Desktop2Stereo: a 2D Desktop to 3D Stereo App

中文版本

A universal real-time 2D to 3D App that supports AMD/NVIDIA/Intel/Qualcomm GPU/Apple Silicon devices on Windows/Mac/Ubuntu, powered by Depth Estimation AI Models

Alternative Download Link

• Hugging Face Bucket
• Quark NetDrive
  Access code: 1vcn
• Baidu Netdisk
  Access code: mr64

Video Tutorials

• Windows
• MacOS
• Ubuntu

Supported Hardware

1. AMD GPU
2. NVIDIA GPU
3. Intel GPU
4. Apple Silicon Chip (M1, M2, M3, M4, ...)
5. Other DirectML devices (Intel Arc/Iris GPU, Qualcomm® Adreno GPU, etc. Windows Only)

Supported OS

1. Windows 10/11 (x64/Arm64)
2. MacOS 10.16 or later
3. Ubuntu 22.04 or later

Preparation and Installation

Windows

1. Install latest GPU driver

   AMD GPU:

   • Windows: Recommend to download the 25.9.2 driver for stable ROCm7 performance: AMD Software: Adrenalin Edition 25.9.2 Windows Download.
   • Ubuntu: Download latest GPU driver from AMD Drivers and Support for Processors and Graphics.

   NVIDIA GPU: Download latest GPU driver from NVIDIA Official GeForce Drivers.

   Intel GPU: Download latest GPU driver from Download Intel Drivers and Software.

   Qualcomm GPU: Download latest GPU driver from Qualcomm® Adreno™ Windows Graphics Drivers for Snapdragon® X Platform.

   Other DirectML devices: Please install the latest hardware driver accordingly.

2. Install Microsoft Visual C++ Redistributable

   Download Visual Studio 2017–2026 C++ Redistributable and install (restart Windows).

3. Enable Long Path

   Double click the long_path.reg in the Desktop2Stereo folder and confirm the warning.

4. Deploy Desktop2Stereo Environment

• Method 1 (Recommended): Use Portable Version

  Download: Quark NetDrive (Access code: 1vcn)

  AMD 7000/9000/Ryzen AI (Max)/etc. Series GPUs with ROCm7 Support: Portable Version is not available due to special deployment process, please refer to Method2.

  Older AMD/Intel/Qualcomm GPU and other DirectML devices: Download and unzip the Desktop2Stereo_vX.X.X_AMD_etc_Windows.zip to local disk.

  NVIDIA GPU: Download and unzip Desktop2Stereo_vX.X.X_NVIDIA_Windows.zip to local disk.

  Intel GPU: Download and unzip Desktop2Stereo_vX.X.X_Intel_Windows.zip to local disk.

• Method 2: Manual Deployment with embedded Python

  1. Download and unzip Desktop2Stereo_vX.X.X_python3_Windows.zip to local disk.

  2. Install Python environment

     AMD 6000/7000/9000/Ryzen AI (Max)/etc. Series GPUs with ROCm7 Support: Double click install-rocm7_standalone.bat. (Check compatibility here: https://rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html)

     Older AMD/Intel/Qualcomm GPU and other DirectML devices: Double click install-dml_standalone.bat.

     NVIDIA GPU: Double click install-cuda_standalone.bat.

     Intel GPU: Double click install-xpu_standalone.bat.

• Method 3: Manual Deployment with system Python

  1. Install Python 3.11

     Download from Python.org and install.

  2. Download Desktop2Stereo app

     Download the Desktop2Stereo.zip and unzip it to local disk.

  3. Install Python environment

     AMD 6000/7000/9000/Ryzen AI (Max)/etc. Series GPUs with ROCm7 Support: Double click install-rocm7.bat.

     Older AMD/Intel/Qualcomm GPU and other DirectML devices: Double click install-dml.bat.

     NVIDIA GPU: Double click install-cuda.bat.

     Intel GPU: Double click install-xpu.bat.

