Using AI to capture and improve quality of VHS tapes?

[Witzker]

In today's discussions, Artificial Intelligence (AI) is a hot topic. Recently, I rediscovered my old Panasonic NV-HS1000 and HD PVR model 1212.

My aim is to preserve the memories stored on old VHS tapes. However, I'm reaching out because I don't want to tackle this task alone.

Call to Action:
Could you lend a hand? I'm seeking your advice and input. Together, let's enhance the quality of these old tapes.

Key Questions:

How can AI improve the appearance of old VHS tapes?

Have you attempted to use AI for repairing old VHS tapes?

Are you interested in joining this project?

If this isn't the right place, do you know of other forums or groups where I could find assistance?

Let's collaborate and safeguard these memories collectively!

[lordsmurf]

Quote:

Originally Posted by Witzker

In today's discussions, Artificial Intelligence (AI) is a hot topic.

Today's discussion is mostly bullshit. It refers to unvetted/untrainer LLM, which has nothing to do with video capture, or really video whatsoever. (Excluding deepfakes, which is still years away from being even remotely passable as realistic, aside from cherry-picked overworked samples.)

The stock market is going ga-ga over AI, and yet it doesn't bring any revenues to anybody (and will not for many years, if ever), aside from Nvidia and maybe AMD. The rest is pure horsepuckey, total speculation by gambler "investors". The "AI" referred to in most quarterly reports is just software that's been in use for years or even decades already. Lipstick on a pig. Sometimes it's just esoteric plans mentioned to juice the stock price for insiders. Seriously, why does Tyson Foods need AI for chicken nuggets or whatever? So much BS out there.

Quote:

How can AI improve the appearance of old VHS tapes?

It doesn't.

Quote:

Have you attempted to use AI for repairing old VHS tapes?

Explain yourself. Use it how, exactly?

There are some Avisynth/Vapoursynth filters that are supposedly based on infantile AI of sorts, but it still doesn't do much. Software from Topaz/etc is just usurping the term for marketing (to gullible newbies), there's not really any AI in there either.

By some definitions, all software is AI.
Similiar to how now all data storage is now a cloud. Even though, at one time, "cloud" had a very distinct definition, which involved failover. Cheapskates, wanna-be's, and "me too's" eroded that definition to now be meaningless. "AI" will suffer the same fate at this current rate.

Quote:

Are you interested in joining this project?

No. Not without some proof that it's not just pixie dust and fairy farts. My time is too valuable to piss away on nothing.

Quote:

If this isn't the right place, do you know of other forums or groups where I could find assistance?

Oh, it's the right place for ther topic of video capture, or digitizing consumer analog formats like VHS. But it's the wrong place to BS others. Just stating "let's use AI" will rightfully be mocked as ridiculous. You must put forth actual ideas on how/what you think can be improved. That said, the hardware will be the limiter, so this is all somewhat moot as a discussion.

[traal]

I've tried using VEAI on VHS but it has little effect other than a very slight sharpening, until you get aggressive with the sliders, and then it creates strange artifacts. So at the moment, it's pretty much worthless for VHS.

It also doesn't noticeably improve any LaserDiscs as far as I can tell, or some DVDs. Other DVDs show moderate improvement. DV footage also improves.

I am hopeful that someday, AI processing will improve analog video. But we don't seem to be there just yet.

[lordsmurf]

Quote:

Originally Posted by traal

I've tried using VEAI on VHS but it has little effect other than a very slight sharpening, until you get aggressive with the sliders, and then it creates strange artifacts. So at the moment, it's pretty much worthless for VHS.

Avisynth LSFmod, which has been around for at least a decade now, is vastly better at sharpening VHS to extremes. Not that you should, but you can, and without artifacts. It doesn't tout "AI" whatsoever.

You probably already know this, but posting just as good example to the conversation.

[dpalomaki]

IMO AI is an attempt to incorporate human analysis. a rule set, and resulting actions into a machine that may be able to apply that human understanding (including experience gained from repeated application) faster and in a more consistent fashion. It hopefully builds on the decision processes of experts to produce software that gives better results than a not-so-expert could do on their own. It is dependent on the skill of the software coder. Lake any product that is sold, it is a sexy name applied to a piece of the evolving man-machine relationship.

Lots of other examples that were just a rebranded evolution of coding processes. A simple example
"GOTO"s became
callable subroutines
that were repackaged into libraries that could be linked into a program
that, due to file & memory size limitations became overlays distributed with the program
and then bundled in to the OS as DLL, then
web apps, and so on.

Not so much new under the sun but evolved from earlier art.

One can make a cake from a packaged mix, or from scratch with raw ingredients. Which is better depends on the skill of the baker and the quality of the ingredients.

[lordsmurf]

Quote:

Originally Posted by dpalomaki

IMO AI is an attempt to incorporate human analysis. a rule set, and resulting actions
Lots of other examples that were just a rebranded evolution of coding processes. A simple example
"GOTO"s became

I would suggest it's nothing more than an advanced IF/THEN.

If/then is essentially infantile machine learning. The machine is human queried, and then it asks itself questions based on data set (current knowledge), in order to arrive at IF/THEN decisions, and take actions. But the problem is always the data set.

LLMs, the current popular/"it" aspect of AI, is just the verbose mode. R2-D2, with his "boops" and "beeps", was more intelligent, and without LLMs for verbose output.

There's no real "intelligence" at hand yet, and probably will not be for more decades minimum, or more likely even centuries.

