so where's all the Topaz Video Enhance AI discussion?

[rrats]

I'm all for progress, I saw some upscaled Britany Spears SD clips on youtube using this new technology and it was an astonishing achievement

The hell are you talking about, all the Britney clips on YouTube have the composite artefacts no matter how old or new they are. And all of them have got sharpening halos. It literally is only denoising and sharpening

[Jarrod1937]

dellsam34, I'm not as hostile as these other guys, haha. While I understand what they're saying, they're also overly hostile for some reason, almost like it's a personal vendetta. Anyways, you can checkout Topaz yourself, the free trial is mostly unlimited, but outputs with a watermark. I'd suggest using Proteus as your first choice and ignoring any and all of their presets. Though, as with most others here, I have found it isn't a magic bullet, but is a good software for an intermediate processing. Success does vary per content type, but it can work nicely.
As for the whole Topaz or not thing, I honestly am not sure why the hostility. Even if alternatives exist that can achieve similar results in specific areas, the software is relatively easier to use than others. I have no problem with full blown programming/scripting, but for others I can see it being a nice alternative for some. Tools are tools, use whatever works for you.

[SaurusX]

Esrgan ftw.

[dellsam34]

dellsam34, I'm not as hostile as these other guys, haha. While I understand what they're saying, they're also overly hostile for some reason, almost like it's a personal vendetta. Anyways, you can checkout Topaz yourself, the free trial is mostly unlimited, but outputs with a watermark. I'd suggest using Proteus as your first choice and ignoring any and all of their presets. Though, as with most others here, I have found it isn't a magic bullet, but is a good software for an intermediate processing. Success does vary per content type, but it can work nicely.
As for the whole Topaz or not thing, I honestly am not sure why the hostility. Even if alternatives exist that can achieve similar results in specific areas, the software is relatively easier to use than others. I have no problem with full blown programming/scripting, but for others I can see it being a nice alternative for some. Tools are tools, use whatever works for you.

Thanks, I'm not really too obsessed about restoration, For me the benefits are not worth my time and effort whether topaz or old school tools, I focus my effort on getting the analog contents out of their medium with the best hardware I can get my hands on, Once digital, de-interlacing is as far as I can go for now, Maybe in the future when these technologies become more efficient and less destructive I may consider doing some of the valuable footages of my family. I like to follow the progress though.

[poisondeathray]

I'm all for progress, I saw some upscaled Britany Spears SD clips on youtube using this new technology and it was an astonishing achievement

The hell are you talking about, all the Britney clips on YouTube have the composite artefacts no matter how old or new they are. And all of them have got sharpening halos. It literally is only denoising and sharpening

I'm wondering which video(s) he was referring to, and if he compared them properly

Actually , some of the good upscaled music videos from that period (there are many crappy ones too) used digibeta masters or high quality intermediates as sources. Sources better than what VEVO official or YT had used, or the VHS or DVD versions. For example, the official version might have rainbows, dotcrawl and look really soft, but the intermediate used for upscaling that someone had access to - would already be rainbow free, clean and have more detail to begin with even prior to upscaling. So as expected, the results are better than the VEVO version even with traditional methods.

You would have to do a proper comparison of sources, outputs, methodology (pre/post filtering used) in order to attribute something to the upscaling algorithm used, because often there are other factors contributing

[rrats]

Yeah but the current remastered music video for "womanizer" has quite a lot of rainbowing effects.

There is literally a smartphone shown in the video

[dellsam34]

Actually , some of the good upscaled music videos from that period (there are many crappy ones too) used digibeta masters or high quality intermediates as sources. Sources better than what VEVO official or YT had used, or the VHS or DVD versions. For example, the official version might have rainbows, dotcrawl and look really soft, but the intermediate used for upscaling that someone had access to - would already be rainbow free, clean and have more detail to begin with even prior to upscaling. So as expected, the results are better than the VEVO version even with traditional methods.

You would have to do a proper comparison of sources, outputs, methodology (pre/post filtering used) in order to attribute something to the upscaling algorithm used, because often there are other factors contributing

Yes I've seen a comparison between different versions from a youtube member, the Betacam ones kicked ass, they looked almost like an original 4K clip. I'm not into the nuts and bolts of this technology, My observations were just as casual viewer.

[ConsumerDV]

Nothing beats the original 35mm film This is not Britney Spears though.

