Topaz Video AI 1.7.2 - Outstanding Performance Improvement in Real-World SD-to-UHD Upscaling

After installing Topaz Video AI 1.7.2, I wanted to share my experience because the performance improvement has been nothing short of remarkable.

My usual workflow consists of a double inference upscale pipeline:

Proteus Standard (SD → FHD) + Proteus Natural (FHD → UHD)

With version 1.7.1, I typically achieved between 8.1 and 8.7 fps. After updating to version 1.7.2, performance immediately increased to 17.1 to 19.9 fps, more than doubling my previous processing speed while maintaining excellent image quality.

Even more impressive, using my MSI Titan 18 HX with its Intel Core Ultra 9 285 HX and its RTX 5090M, the software now appears to use hardware resources much more efficiently:

  • GPU utilization: ~66%
  • GPU temperature: 63-68°C
  • CPU utilization: ~57%
  • CPU temperature: 72-76°C

The system remains stable, cool, and responsive during long processing sessions.

As a practical example, I am currently restoring the complete TV series The Last Ship. The first episode, with a runtime of 43 minutes and 14 seconds, was processed in only 1 hour, 4 minutes and 47 seconds. For a double-inference SD-to-UHD workflow, this is by far the fastest result I have ever achieved.

What makes this particularly noteworthy is the type of content involved. The Last Ship is an extremely demanding test case for any super-resolution model:

  • Digital camouflage uniforms
  • Highly detailed military equipment
  • Smoke and atmospheric effects
  • Explosions and debris
  • Fast action sequences
  • Detailed close-ups of faces
  • Complex wide shots with fine textures

In my opinion, this kind of material is often far more challenging than many period films or slower-paced productions.

Over time I have also simplified my workflow. I completely stopped using Recover Details because, although it can occasionally restore useful information, it also tends to bring back source defects such as compression artifacts, halos, and unwanted noise. Eliminating this step has improved both processing speed and image cleanliness.

I also stopped adding artificial grain. Working entirely in a 10-bit pipeline provides smooth tonal transitions and natural-looking skin textures without the waxy appearance that often motivates users to add grain in the first place.

The result is a cleaner workflow, faster renders, lower hardware stress, and excellent image quality.

I know Topaz products sometimes receive criticism, but I believe it is equally important to acknowledge the progress being made. As someone who regularly restores large collections of SD content to UHD, I can clearly see the substantial work the development team has accomplished in this release.

For me, version 1.7.2 feels like an early Christmas present.

Congratulations to the Topaz Labs team, and thank you for the impressive improvements.

Kind regards, Vincent.

We’ll see about that…

In fact, the dramatic improvement in rendering speed was due to the Auto option in the second processing pass!

As soon as I was able to select x2 as usual, times went back to normal! It was too good to be true!

I also stopped add grain into Topaz into my workflows. Adding grain should be the last step if you want have it, you can do it outside Topaz later. FFmpeg is limited adding grain, but this one give me suprisly good fine one! “noise=c0s=4:c0f=t+u:c1s=1:c1f=t+u:c2s=1:c2f=t+u” Note: Grain Size also pends on your Video gets scaled when play, watching on TV big Screens or not.

I also stopped years ago using “Rec. (Orig.) Details” as already wrote above it does not really recover, it pronounce existings including artifacts and is invasive.

I agree ! :+1: