{"id":38738,"date":"2025-08-09T13:53:20","date_gmt":"2025-08-09T20:53:20","guid":{"rendered":"https:\/\/www.pugetsystems.com\/?post_type=case_study&p=38738"},"modified":"2025-08-09T13:53:21","modified_gmt":"2025-08-09T20:53:21","slug":"custom-machine-learning-solves-the-impossible-for-dune-2","status":"publish","type":"case_study","link":"https:\/\/www.pugetsystems.com\/featured\/custom-machine-learning-solves-the-impossible-for-dune-2\/","title":{"rendered":"Custom Machine Learning Solves the Impossible for Dune 2"},"content":{"rendered":"\n
\"Banner<\/span>
\n\t\n\n
\n
\n

Custom Machine Learning Solves the Impossible for Dune 2<\/em><\/h1>\n\n\n\n
<\/div>\n\n\n\n

Can you make their eyes blue? All of them. In all 1000 shots.
Oh yeah, and this is for Dune 2<\/em>, so we need it done yesterday.<\/strong><\/p>\n<\/div>\n\n\n\n

<\/div>\n<\/div>\n\n<\/div>\n<\/div><\/div>\n\n\n\n
<\/div>\n\n\n\n

Such was the challenge placed before Academy Award-winning VFX Supervisor, Paul Lambert. Paul is no stranger to creative challenges and has the track record to show for it. Having won Academy Awards for Best Visual Effects for such blockbuster films as Blade Runner 2049<\/em>, First Man<\/em> and Dune 1<\/em>, his Herculean efforts on this latest challenge for Dune 2<\/em> won him his 4th statue!<\/p>\n\n\n\n

Puget Systems is very proud to have played a role in his latest achievements in Dune 2<\/em> – not only providing the hardware, but working side-by-side with Paul to dissect the challenge, brainstorm a workflow, troubleshoot the challenges along the way and ultimately deliver a workstation that would perform precisely as he needed it to. And as Paul will attest, precision is everything when it comes to high stakes, complex visual effects in Hollywood.<\/p>\n\n\n\n

We spoke with Paul Lambert about his latest achievement with Dune 2<\/em>, and the path it required to get there.<\/p>\n\n\n\n

\u201cI had already begun prepping for Dune 2<\/em> with the understanding that one of the larger, more tedious and time consuming tasks was focused on a group of people central to the story in Dune – the Fremen.\u201d <\/p>\n\n\n\n

For those who might not be familiar with the Dune series, the Fremen are a group of people in the fictional Dune universe who inhabit the desert planet Arrakis – also known as Dune. And their distinguishing characteristic: their eyes are blue. Not just their iris, but their pupils and the whites of their eyes are all a slightly different shade of blue.<\/p>\n\n\n\n

<\/div>\n\n\n\n
\"Frame<\/figure>\n\n\n\n
<\/div>\n\n\n\n

Paul continued, \u201cThe amount of work required to create the effect – which started in Dune 1<\/em> – was astronomical. In the first film, to accomplish that effect, I had 260 shots manually rotoscoped in Nuke. And to be clear, I\u2019m not just talking about the outline of the eyes – I literally had to have hand-roto of the outline of the eye, the pupils, iris and around the whites of each eye! From these 260 shots, I created Open CV code to handle face detection, and tested it on each face. In the end, I ended up with 30,000 pairs of eyes for the Fremen.\u201d<\/p>\n\n\n\n

\u201cThere was simply no way to do all of this by hand. So I started to tinker. I have some experience building workstations, so I knew some basics about how to build a system for visual effects work, but not too much beyond that. It\u2019s also important to know that during this time, Machine Learning was also exploding, and the use cases for the kind of work I was facing were very compelling. So I had to take a closer look to see if and how it could play a role in streamlining this workflow. After all, 30,000 pairs of eyes represents a data set on which machine learning can be trained – in particular data sourced only from the first movie! I had a rudimentary system to help me start testing out machine learning to see if it would work for this challenge.\u201d<\/p>\n\n\n\n

<\/div>\n\n\n\n
<\/span>
\n\t\n
\n
\n\t

There was simply no way to do all of this by hand. So I started to tinker… <\/span>Machine Learning was also exploding… So I had to take a closer look to see if and how it could play a role in streamlining this workflow.<\/em><\/p>\n <\/div>\n Paul Lambert<\/cite>\n <\/blockquote>\n\n<\/div>\n<\/div><\/div>\n\n\n\n

<\/div>\n\n\n\n

\u201cRemember, Dune 1<\/em> had 260 shots – but Dune 2<\/em> had over 1,000 with many many more Fremen whose eyes required roto. It was a much bigger exercise to pull this off. I asked myself, is there a way to leverage what we created for Dune 1<\/em> and simply transfer it to Dune 2<\/em>?\u201d<\/p>\n\n\n\n

\u201cIt was around that time also that Nuke had just introduced \u2018CopyCat\u2019 and it was getting much better, but it was optimized to run on multiple GPUs and multiple CPUs. As I was testing it all out, it was becoming an obsession! It just worked! But I needed to be able to fit all of the sequences, all of that work into the system. This is where I needed help, I needed \u2018more system.\u2019 So I called Puget Systems.\u201d<\/p>\n\n\n\n

<\/div>\n\n\n\n

Puget Systems Steps In: The Specs!<\/h2>\n\n\n\n
<\/div>\n\n\n
\n\t\n\n
\n\t\t\t\n\n

The Puget Systems team has invested significant time and resources in developing a deep understanding of professional workflows in high-end VFX productions. And Paul\u2019s case was in perfect alignment with the consulting team\u2019s expertise. Since Paul was working in Nuke, and more specifically, CopyCat, he not only needed a system with plenty of GPU power, but he also needed significant RAM to be able to process the tens of thousands of Fremen eyes he had created. Here is the system the team came up with, specifically tailored to Paul:<\/p>\n\n\n\n