MacOS

1. Install Python 3.11

   Download from Python.org and install.

2. Download Desktop2Stereo app

   Download the Desktop2Stereo.zip and unzip it to local disk.

3. Install Python environment

   Double click install-mps executable. (Please allow open in Privacy and Security Settings). If you cannot run the executable, do the following first:

   chmod a+x install-mps
   chmod a+x run_mac
   chmod a+x update_mac_linux
   

Ubuntu

1. Install latest GPU driver

   AMD GPU: Download latest GPU driver and ROCm from AMD Drivers and Support for Processors and Graphics.

   NVIDIA GPU: Download latest GPU driver from NVIDIA Official GeForce Drivers.

2. Install Python 3.11-dev

   sudo add-apt-repository ppa:savoury1/python
   sudo apt update
   sudo apt-get install python3.11-dev python3.11-venv
   

3. Download Desktop2Stereo app

   Download the Desktop2Stereo_vX.X.X.zip and unzip it to local disk.

4. Install Python environment

   AMD 7000/9000/Ryzen AI (Max)/etc. Series GPU with ROCm7 Support: Check compatibility here: https://rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html

   bash install-rocm7.bash
   

   Older AMD GPU: Run install-rocm.bash:

   bash install-rocm.bash
   

   NVIDIA GPU: Run install-cuda.bash:

   bash install-cuda.bash
   

Run Desktop2Stereo

Quick Run

1. Choose one of the Run Mode in Desktop2Stereo: OpenXR Link (default), Local Viewer, MJPEG Streamer, RTMP Streamer, Legacy Streamer, 3D Monitor (Windows Only)
2. Select the Computing Device
3. Select target Monitor/Window
4. Just use the default settings and click Run.

OpenXR Link Mode

Tip: OpenXR Link mode (Windows/Ubuntu) is best for wireless streaming to OpenXR compatible client devices/apps with video and audio together, like Virtual Desktop VDXR Meta Horizon Link or Pico Connect with OpenXR support.

1. Connect your OpenXR Device to your PC via Wired/Wireless Connection. For Wireless Connection, recommend using Virtual Desktop VDXR Runtime for best experience on Wireless Connection.
2. Choose Run Mode as OpenXR Link.
3. Select the Computing Device and target Monitor/Window.
4. Click Run button to run, the 3D display will be launched on your OpenXR device.
5. (optional) For passthrough view, set Virtual Desktop VR passthrough settings as follows. Use button X on the Left VR Controller to toggle the passthrough view on/off.

OpenXR VR Controller Shortcuts

Crop Mode (OpenXR Link)

Crop Mode removes unwanted borders (e.g. movie letterbox black bars) from the captured frame so the video fills the virtual screen. Choose the mode from the Crop Mode dropdown in the GUI, or cycle it in VR.

• Auto — Automatically detects and removes letterbox/pillarbox black bars each time the content changes. Best for movies with baked-in black bars.
• Manual — You set a custom crop rectangle with the left controller. Use this when Auto misjudges the bars or you want a specific framing.
• Off — No cropping; the full captured frame is shown.

Manual crop gestures (left controller, laser pointed off the screen):

1. Hold the left trigger for 3s to cycle Auto → Manual → Off. An on-screen indicator shows the current mode.
2. In Manual mode, double-tap the left trigger to enter crop-adjust (the frame freezes and a pixel-size OSD appears). Double-tap again to apply and resume.
3. While adjusting, push the left stick: left/right (X) crops the sides, up/down (Y) crops the top/bottom. Only the dominant stick direction is applied per movement, so you never crop both axes at once. Cropping shrinks that axis of the screen without changing screen distance.

Manual crop rectangle and the selected crop mode are saved with your runtime settings.

Local Viewer Mode

Tip: Local Viewer mode is best for low-latency usage with SteamVR/Virtual Desktop/AR Glasses as wired display.

1. Choose Run Mode as Local Viewer.

2. Choose capture target by Monitor or Window mode. You can use Refresh button to update to the latest list of Monitor or Window.