These current "AI" can get confused really quickly (bad data sets, stupid self-query, "does not compute"). Especially given how most are being trained on random input (ie, internet data), as opposed to specific expertise and experience. Specific to video, that presents as artifacts, mush, and mess.

The real kicker is that even when fed extremely well-vetted data, the machine often does some really dumb things. So fixing the input data isn't enough.

Neural networks are essentially just the IF/THEN being refined with new data sets in real-time, forming some low-level degree of cognitive function (stored in tensors). But "learning from your own mistakes" is still a challenge. Most AI researchers are still not sure how/why the AI makes certain decisions. To a degree, the mind of an AI is almost like the mind of a psychopathic amoeba.

A lot of AI people get all pissy when their magical box is dumbed down to an IF/THEN base. "No, it's different!", and then they vomit jargon all over themselves, or scribble out complex math equations. But if they can't explain it to a normal person, they don't understand it themselves.

And then most items touted as "AI" aren't any of the above anyway. Marketing meets gullible.

Way, way back in school, I was influenced by Transformers (G1). I wanted to get into robotics, either the engineering side, or the programming side. But I lacked the math skills. I'm just not BS'd easily on this "AI" stuff. When most people think of AI, they think of Arnold's Terminator and Skynet. But I think of Optimus Prime, which came out 1 month earlier in 1984! Or Asimov books, and I read several in the 80s.

[whatastory]

I have AVC Video Enhancer AI, it is terrific although I've noticed that it is heavily dependent on your computers hardware. I use it with 10th gen i5, 8GB RAM, and an RTX 3060. On average a 1 - 2 hour video takes 8 - 12 hours. You can just run it and do something else. From my observations, the best enhancement is the denoise and it is a night and day difference. It cleans up the video really good and improves it tremendously. An alternative is Topaz. They both do the same. I don't think theres any significant difference in upscaling directly and you might have to look really hard to see. But AVC offers the denoise seperately as their AI model along with outputting any resolution you want. If viewing on a modern TV or monitor don't make the resolution 480p as I feel like that can skew results. Output to a 16:9 aspect ratio because then it won't need to be upscaled further and will appear as is. I've prefer AVC over Topaz because its more simplistic. Topaz overwhelms me with so many features and its very hard to configure settings to specific videos. AVC has given me thorough results with just simple clicks without having to figure anything out and waste time experimenting.

Also deinterlacing i've found through AVC to be excellent. I've compared the deinterlacing with other software and this does the best. AVC offers monthly for one license on 1 PC and lifetimes subscription to 1 PC. The denoise and deinterlacing is top quality and no other software IMO can top these as it goes frame by frame and makes sure every single frame is cleaned up well. The time it takes is long, but I believe that to achieve excellent quality a long processing time is needed to not half ass results.

[timtape]

Often not grasped in this context is that speaking of intelligence, the most intelligent detective in the world, whether human or artificial, still needs evidence (relevent data) to have a hope of solving the crime. Just as humans (who create AI) cant get around that basic problem, neither can AI.

Talk is cheap, as always, but what lasts and gains respect in the long run is consistent, repeatable results.

[dpalomaki]

Quote:

...overwhelms me with so many features and its very hard to configure settings to specific videos..

If that is the case it is not very smart. The intelligence is in making the rule set; what to look for and what to do when it is found, and what is a satisfactory end result.

System learning may be may be a matter of the system counting the human feedback for each case and then in the future acting based on what the human said was right most often.

Nothing magic there. Also, as time goes by we learn more secrets about the "AI filtering" schemes used by social media that arguably favor some ideas and search results over others, or perhaps steering users to the sources of the services income. In the 1950's rock-and-role music industry it was called payola.

[lordsmurf]

Quote:

Originally Posted by whatastory

I have AVC Video Enhancer AI,
Topaz AI
overwhelms me with so many features and its very hard to configure settings to specific videos..

It's because none of this stuff is actually AI by any true definition. It's just a BS label slapped on the box. ("Now with 20% less fat!" ... because it's 20% smaller!!!) Misleading jabberwocky. It's the meme of the moment, and it makes money for that reason alone.

- One is a dummy GUI, so it hides settings.
- The other gives you all the options ... because you're the "I" (hopefully, though sadly not usually), and there is no "A". Your choices = your video settings. The AI there would be as dumb as a dog operating a VCR.

"AI" has such a loosely defined term that everybody is claiming it now. All software is AI?

Avisynth will likely be vastly more impressive if you ever learn it. No AI, no BS.

Quote:

Originally Posted by dpalomaki

Also, as time goes by we learn more secrets about the "AI filtering" schemes used by social media that arguably favor some ideas and search results over others, or perhaps steering users to the sources of the services income. In the 1950's rock-and-role music industry it was called payola.

This is already highly suspect, a topic making the rounds under the surface. The reason is because LLM/generative AI is potentially (not just yet) eating into revenues from the same company doing both. So to lose in one, it must be made up in the other. Enter sleight of hand.

Many "AI" have already been exposed as human-influenced "AI" (choke collar for ulterior motives).

[dpalomaki]

LLM, brings to mind the game from years ago where you had three stacks or words, one of nouns, one of verbs, and one of one of objects. You picked one from each stack, strung them together, and it made sense (sort of). (E.g., who eats dodo.) and with 50 items in each stack you could get 125,000 unique responses!

An simple example of an "AI-ish" noise reduction that would be relatively easy to program perhaps.
0. Point the program at a video clip
1. It finds a segment with minimal luma levels and luma changes.
2. Assume it is all noise.
3. Use it as the profile for noise reduction software.