundamentally expected behavior but the results can vary wildly based on how much of a gap there is between available information and the 'resolution' of the desired output.
I feel like managing the level of information available and judging if its sufficient to make an extension of details lead to a reasonable outcome is the responsibility of the human/user more than the AI. If half the image is destroyed, it is less of a restoration and more of an imagination but the process is the same as if it just had some scuff marks to buff out. Its on the person using the tool to recognize the extent of impact that using the tool is having and to communicate that extent when discussing the results.
otherwise i could give the "restoration process" a torn corner of a photo from my great grandmas house and after it completely makes up a photo i call it restored
But this is exactly what you expect when you ask intelligence (machine or human) to fill in missing information. It picks some plausible, middle-of-the-road filler. If you were a restoration artist with an airbrush (or photoshop) with just the instruction to make the picture nicer, that's what you'd do. If you got told the right colour, you use it, and if not, you guess.
> But this is exactly what you expect when you ask intelligence (machine or human) to fill in missing information.
I haven’t read the blog post since it’s currently hugged to death (so I’m admittedly filling in missing information), but one could alternatively expect the intelligence to research the correct missing information, and to report back if it is unable to.
If it's possible, sure. But sometimes it's not possible, as with the brother's eyebrows or the horizontal slats on the gable in the first picture, both of which were invented by the AI.
Your last sentence is the important one, yeah? On a spectrum of intelligence or expertise, the higher end would be defined (ish) by knowing the right color or having better guesses anyway.
Like someone restoring the Sistine Chapel would ideally have some of: knows what it looked like before because they've been working on it for thirty years, because they've been researching and/or have access to photos or some other historical documents of it's priors, because they're educated on the types of materials and pigments from the time, maybe they know something about the daily use/tourism of the place and how/whether it should change some parts of the restoration. In the absence of those things, then infer that something should be purple-ish because it's blue to the left and red to the right. This is good, but perhaps below great. Maybe the gap should have a hard line between the two colors or maybe there was an object in there.
And of course plenty of people maybe just don't care, aren't getting paid enough, don't have enough discipline to get the right purple, or are playing to their supervisor's eye rather than their trained intuition. The last one having a lot to do with the notion of "plausible". This effort aspect is I guess different, but still related/adjacent to what comes out of or how that "intelligence" is deployed.
A restoration artist would not invent eyebrows because they would have worked under a theory of the sort "human faces are especially sensitized to human interpretation so if the noise image does not strongly support a feature then it is better to leave it vague and noisy than to try and pick one out of a large space of eyebrows belonging to other human faces". This is an entirely reasonable, intuitive theory.
I bet that ChatGPT actually would regurgitate exactly this notion but cannot effectuate it.
I think ChatGPT will just be calling out to a diffusion workflow to generate the new image, so this is using a general tool for a precise problem. It's not going to be great output.
I think, given you know some details about the film you're scanning and how it reacts to certain hues/brightnesses, you can remap colours from greyscale to a few possible colours right? Obviously it's not a 1:1 mapping since there's some data lost. (A dark blue is going to be a similar grey to a brown or something) but I'm almost certain there's old ML technology that essentially takes the most likely colour based on the context of the entire scene/some convolution matrix chunk of the scene, trained for specific film formulations.
This reminds me of an old problem I used to run into with duplicate data.
Duplicates are not always what they seem - duplicates have a relative quantity of how duplicate they are, for example - 100% duplicate identical is what it sounds like, but then what is a 50% duplicate? Well.. that could be a duplicate where semantic meaning that only a human is aware of (missing information) could be used to determine that these two datum are in fact duplicates in different forms, where 50% is identical, but the other 50% is semantically identical. Or 50% truly identical could, in some case, be "good enough" to be considered identical. These are just examples.
Why this example is relevant is because, when deduplicating data, we often have to consider which value(s) we keep as the true singular representation.
The problem is, long after the data is created, this may be impossible to do.
The solution? Often times it is no better than flipping a coin and guessing, i.e. randomly keeping one datum and discarding what we consider to be the duplicate. Or you can try to be fancy and merge the data or choose, based on some rule, what you assume is the data with the highest valid signal and discard or ignore the noise.
This happens every day in practical information systems with messy data.
I always come to the conclusion that in the end, one cannot truly replace missing signal precisely - one can only infer the correct signal. If the information was not recorded and or compressed or corrupted beyond recoverability, there is no way to get it back - only to infer or guess with some degree of confidence, and accept or reject the outcome.
This to me actually clarifies something deeply about the conversion between the real world to the digital one as data, and the conversion of data into information- it is a real physical process.
In my mind it makes clear that the recording of data (or generation of data) is a real physical process of transfer, from an instrument making observations to physical bits in a physical system, which become the raw data.
Without the raw, precise, complete data, your observation is lossy, and therefore everything above it is as well- e.g. information/interpretation, etc.
This isn't a negative. If stated correctly, the above just helps to explain how we imperfectly navigate through the real world with our tools.
All the same, we can do some pretty amazing things in the absence of data, even if some of it is "false" or "wrong".
I spent some time working in the data industry and was fascinated with this topic as well
dont forget the data itself is already a compression of reality. text data has lost tone, situational context, emphasis, etc.
then youve also got different data that yields the roughly same information that isnt exact but serves the same purpose. for example, birthday vs age; one can be used to infer the other although it is more accurate moving in one direction than the other.
youve got resolution of data. for example, A specific hexcode value vs "deep navy blue" vs "blue". sometimes the signal is low resolution but the data type is high resolution and you end up with some tells like when a lot of date data is the 1st of the month or 1st day of the year; implying the signal was at the month level or year level, no day signal available. which to call back an earlier example, can be caused by getting a signal for age but storing it in a birthday data type field
to your point, often the "data" signal is not a single field but set, and then youve got your source updating every month or year, and its not always the same format or exact set, maybe it was collected differently or two or more different sources contributed. its hard to apply any rules without assessing the type of data being considered. newer data that is missing older data, sometimes you carry the old pieces forward and sometimes you dont. sometimes you change the shape of the data by inferring events. job title changed? well the old job title isnt wrong data, its just not their current title and now can be moved to a past job titles" field. maybe past job titles arent something that add value to your use case and it can be dropped, etc.
Occasionally I've advocated to product-owners that their design needs a business/domain-concept for uncertainty or data-quality, rather than treating it as an implementation detail.
This is especially true when you've got data where even a human specialist would throw up their hands and say: "It could be this, or that, I can't tell."
The human brain does the same thing. It fills in details that aren't really there. The eye can only resolve detail in a fairly small spot at the center of vision. But we normally don't perceive that. Our entire field of view seems equally sharp. The brain is filling in the detail.
>I went back and explicitly told ChatGPT not to resolve details that didn’t exist in the file.
I don't know what the author thinks this could have meant, other than "give me the original". Making up details is what he is asking it to do.
True, you can't see the eyebrows in the original, but you also can't see (IMO) an absence of eyebrows and since far more people have eyebrows than not, I would be extremely surprised if it had come up with a version without them.
I feel like the basics of information theory are a valuable thing for people to learn, especially as AI that makes up information starts to become more prevalent.
[dead]
Maybe you forgot to put "Enhance" in the prompt?
/s
I say "f*ck you; pay me" when the app begs me for my data; that's what.
(sorry for the censored swear)