I Am Sam (2007) - AI Upscale Project

[pushler]

Hi everyone.

A little backstory.

About three years ago, I got into Korean dramas. I like this genre because it doesn’t contain immorality or vulgarity. They are like romantic fairy tales for grown-ups.
Last year, I watched some series with Park Min-young: «What’s Wrong with Secretary Kim», «Healer», and «Sungkyunkwan Scandal». I loved her acting, and because of that, I wanted to see her very first lead role in «I Am Sam». This drama has been available on YouTube for many years. However, I couldn’t find a version with acceptable quality anywhere online, which made it hard to enjoy.

That got me interested in ways to improve video quality. Recently, video upscaling with AI has become a hot topic. However, most suggested solutions were either unsatisfactory or too expensive.

I decided to try some projects from GitHub, as well as Topaz Video AI. Topaz delivered quite satisfying results, and I wanted to share my upscaled version.
I’m curious—has anyone else explored this topic?
Has anyone tried upscaling full-length movies or TV shows?
What challenges have you faced?
I’d love to hear about your experiences!

[pushler]

As I mentioned earlier, I was looking for the 2007 drama "I Am Sam". It’s available in the TPB database with a resolution of 1280×720. However, the total size of 7 GB for 16 episodes made me doubt the quality—it likely has a very low bitrate. After a month of trying, I still couldn’t download it.

I searched on other sites and found the best available version here:
https://rutracker.org/forum/viewtopic.php?t=2586592.
Its resolution is 624×352.

I decided to upscale this drama. First, I extracted a short fragment to quickly test different AI enhancement models. After finding the best one, I applied it to an entire episode.

This is where I ran into a challenge. The test clips I used were short (15–30 seconds), shot in a single location with stable lighting and a consistent background. But a drama is much more complex.

Scenes vary widely—daytime, nighttime, bright, dim, close-ups, long shots, fast and slow movements. Sometimes, facial tones blend into the background, making it hard for AI to recognize edges. This led to issues like blurred or even missing facial features (e.g., the nose or neck).

Another problem was that settings optimized for close-ups often made distant faces look cartoonish and unnatural. Some call this effect "AI hallucinations." The main challenge was to find parameters that worked well across all types of scenes.

After a month of testing, I finally found the best balance. In the end, I upscaled the entire drama 2× to a resolution of 1248×704.

---

I uploaded the upscaled version here:
https://thepiratebay.org/description.php?id=78101283.

If you're interested, you can compare the original and enhanced versions by downloading a single episode.

I used Topaz Video AI 5.3.3 for this project. There’s still a lot of room for improvement in AI upscaling, but maybe this work can serve as an example of how the technology is evolving.