3. Click the Stereo Viewer window. Use ← Left or → Right arrow keys to switch the Stereo Viewer window to second (virtual) monitor display.

4. Press Space or Enter or XBOX game controller button A to toggle fullscreen mode (On MacOS you may have to quickly press twice).

5. Now you can use AR/VR to view the SBS or TAB output.

   • AR needs to switch to 3D mode to connect as a 3840×1080 (Full Side-by-Side, Full-SBS) display.

     !Full-SBS

   • VR needs to use 2nd Display/Virtual Display (VDD) with Desktop+[Steam VR] or Virtual Desktop[PC/Standalone VR] or OBS + Wolvic Browser [Standalone VR] to compose the Half-SBS (Half Side-by-Side) / Full-SBS (Full Side-by-Side) / TAB (Top-and-Bottom) display to 3D.

     • You can use Tab key to toggle Half-SBS/Full-SBS/TAB/Depth Map mode.

     !Half-SBS !TAB

6. Real-time modification of depth strength.

   Use ↑ Up or ↓ Down arrow keys to increase/decrease the depth strength by a step of 0.5. To reset press 0 key.

   The definition of depth strength is in the detailed settings section.

7. Press Esc to exit the Stereo Viewer.

TIP: The Depth value will show below the FPS indicator if Show FPS is ON.

RTMP Streamer mode

Tip: RTMP Streamer mode is best for wireless streaming with video and audio together to client devices/apps by capturing the local Stereo Viewer window, like VLC Player, Wolvic Browser, etc., but it may have a latency of 1~3 seconds.

1. Choose run mode as RTMP Streamer.

2. Choose a Stream Protocol: recommended to use HLS.

3. Select an audio device

   • Windows

     Select the Stereo Mix as Stereo Mix (Realtek(R)), and select Realtek(R) HD Audio as the system Sound Output device.

     If your Windows device does not have the Stereo Mix (Realtek(R)), please install the Screen Capture Recorder and select the Stereo Mix as virtual-audio-capturer.

   • MacOS

     Install one of the following software containing the audio capture driver:

     a. BlackHole: https://existential.audio/blackhole/
     b. Virtual Desktop Streamer: https://www.vrdesktop.net/
     c. Loopback: https://rogueamoeba.com/loopback/ (Commercial)
     d. Or other virtual audio devices

     Select the Stereo Mix as BlackHole 2ch or Virtual Desktop Speakers or Loopback Audio or other virtual audio devices accordingly, and select the system Output device with same name.

     

   • Ubuntu

     Select the Stereo Mix device name ended with stereo.monitor i.e. alsa_output.pci-xxxx_xx_1x.x.analog-stereo.monitor, and select Output Device as Digital Output (S/PDIF)-xxxx in system sound settings.

     

4. Set a Stream Key, default is live.

5. (Optional) Adjust the Audio Delay. negative value means play the audio in advance before the video, positive value means delay the audio play after the video.

6. (Optional) It is recommended to use a second (virtual) screen with a resolution equal to or larger than the main screen to place the Stereo Viewer window.

7. The other settings are the same as the Local Viewer, click Run button to run.

8. On client device, key in the streaming URL according to the Stream Protocol.

Tip:

• AR: Use VLC Player to open the HLS M3U8 URL directly with Full-SBS mode.
• VR / Huawei AR: Use Wolvic Browser to open the HLS URL directly with Half_SBS / TAB mode.
• For MacOS, you can also use WebRTC URL.
• Other RTSP, RTMP, HLS M3U8 protocol may be chosen for VLC Player [i.e. extend screen mode for AR glasses] / VR Video Apps (DeoVR) on client devices.

If using Full-SBS output at the same resolution as the main screen, you will need a screen with twice the width of the original screen. For example, if the main screen is 4K (3840x2160), the second (virtual) screen needs to be 8K (7680x2160).

