This 4K video was recorded to a cassette tape
quite a few years ago - and yet for the most part it just looks like a meh phone recording, or
with a little tweaking, it looks pretty solid. That’s the power of AI at work. VHS tapes are a fun medium. Different lengths
of time available based on quality - very similar to video compression
today - no unskippable ads, it remembers where you left off, and forever,
and held everything from intimate family memories to bootleg music videos and TV
recordings to full commercial releases. But as much fascination as I may hold for this
medium, it’s quite fragile and the tapes you may have stored in a box are expiring fast,
if they haven’t already. I’ve been working on my VHS archiving guide I promised, but in
the meantime I got distracted searching for the best way to upscale and preserve
these captures in a more modern way. This asks the question: can Artificial
Intelligence help us preserve VHS captures? I’m EposVox, the stream professor and I’m a VHS
hoarder. Movies and TV shows from my childhood, but what I’m really interested in are tapes of
game previews and demos that were sent to press, storefronts, or included with magazines.
If you have any, please email me. There’s a lot of considerations when it comes to
the hardware side of capturing tapes, which we’ll cover in a future video, but regardless you end up
with a 480i analog-degraded image. It doesn’t look great blown up on a modern TV. What if we could
fix that? Or at least give it a little life boost? Upscaling is only part of the battle: deinterlacing in a pleasing way
makes a huge difference, as well. Today’s video focuses on comparing
3 ways of upscaling VHS captures: basic upscaling in a video editor if you just
ingested the footage at native resolution, normal bicubic upscaling with the QTGMC
deinterlacing algorithm for super smooth motion, using DaVinci Resolve’s GPU-accelerated “Super
Scale” tool with their new “Neural Network” AI deinterlacing algorithm, and lastly, a full
AI solution using Topaz Labs Video Enhance AI. Topaz Labs makes a full suite of AI tools for
upscaling and working with photos and videos. I’ve been playing around with Video Enhance AI since last summer and wanted to see
how it handles interlaced content. Well, thankfully they added a neural model
for this, called Dione. This is designed for interlaced content with two different versions
based on capturing cleaner DV tape sources or older analog tapes such as VHS and 8mm. We’ll
mainly be looking at the TV option, including the secondary “robust” algorithm - but I also ran
a DV tape capture through the DV model, too. The results here are quite interesting. You're
going to see better results all around with cleaner and higher-quality tapes, but these are
what most people will be capturing: Home videos from the 80s and 90s from VHS, 8mm transfers,
and so on. Things that aren't super high quality. For this, we're taking native 720x480 or 720x486
interlaced captures and scaling them by more than a factor of 4 to 2160p high for 4K. Plus,
deinterlacing. That's a lot to push this footage. Overall, I was honestly shocked to see that
Resolve's "Super Scale" amounted to nothing more than just scaling up the original footage
on the timeline. I didn't have high hopes, but I expected... something. It's so much
slower and seems to be doing something, but realistically the results aren't much
different than just scaling the footage up normally. For this, I had "Sharpness" and
"Noise Reduction" set to Medium, the default, but realistically not much would
change with the other settings here. It didn't matter if I was scaling animation,
a higher-quality tape, or an old 8mm transfer, Resolve Super Scale looked marginally
sharper than a native scale, but that's it. I still had to test it, and
even setting "Sharpness" and "Noise Reduction" to High.... didn't make
a big enough difference. There's s SLIGHT, SUBTLE change compared to Medium, but still
not competitive with the other options. StaxRip's Bicubic scaling with QTGMC deinterlacing
actually provided really compelling results in the animation tests. There were some situations
where the perceived aliasing or "jaggies" on lines were almost entirely removed with this scale, with
noise reduction competitive with the AI upscaling. The Dione AI upscaling and deinterlacing
provided a cleaner image on the animation footage, but did not handle line
artifacts as well. I'm surprised. ***[Brilliant Ad]*** If you want to develop your own
AI-based tools like this one, you'll have to learn to code, and learn
about neural networks. Where do you start? If you're interested in learning computer
science, programming, or other STEM topics, you could watch more educational videos
on YouTube - you're already here, cool. But in order to learn these kinds
of things, you have to actually DO it. Brilliant is a website and app
built off of this very principle: You learn best while doing and solving in
real-time. Jump right into solving problems and be coached bit-by-bit until you've learned a
new STEM subject before you've even realized it. Pick a course you're interested in, get started,
and don't worry about memorizing long formulas or fact sheets. That's not how you learn a skill.
Working through problems, reading explanations for mistakes or places you get stuck on, and
learning at your own pace - that's education. Brilliant has something for everybody - whether
you want to start at the basics of math, science, and computer science, or dive right
into cutting-edge topics like Neural Networks - the stuff building tools like this.
If you'd like to join me and a community of 8 million learners and educators today, click
the link in the video description or head to brilliant.org/eposvox - the first 200 who do so
will get 20% off the annual Premium subscription! ***[end Brilliant ad]*** Switching to live action footage,
though, and even other animation clips, and generally speaking the Dione AI
upscaling provides smoother results. Less noise, clearer shapes and colors, and so on. It's not
magic, you're not getting detail that was never present, and the more degraded your tape is,
the less you have to work with, but this does look significantly better blown up to a 4K TV
than any of the other options, in most cases. I will say that with the 8mm car tape, I found the
"Robust" version of the Dione AI to look the best, the standard one produced a
bit of ghosting - but this was the only sample I found that result to matter. Using the Dione DV AI on my widescreen miniDV
tape produced remarkable results... Like WOW. Noise is mostly gone, colors are
back a bit, sharpness is far improved. This is beyond impressive, especially compared
to all of the other options. Add a bit of a color grade and this isn't terrible-looking footage for
what it is, considering the terrible lighting. Letting curiosity get the best of me, I also
decided to test the Dione DV AI on a couple HDV clips I had from some experimenting back in 2017.
Something with better light and less baked-in noise. So here we're taking 1440x1080i footage
and deinterlacing while also quadrupling up to 4K. Honestly, the AI upscaling doesn't do anything
magical here, either. Deinterlacing is done quite smoothly, but the results via
StaxRip and QTGMC are mostly identical. Also, Video Enhance AI kept crashing on
the MPEG capture for one of these clips. In fact, there were quite a few issues that
I found with Video Enhance AI for myself, personally. On top of it really not liking my MPEG
HDV files - which a simple copy to AVI via FFMPEG fixed - you have no control over the aspect
ratio. By default if you scale a 4:3 image, it will crop to fill frame, but you have
an option to disable that. Cool. However, after that, you have no control over how to
handle the aspect ratio. So one of my HDV clips was detected properly in 16:9 via a wider pixel
aspect ratio since the footage was 1440x1080. But the second clip, same resolution and pixel
aspect ratio, was detected as 4:3, so I had to stretch it out to fill the proper ratio after
the fact. Plus, the actual 4:3 scaled results here came out a tad wider than my normal scaling
results in Resolve, though closer to StaxRip. I'd honestly lean on suggesting Video Enhance AI
might be more correct here, but it's still weird. Lastly, framerate detection was... all over the
place. Since I'm working with interlaced clips specifically here, that's understandable, but
even between two different PCs with the same version of the app would have different
results between 29.97fps and 59.94fps. And then you have the MPEG clip, which I copied
to an AVI container, which it decided was 120FPS and exported as such for some reason?!
Give me framerate controls, please. Those were quite a few complaints, but really -
I've been playing around with this software, and even using it in videos without saying, for nearly
a year now and it's been a blast. You can do some wacky stuff with Video Enhance AI and it has
models for computer graphics, real-life footage, faster stuff, slower stuff, lots of details to
play around with. If you're a Digital Foundry patron, they've been using it for upscaling old
lossless game reveals and such that they have. There's also a trade-off, where Video Enhance AI
takes exponentially longer than normal solutions, requiring a decently-high-end graphics card
and measuring render times in seconds-per-frame rather than frames-per-second, but
doing hard work takes time, I suppose. To answer the main question, though: Can
AI help preserve VHS captures? I'd say yes. The natural-looking noise removal, the extra
enhancement of detail and sharpness, without doing anything too wonky, along with near-perfect
deinterlacing most of the time means that my old family videos - or even just fun, commercial
tapes - can look better than ever blown up on the big screen. Sure, you're technically losing
"accuracy" from how the tape actually looked - but in many of these cases, that's not accurate to
how it was originally recorded anyway either. Thanks for coming along on
this journey with me. Links to StaxRip and Video Enhance AI will
be linked in the video description, if you're interested. Let me know
what fun you've gotten into with your tape upscaling in our Discord server at
Discord.gg/eposvox. Remember: Be kind, rewind.
