Monday, 26 October 2020

Topaz Video Enhance AI Is Awesome, but Essentially Early Access Software

This is now.

Note: This is not a full-fledged formal review and comparison of Topaz Video Enhance AI against other in-market applications, but a discussion of one particular application’s strengths and weaknesses.  Model quality and capability are still changing significantly from version to version.

Over the past 10 months, as I’ve worked on upscaling Star Trek: Deep Space Nine and Voyager, I’ve relied on one piece of software above all others to perform the work: Topaz Video Enhance AI ($199.99, for Windows and Mac). I have recommended this product both indirectly and directly, and I’m going to continue to do so. But I need to make certain it’s clear what you are getting into.

Up until September, virtually all my focus had been on improving the quality of my video pre-processing steps. After publishing “What No Fan Has Seen Before,” I decided to turn my attention to the upscaler side of the equation.

Here’s the good news: Topaz Video Enhance AI is, hands-down, the best AI video upscaler I’ve tested. Some of its models are tunable and it can improve a broad range of video. I’ve seen it breathe new life into Grateful Dead shows, old VHS tapes, Star Trek, Stargate: SG-1 and a number of other types of content.

I cannot say that the entire community has been happy with the pace of development, but given the complexity of video editing software and the need to keep continually improving the underlying AI, I feel like things have been moving along at a reasonable clip. AI video processing is an incredibly new market, and Topaz is way out in front of any of the video editors I’ve tested (though I’m always happy to hear suggestions for other programs to test). To the best of my knowledge, TVEAI is the only application that does what it does as well as it does it.

I don’t have a lot of new footage to show just at the moment — I’ve been working on Voyager, but I’m not ready to show work. What I’ve got on tap is a version of the DS9 credits I built immediately after “What No Fan Has Seen Before” went live. This footage was produced using AviSynth+ for initial upscaling and TVEAI for processing. For those of you wondering where the tutorial I promised is — I’m still working to hammer out a workflow that’s going to deal with the full range of the show and improve the first seasons more than the current method does.

But users should be aware that there are some inconveniences to Topaz Video Enhance AI as well. First, it is not kind to system stability. It never fails to start properly when a machine is fresh from reboot, but if you’ve been using your GPU for other applications, the program may need a reboot to work properly. Stopping and starting it again pretty much always works the first time. It probably works a second time. You’ve even got a solid shot at three times. Asking it to render more than three videos in a row? That’s pushing your luck. Eventually, either the AI engine will fail to initialize or the application will crash and a reboot will be required.

Topaz Video Enhance AI seems to become unstable faster if other applications like StaxRip are running multi-threaded AviSynth+ encodes at the same time. It is not overly fond of sharing the GPU, though this behavior has improved in recent months. I’d recommend rebooting every 1-2 days to minimize the chances of a crash, especially if you’re running multiple videos in that time. Loading videos increases the chance that the next video will cause a crash, though the software can also crash on very long encodes. Preprocessing is often required for maximal upscale effectiveness, however. Personally, I just grit my teeth and reboot a lot.

One of the ways Topaz has addressed this instability issue is by creating an AutoSave mode that reloads your previous video and remembers (mostly) what frame you were on. This mitigates the hassle of rebooting four times in a single day when upscaling a great many short clips.

Model Quality Is Still a Moving Target

The quality of each AI model varies from release version to release version. One of the suggestions I’ve made to the company that it’s said it will implement is making it easier to find older versions of the software. This can be a necessity when testing to see if a given model works better in an older or a newer version.

The reason that model quality can vary is that the company is continuing to refine and train its models. There are a lot of moving parts in the equation, and Topaz has been trying to improve their application on multiple fronts simultaneously, which means some models have gotten worse and then been rolled back or repaired over time. Some spots that were hard for Topaz to upscale in February are still hard in October.

If you’re interested in this software, take advantage of the free trial to make certain its models can address your content before pulling the trigger on it.

Be advised that Topaz Video Enhance AI is not a magic bullet. Here are two very different versions of Benjamin Sisko:

Season 2. This is upscaled. It’s an older shot, and I’ve improved on it, but I haven’t fundamentally transformed it.

This is from Season 2, which appears to have been transferred to DVD in a manner intended to preserve every bit of the VHS “Super Long Play” viewing experience. Here’s Benjamin Sisko from Season 4:

Season 4 Captain Sisko. This is purely the difference in how the show was mastered to DVD, nothing I’ve done.

I’m still working to discover if the top frame can be fixed, but the fact that the image doesn’t look great is not TVEAI’s fault. There’s a limit to what the application can handle, and even some commercial video isn’t very good right now.

Topaz’s Ownership Model

Topaz has an interesting software model. When you buy Video Enhance AI, you get access to the application as it exists today and all future updates, major and minor — for one year. If you want updates thereafter, another year is $49.99. It is not clear if you are allowed to skip years while keeping the $49.99 price, or if you must keep resubscribing every year or buy the entire program again.

In the event that you stop paying the yearly update fee, you keep full access to the program as it exists the day you bought it. The question of how good of a deal this is really comes down to how rapidly Topaz improves the application. So far this year, it really has improved a lot — but it’s also got a long way to go.

Anyone considering this application should download the free trial first and test how it performs on your content. Be aware that you might wind up climbing under the hood with a wrench to bang on the video before you run it through the program. If you need to deinterlace footage, for example, that needs to happen outside the upscaler.

Topaz Video Enhance AI is effectively early-access software — in effect, if not in name. I don’t think there’s anything wrong with the early access model, especially when the entire market is so new, we don’t have effective competition yet. But I also don’t want to plug this application as being astonishingly effective for my purposes without also noting that this program is not, strictly speaking, all that newbie-friendly. The MP4 encoding option tends to drop audio, which means it’s best to learn how to use FFMPEG to reassemble video from constituent frames. Use Topaz Video Enhance AI to render out to PNGs or JPG, and there’s no problem at all (this method also recovers more easily, since interrupted MP4 encodes can’t be resumed).

Topaz Video Enhance AI is a unique, interesting application and I’m eager to see where it goes, but I don’t want to paint the picture rosier than it ought to be. There’s not so much a learning curve around the program as there is a learning curve around the other things you need to do to use Topaz Video Enhance AI at peak capability. Either way, be aware it exists. If your content aligns well with its capabilities, the results can be incredible. If it doesn’t, it might not be worth investing in until the app gets some additional development time under its belt. Take advantage of the free trial and allow for the idea that you might wind up astonished. Also, allow for the idea that you hope to be astonished in about two years, once all the kinks are worked out.

Now Read:



No comments:

Post a Comment