[dellsam34]

What 35mm? You mean this? That has nothing to do with software restoration, the footage originally shot on film.

[ConsumerDV]

What 35mm? You mean this? That has nothing to do with software restoration, the footage originally shot on film.

I know. It was a joke. Sorry for hijacking the important discussion.

[poisondeathray]

Yeah but the current remastered music video for "womanizer" has quite a lot of rainbowing effects.

There is literally a smartphone shown in the video

Yes, there are some bad versions that probably use the same crappy source. But there are other ones that use a better source without rainbows. For many music vids, not just Womanizer, or "Britney". There are various fan sites that share less compressed version than YT as well.


official vevo, has rainbows
https://www.youtube.com/watch?v=rMqayQ-U74s

"1080p" webm version at 1920x1032

[Attachment 64282 - Click to enlarge]


somebody's upscale - No rainbows. I wouldn't say it's great, it's definitely not UHD quality, way too soft, smooth textures (partially from YT compression), weird unbalanced sharpening, but it looks slightly better overall than the VEVO version
https://www.youtube.com/watch?v=gXzKoeT8REY

"UHD" version downscaled to 1920x1032 to compare (framing is different, this video used director's cut master as the source)

[Attachment 64283 - Click to enlarge]

Yes I've seen a comparison between different versions from a youtube member, the Betacam ones kicked ass, they looked almost like an original 4K clip. I'm not into the nuts and bolts of this technology, My observations were just as casual viewer.

"almost like an original 4K clip" is a massive overstatement - You definitely cannot get "almost" or close to real 4K resolution (as in measurable lines of resolution) from SD sources, even under "perfect" conditions (downscaled from a real UHD source to start with, not betacam), even using "perfect" models (same training sources). The best you could hope for is real ~720p to <1080p with current GANs. And you almost never have ideal conditions. Supposedly there is better stuff around the corner such as "diffusion models" from Google, but I haven't been able to test them yet

[lordsmurf]

I'm not as hostile
I honestly am not sure why the hostility.

You're confused.

hos·tile =
- unfriendly; antagonistic.
- of or belonging to a military enemy.
- opposed

I'm entirely friendly.
Only trolls antagonize.
Opposition for the sake of being argumentative is ridiculous.

That doesn't describe this thread at all.
The crux of the situation is this: Topaz touts software as doing the yet-impossible, usurping terms (AI), and puffing itself up with self-fellating bellicose marketing. It's a video software version of a banty rooster. This extends beyond their video software, and applies to all of their half-baked software. Complete with cherry picked samples. Also noting that their seemingly glowing texts has lots of weasel phrases, to give them an out when quality of the output is awful. Which, sadly, is more common than not.

Know-nothing and know-little newbies are
- too easily fooled by marketing
- often trying to defend their purchases
- not objectively seeing the extremely problematic flaws; willingly wearing blinders
- sometime stubbornly refusing to acknowledge that other software can do the same or better
Unfortunately, that doesn't lend itself to wisdom, or knowledge growth, so they shall forever remain newbies that don't know much. And samples of their work is quite obviously bad, or cherry picked snippets.

That doesn't sit well with experienced video folks that see it as the BS it is. So we vehemently disagree.

ve·he·ment =
- showing strong feeling; forceful, passionate, or intense.

You're at a site about video, with a large population of pros and serious users. We are passionate about our hobbies, and careers. Unlike politics, some communities still prefer facts and truths. Not "alternative facts", nor misleading marketing. And these days, to hammer BS out of existence, it often does need to be forceful.

So hostile? No.
Vehement disagreement? You betcha!

The onus of proving Topaz software is good, or AI, is on Topaz. Thus far, it fails to do either.

[SaurusX]

Topaz touts software as doing the yet-impossible, usurping terms (AI), and puffing itself up with self-fellating bellicose marketing.

I don't think you know what the definition of AI is. It's not a self-aware Skynet.

[rrats]

I really love seeing lordsmurf stuck in time fight with newbies blinded by the marketing

[lordsmurf]

Topaz touts software as doing the yet-impossible, usurping terms (AI), and puffing itself up with self-fellating bellicose marketing.

I don't think you know what the definition of AI is. It's not a self-aware Skynet.

Apparently "AI" is a loosey-goosey wide term like "TBC", and can mean whatever anybody wants it to mean. I reject that notion. There has to be a standard, otherwise why even have the term?

