A completely rewritten version of Deep Desktop
After about a year of work, Deep Desktop 2.0 is here. The objective of this new version is to address many of the pain points that Deep Desktop has had since its release. It comes with a complete redesign of how the menus look and work, a switch to the Godot game engine, and a completely redesigned depth estimation system. This replaces the separate installation systems for each GPU brand, which were unstable and took up a lot of disk space with duplicated files.
For existing users
If you are an existing user of Deep Desktop 1.1.0 or the 1.2.0 beta, the new system is entirely incompatible with the old model installations, so you will need to reinstall your models. Some models are already packaged with the new version, so you may only need an initial installation of around 1 minute. Subsequent updates to the model runtimes should only take a few seconds to apply. You can uninstall the old Deep Desktop model installations using the "Manage models" menu if you wish.
If you want to go back to the older versions, they are still available as beta versions in the Steam menus under the names "1.1.0" and "1.2.0". The old and new model installations should be able to coexist.
New menus
In 2.0, the menus have been completely redesigned to provide a more straightforward experience when you first open the program.
The program now centers around three main menus. Version 1.1.0 had a much larger number: the Quick Start menu was added in a hurry, followed by the Video menu, which made navigation increasingly confusing:
Display mode menu: The first menu you encounter when opening the app lets you choose how you want Deep Desktop to be displayed: VR, SBS, Anaglyph, or 2D Display. For SBS and Anaglyph, a second menu shows your screen setup and lets you select which screen the app should use for full-screen display.
Setup menu: After making your selection, you arrive at the Setup menu, which brings together all the settings that used to be scattered across the Quick Start, Video, Setup, and Options menus. It has simple and advanced modes so that new users are not overwhelmed by the number of options. To get started with the default settings, just press the "Install" button. Once the model files have been installed, the button changes to "Use selected source", which takes you to the depth screen. If you want additional settings, select the "Advanced Options" checkbox to see the full range of settings available. Deep Desktop now also has a built-in file and folder browser that you can use with your VR controllers, instead of having to switch to your mouse.
Depth Screen: You then arrive at the Depth Screen, which has a much simpler interface than in version 1.1.0. Here, you will find the movement controls and an Options button that opens a menu with settings for depth estimation and the screen. A few new options have been added, notably horizontal and vertical crop positioning and the new export buttons. In video mode, you can now load MKV files and switch between audio and subtitle tracks.
New inference engine and new game engine
Deep Desktop has switched from Unity to Godot for many reasons. The two main ones were performance and the ability to modify the underlying game engine code to implement the new depth estimation system.
This new depth estimation system uses the Vulkan graphics API instead of CUDA or DirectML. Replacing the old system was very important for several reasons:
Overreliance on third-party update cycles: When Deep Desktop launched, it broke on the first day because of an update to NumPy, a popular Python package, that had been released a few days before launch. Then, when NVIDIA's RTX 50-series GPUs launched, there was a one-month wait before a compatible version of PyTorch became available for those cards. Support for older GPUs, notably the GTX 10-series, is also being phased out in those systems because they are considered too old to support. Finally, the PyTorch DirectML package has not received a new release in almost two years. I am also increasingly concerned about supply chain attacks within the Python package ecosystem. It is therefore important to me to reduce the risk of third-party software updates breaking Deep Desktop.
Performance lost in communication with the game engine: The old system relied on a slow communication pipeline to send the screen capture or video feed to the Python depth estimation system and receive the results back. Around 10–15 ms per depth estimation was spent transferring data. Exceeding 40 depth estimations per second was effectively impossible, even with high-end GPUs. The new system uses Vulkan–DirectX 12 interoperability on Windows, while Linux runs entirely on Vulkan. It keeps the data needed for depth estimation entirely on the GPU, from the moment the screen capture is produced to the moment the depth screen is deformed using the estimation result.
Much smaller installation footprint: One major issue with the CUDA installations was that each required around 3 GB of NVIDIA's proprietary libraries. This took up a lot of space if you wanted to try multiple depth models. In the new version, the model weights account for most of the space used, as the runtime that performs depth estimation has been reduced to a few MB per model.
Offline installation system: Deep Desktop now ships its inference runtimes directly, making offline installation possible for the bundled models while keeping the Deep Desktop package size at around 1.6 GB. For those of you with unreliable internet access or download restrictions, this makes some previously inaccessible models available. The new installation folder contains an offline-payload directory. If Deep Desktop fails to download a model file, it will display the filename. You can then download that file from a third-party source and place it in the offline-payload directory for Deep Desktop to use instead.
Proper GPU selection: If you have multiple GPUs, the new system lets you select which one to use for depth estimation before starting. This includes the integrated GPU found in most modern CPUs.
Linux version
Deep Desktop is now available natively on Linux, with the same redesigned interface and Vulkan-based depth estimation system as the Windows version.
The main difficulty in developing the Linux version is the variety of distributions and the different system components they provide. Building something that works across them requires finding a common baseline. The switch to Godot was instrumental because it allowed me to adjust the game engine's build process to match the Linux versions being targeted.
I have tested Deep Desktop on the following distributions, which I hope cover a representative range of the preinstalled libraries, desktop environments, and display servers (X11 and Wayland) found in recent Linux distributions:
Ubuntu 22.04
Ubuntu 24.04
CachyOS 260809
Arch Linux, August 2026, with the KDE Plasma desktop environment
Fedora 42
Bazzite (I don't have the exact version)
If you are using an NVIDIA GPU, please make sure that NVIDIA's proprietary drivers are installed, as the Nouveau driver did not work properly in my testing.
Conclusion
This version is an almost complete rewrite of Deep Desktop, both in how it works and how it is used. Despite the testing I have carried out over the past few weeks, I may have missed critical issues that did not occur on the hardware and operating system combinations available to me. If you encounter any issues, please visit the forums (https://steamcommunity.com/app/2947590/discussions/) so that I can look into them.
Finally, I want to thank everyone who showed interest in what started as a small side project and grew beyond my expectations. Version 2.0 is much closer to what I originally envisioned for Deep Desktop but did not have the time or resources to build at launch. It has taken a lot of work to get here, and I believe it provides a more reliable foundation for the future.