The Art of Calm video art being played on the ultra-wide theater-sized screen at UHealth Doral.
Walk past the lobby video wall at UHealth Doral Medical Center and you might catch a wave rolling in slow motion across a stretch of pale sand. Walk past the one at UHealth SoLé Mia Medical Center and it could just as easily be morning light drifting over the Everglades, or a mountain stream easing into a still, glassy lake. None of it repeats in any way you would notice. There is no music, no ticker, no cable news, just light and time—moving at the pace of a very deep breath.
You might wonder how someone captured such breathtaking footage. The answer is, they didn't. Every scene is AI-generated as part of "The Art of Calm" initiative, UHealth's new custom digital video experience at its newest clinical facilities. Each video was built around what a specific location actually asked for, not pulled from a generic template. Together, they replaced a licensed video contract that had quietly cost the health system thousands of dollars a year.
A Simple Idea, A Complicated Build
On paper, the concept seemed straightforward: take the video walls that had been running licensed stock footage on a recurring contract, and replace them with content generated in-house using AI, so the health system keeps full ownership of the content, can tailor it to each location, and stops paying an outside vendor year after year.
The clinical case was just as important as the financial one. Hospitals can be stressful environments for patients and families, particularly during periods of waiting and uncertainty. The idea behind "The Art of Calm" was to give those spaces something calmer to look at: slow, predictable, non-narrative motion instead of a static image.
A small group, part of University of Miami Information Technology & Systems (ITS) AI team, set out to build a library of ambient videos, each with its own nature-inspired visual theme, like beach waves, underwater scenes, and the Everglades—produced for both the Doral and SoLé Mia displays. None of the videos was a copy-paste of another, and each theme was chosen and shaped around what that specific location stakeholders were looking for, from the mood they wanted to the pace and color palette that fit the space. Every clip, whatever the scene, was generated by Veo 3.1, an AI video model, and stitched from 8-second segments into a seamless, silent, infinitely repeating loop.
The breakthrough isn't that AI can generate video, it's that the cost of creating studio-quality content has fallen dramatically. "What people don't see is the work between the prompt and the final result. These models are getting better every month, but producing something that is seamless, reliable, and ready for a real-world environment still takes far more than pressing a button," said Dr. David Wayne, Chief Artificial Intelligence Officer.
An example of a photo-realistic AI-generated video.
When "Simple" Wasn't So Simple
The AI team's first assumption was that generating a calming ambient scene would be the easy part, and the real work would be the deployment. It turned out to be almost exactly backwards.
Creating an eight-second AI-generated video was one thing. Turning a series of those clips into a seamless experience that could play continuously for more than ten minutes was something else entirely.
Early tests revealed inconsistencies that became more noticeable as clips were extended; a limitation grounded in video generation models. A camera that should have moved at a slow, steady pace could suddenly speed up or change direction from one segment to the next. Water, clouds and other elements did not always move naturally, and visual quality could begin to deteriorate after several consecutive generations, in some cases after just 24 to 32 seconds.
The team also encountered issues with video artifacts, including a phenomenon called "ghosting." During testing, moving objects could disappear, change shape, or look different from one clip to the next. Motion blur could compound the problem when a frame from one clip was used as the starting point for the next. Brightness and color presented another challenge, sometimes shifting noticeably between clips.
For a traditional video, some of those differences might go unnoticed. For "The Art of Calm," they mattered. The goal was to create slow, continuous motion that blended naturally into the environment. A sudden change in speed, direction, image quality or brightness could break that sense of continuity. And any minor flaw gets amplified when displayed on a large screen.
Addressing those issues required a combination of iterative experimentation. The team tested different approaches to controlling camera movement, such as seeding the video generation with carefully curated start and end frames and slowing down videos based on measured motion speed between clips with custom computational methods. Clips that did not meet the desired quality were regenerated, while subtle transitions helped minimize visible changes in brightness between segments.
There was no single solution that worked for every scene. A technique that improved one video did not necessarily work for another, and some challenges could only be addressed through repeated generation and manual review. Over time, that experimentation gave the team a more reliable process for determining which approaches worked best for different types of scenes.
The displays themselves added another layer of complexity. Veo3.1 natively outputs video in a 16:9 format. The screen at Doral is ultra-widescreen at 30:9, and SoLé Mia is closer to 5:4. Creating prompts to instruct the model to output in those aspects ratios was unreliable, so we had to use other techniques such as feeding it an image in the correct aspect ratio as the first frame of the video. Moreover, with each facility having different screen formats, we had to ensure the same visual concept worked in two different aspect ratios. The team explored whether one version could simply be cropped or reformatted for the second display, but the resulting loss in quality was too noticeable. Instead, the videos had to be generated separately for each format—effectively creating two productions for every scene.