Reminds me of "cloud", now a nothingburger term due to constant misuse in the past decade. AI is the next cloud, magical computer stuff for the masses. Nevermind if it's not actually textbook AI, or cloud.

Many folks literally write/say that Topaz software "works like magic". They have no idea what it's doing, nor care. Topaz said "AI", so that must be it. Magic.

But in actuality, it's really not any different than anything else currently available. Aside from often looking worse.

I really love seeing lordsmurf stuck in time fight with newbies blinded by the marketing

If by "stuck in time" you mean resisting the "current" trend to believe in whatever random crazy crap you want to (always seeded by those with ulterior motives), then yes. I prefer the days when facts mattered, when terms mattered. Mostly referring to science, medicine and politics here, but sadly it extends into other things like computing and video.

In this exact case, ulterior motive = marketing = buy our inferior BS software ("it has AI!!!!")

[litmus]

Lordsmurf, you reject the term hostile, but embrace the term vehement as a description for your arguments against TVEAI.
Vehement implies emotionality.
Emotionality implies a lack of impartiality.

Your arguments basically paint any TVEAI supporter as a newbie blinded by marketing, or (sometimes) a shill deliberately spreading misinformation. This comes across a bit paranoid.

"The proof is in the pudding" as you've said, the problem is that video quality to a certain extent is in the eye of the beholder. If you are heavily invested in one side of the argument, you are likely to take whatever evidence is presented and fit in into your current narrative rather than change your opinion. Any evidence you've found hard to discount you've dismissed as "tailor made examples".

Some of us have hundreds of hours of working with this software. We aren't all newbies. We aren't shills. The software has fans because it's producing good results for people.

I'll add that a quick google will tell you that Topaz Labs has 24 employees and a revenue of 8 million per year. They are hardly Apple or Time Warner or some other company with a huge reach. I personally can't recall ever seeing a single ad for TVEAI, I became aware of the software by positive word of mouth.

The idea that Janus Films using Topaz for their restoration of Inland Empire is a paid sponsorship is laughable.

What could be happening in this forum and is happening elsewhere on the internet is positive discussion about results obtained via Topaz, what settings to use for specific examples, what kind of pre-processing might be beneficial. There is a tiny bit of that in this thread but it quickly turns into a put down fest ringleaded by Lordsmurf because how dare anyone put money into Topaz's pockets when he thinks that AviSynth could do just as good of a job and how dumb you are for falling for marketing because you must be a newbie and aren't you stupid. It's childish and dumb.

This used to be my go-to forum for video stuff but I wouldn't bring any serious discussion here knowing that only certain points of view are to be supported.

[lollo]

how dare anyone put money into Topaz's pockets when he thinks that AviSynth could do just as good of a job

That's the key point!

As we said many times, in general there is nothing that Topaz VEAI can do that cannot be matched or beaten by AviSynth/VapourSynth; however, a full AviSynth/VapourSynth processing including Topaz VEAI as final step may have an interest in some cases.

[ConsumerDV]

Apparently "AI" is a loosey-goosey wide term like "TBC", and can mean whatever anybody wants it to mean. I reject that notion. There has to be a standard, otherwise why even have the term?

As in any new field you don't have strict definitions, but using neural networks for finding patterns is a good indicator of something being AI (or "AI" if you prefer).

What could be happening in this forum and is happening elsewhere on the internet is positive discussion about results obtained via Topaz, what settings to use for specific examples, what kind of pre-processing might be beneficial. There is a tiny bit of that in this thread but it quickly turns into a put down fest ringleaded by Lordsmurf because how dare anyone put money into Topaz's pockets when he thinks that AviSynth could do just as good of a job and how dumb you are for falling for marketing because you must be a newbie and aren't you stupid. It's childish and dumb.

This used to be my go-to forum for video stuff but I wouldn't bring any serious discussion here knowing that only certain points of view are to be supported.

@litmus, Jarrod1937, please do not drop out of the discussion because of naysayers. Non-productive noise can be ignored.

[rrats]

Ai does exist. Even vaporsynth has it. It's called VSGAN

[dellsam34]

We had dust and scratch removal software and character recognition for almost 3 decades now, AI is all over the place. it's just a matter of few more years before we can make perfect deep fakes.

[rrats]

That kinda scares me tbh

[lordsmurf]

video quality to a certain extent is in the eye of the beholder.
you've dismissed as "tailor made examples".
positive discussion about results obtained via Topaz, what settings to use for specific examples, what kind of pre-processing might be beneficial.

Quality is only up for debate from the artistic side, not the technical. Topaz barfs out artifacts at a high rate. There's no quality there, unless you're seeking to make an example of "what not to do".

It's really easy to see a cherry-picked example, especially when it comes from the manufacturer, or is compared against the norm.

Topaz needs to earn positive comments. So far, every example shown in this thread has been mediocre to completely craptastic. And yet, Topaz lovers gush over the middling results -- even when better results from Avisynth are shown. Youtube is the same. This is the definition of fanboy. Undeserved devotion, and only positive comments may be uttered.

how dare anyone put money into Topaz's pockets when he thinks that AviSynth could do just as good of a job

That's the key point!
As we said many times, in general there is nothing that Topaz VEAI can do that cannot be matched or beaten by AviSynth/VapourSynth; however, a full AviSynth/VapourSynth processing including Topaz VEAI as final step may have an interest in some cases.

Yep.

But not just that. Almost all of the Avisynth used doesn't claim or market itself as AI, and yet can do the same (or better!) as compared to the Topaz software. Nobody ever explains that, it's just glossed over. Again: "Topaz said so, so it must be true!" ie, gullible to marketing. The takeaway here is that very little, if any, is actually AI (or more likely "AI", the sorta-kinda-not-really type of AI).

I do agree, a combo of software is always interesting, complement. However, as many on both Reddit and Twitter are seeing, the often required pre-process in non-Topaz makes using Topaz somewhat moot, if comparable gains can be had with the existing program. Why buy Topaz for a tiny % quality boost? The kind that can be imperceptible, even to experienced eyes? No ROI there. So far, a lot of "Topaz is bestest ever!" samples have but middling at best, too often actually worse than the source when accounting for artifacts (ie, the user is fooled by sharpness alone).

Topaz is a show-me software.
Topaz has a history of not engaging customers, and short-lived software. If this still exists in 5 years, I'd be shocked.

As in any new field you don't have strict definitions, but using neural networks for finding patterns is a good indicator of something being AI (or "AI" if you prefer).

Hmmm...
On one hand, I can see that.
On the other, AI does exist, and isn't really a new field. Sure, new comparable, but not new in absolute terms. I guess that could only be argued if you put numbers on it, quantify what makes it a "new" field. And then further cement the posit by giving examples of other fields that were fast-and-loose on terms.

We had dust and scratch removal software and character recognition for almost 3 decades now, AI is all over the place. it's just a matter of few more years before we can make perfect deep fakes.

Yep. Dragaonspeak was called "AI" for a time, late 90s or early 00s. It wasn't.

[naimes’ə’dʒoû]

Ok, you just read the title.

To my understanding, the key point was that the expert people you mentioned stated that you can obtain an equivalent/superior result with AviSynth filters. I agree with them.

My post was for people who were looking to get the best results for their needs and I think I demonstrated by the provision of evidence that the product should be certainly be considered. Am I saying it is the only solution? No. Am I saying the product performed great for me? Yes. Have I proved my case? I believe so. Let the public decide for themselves.

Matt.

Agreed, and the comment you quoted is bizarrely categorical in the context of video upscale, or in the context of upscaling and interpolation in general. I've spend a lot of time experimenting with lots of different software, and results can vary wildly, and often surprisingly, with application. Sometimes, AI-based upscaling performs remarkably, but sometimes, even outside of their intended use-case, something like XBR or reverse-AA will either produce something that more closely approximates my aim or even more-faithfully approximates, at least in my estimation, the higher-fidelity reference I'm attempting to reconstruct.¹ Conversely (and far less rarely), AI upscalers can easily produce a result which I prefer over the results of pixel-art algorithms ('𝑃𝐴𝐴') in what appears to be an ideal application for the latter. Notwithstanding, this doesn't occur universally, and, even were I to be able to frequently achieve equivalent or superior results using the AI in this application, the mechanisms each employs to produce those results are so different from one another as to preclude a rational basis sufficient to support my asserting that the 𝑃𝐴𝐴 are redundant and wholly replaceable (with respect to results) by the AI tools I've used to successfully replicate or surpass their results (assuming, of course, that no trivial redundancy exists, i.e., that any coincidence in outcome is the result of the AI algorithm replicating the results of the 𝑃𝐴𝐴, not the AI using a 𝑃𝐴𝐴 itself, in which case the 𝑃𝐴𝐴 would effectively be a subcomponent of AI tool)—the vast majority of the variability in upscaling performance is a function of the particular subject matter being upscaled, and, even when that subject matter is superficially similar in fidelity or type of content, confounding variables, such as grain, noise, decimitation, and other artifacts, are abundant and, even small changes in these variables, subtle or nearly undetectable to a user, can produce in significant differences in outcome between results from the same algorithm, and expected variance increases, at least at a polynomial rate, as additional algorithms are considered, so that the outcome becomes more and more ambiguous (presumably, empirical study, employing reliable objective measurements on large data sets could help to mitigate uncertainty and facilitate predictability).

Finally, considering outcome, alone, is insufficient to a conclusion of redundancy—if we consider the instructive limit case that, regardless of how good an algorithm is, given enough time and resources, we would always be able to out-perform it by, e.g., hiring a team of artists who manually color each pixel to produce an upscaled image. While effective, I don't think the existence of this option implies that every upscaling algorithm is redundant.² The point is that, even if it is possible to mimic the results of one algorithm using another, doing so may require more skill, take more time, or be less reliable. I've used plenty of avisynth/vapoursynth filters, as well as MadVr for years, and there is definitely a noticeable difference in results between something like Waifu2×, DAIN, RIFE, Topaz Gigapixel or Topaz Video Enhance, and, e.g., SVP (spatial or temporal) or MadVr's NGU algorithm. For example, SVP has recently added support for a RIFE vapoursynth filter, which, although largely unrealistic to use in real-time with high-res video on modern hardware, clearly produces superior results (pre-upscaled—as mentioned, real-time isn't realistic right now) to the native SVP temporal interpolation (I suspect this was SVP's motivation for including the filter, despite its application being practically limited to pre-encoding). Perhaps, though, it is possible for a human to analyze some video and build a custom stack of filters that produce a more appealing result than these AI-style tools, but, at that point, we're starting to approach our team of artists we discussed earlier.

────────────────────────────────────────────────── ──────────────────────────────
¹ I should note that the set of upscaling problems we might expect to be best-solved by a pixel-art algorithm tend to be exceptional, but have distinctive characteristics these algorithms excel at addressing ( e.g., lower resolutions, with an indexed or otherwise limited color palette and the resulting aliasing, defined edges / demarcation between flat regions of different color, 🙰 𝒄.). These circumscribed conditions make appropriate applications more obvious and tend to motivate their use. They can surprise you in unorthodox applications from time to time, though; also, depending on our goal, it might be possible to prep an otherwise discordant image (e.g., by indexing color, cleanup, tracing edges) to be a good candidate for pixel-art algorithms.
² For completeness, the team of artists would be relatively expensive and slow when compared to an algorithm running on a modern processor.

[lollo]

Agreed, and the comment you quoted is bizarrely categorical in the context of video upscale

I did some experiment in upscaling, and showed the results (BTW you should do the same to prove your concepts, if any). That's all I care. Reading the rest of your long post is useless for me and my goals. Good luck with your upscale routines!

[rrats]

Here are two upscaled I did using the VSGAN (vapoursynth gan). I am pretty sure that the topaz software is in some way based on it, since they had never specified as to how does their software work (and also because of their sh##ty marketing). Ai is, in some way becoming useful in video filtering. I am pretty sure these kind of results wouldn't have been possible with only the basic filtering scripts



https://youtu.be/HYKLZOo3DM4
https://www.youtube.com/watch?v=OtcNMnPbyps

[lollo]

Thanks for your samples! You should also provide the original videos, so that anyone willing to do so, can experiment using different approaches and compare

[lordsmurf]

Cel animation is some of the easiest footage to upscale. It's almost entirely advanced sharpening, such as warpsharp. And yet, Topaz software still chokes on it.

On one of those Youtube samples above, around 1:06, you can see where some frames turned to blurry mush. That wasn't due to Youtube compression.

[lollo]

LS, s-mp is using VapourSynth, not Topaz

[rrats]

LS, that's flash animation and it's sourced from a 384px compressed file. Both of them.

https://sp.nicovideo.jp/watch/sm6157102 this is the place i sourced the stupid kasane teto video from btw. It's the highest quality source I know.

[lordsmurf]

LS, s-mp is using VapourSynth, not Topaz

I know.
I added line spacing, I guess it wasn't clear.

But still, mush at 1:06. The second point was it's easy to overdo filters.