MJPEG Streamer mode

Tip: MJPEG Streamer mode is wireless streaming with video only to client devices/apps with lower latency, like Wolvic Browser, etc. For VR or Huawei AR: Wolvic Browser (Chromium Based) is recommended to open the HTTP MJPEG link.

1. Choose run mode as MJPEG Streamer.
2. Assign Streaming Port, default is 1122.
3. The other settings are the same as the Local Viewer, click Run button.
4. On client device, key in the Streamer URL to access the video.
5. For audio, please use Bluetooth or Headphones connected to your PC or Mac.

Legacy Streamer mode

Tip: Legacy Streamer mode is a legacy MJPEG streaming mode, which uses PyTorch method to generate left and right eye scenes. The main usage is the same as the MJPEG Streamer mode.

3D Monitor mode (Windows Only)

Tip: 3D Monitor mode is a special Local Viewer mode dedicated for a 3D Monitor, no virtual display driver needed for this mode. It can only run as fullscreen and be used locally as the screen capture attribute for the Stereo Viewer window is disabled globally. In 3D Monitor mode, please use the passthrough cursor on either left or right scene to control your PC.

Full Keyboard Shortcuts

Tip: Need to click the Stereo Viewer window/tab first to use.

Detailed Settings Guide

All optional settings can be modified on the GUI window and saved to the settings.yaml. Each time you click Run, the settings will be saved automatically, and clicking Reset will restore the default settings.

1. Run Mode
   6 Run Modes on Windows, 5 on other platforms. The default is OpenXR Link. Available modes: OpenXR Link, Local Viewer, MJPEG Streamer, RTMP Streamer, Legacy Streamer, 3D Monitor (Windows Only).

2. Set Language
   English (EN) and Simplified Chinese (CN) are supported.

3. Monitor or Window mode

   

   Default is your Primary Monitor (mostly shall follow the monitor numbers in your system settings). You can toggle to Window capture mode as well, the optional menus will include all the active window names.

4. Computing Device
   Default shall be your GPU (CUDA/DirectML/MPS), or CPU if you don't have a compatible computing device.

5. FP16
   Recommended for most computing devices for better performance. If your device does not support FP16 DataType, disable it.

6. Show FPS
   Show FPS on the titlebar of the Stereo Viewer and as an on-screen indicator on the output left and right eye scenes. The capture FPS is auto-detected from your monitor's refresh rate.

7. Capture Tool (Windows Only)

   • DXCamera: Based on wincam using DXGI Desktop Duplication API, it has the highest FPS but higher CPU temperature.
   • WindowsCapture: Based on Windows-Capture Python using Graphics Capture API, it has slightly lower FPS but lower CPU usage and temperature.
   • WindowsCaptureCUDA: NVIDIA CUDA-accelerated variant of WindowsCapture for reduced overhead.
   • WindowsCaptureROCm: AMD ROCm-accelerated variant of WindowsCapture.
   • DesktopDuplication: DXGI Desktop Duplication API via windows-capture. Window capture mode disabled.

8. Fill 16:9
   Enabled by default. If the aspect of input source is not 16:9, the black background will be applied to fill it to 16:9.

9. Fix Viewer Aspect (Local Viewer mode Only)
   Disabled by default. This option is to lock the window of Stereo Viewer, which may be useful for the upscaling and frame generation apps like Lossless Scaling.

10. Depth Resolution
    Higher Depth Resolution can give better depth details but cause higher GPU usage, which is also related to the model training settings. Default Depth Resolution is set to 336 for balanced performance on Depth-Anything-V2 models. The Depth Resolution options vary among different depth models.

11. Depth Strength
    With higher Depth Strength, 3D depth effect of the object would be stronger. However, higher value can induce visible artifacts and distortion. Default is set to 2.0. The recommended depth strength range is (1, 5).

12. Anti-Aliasing
    This can be effective to reduce jagged edges and artifacts under high Depth Strength, default value is set as 2 for most cases. Higher value may reduce the depth details.

13. Foreground Scale
    Default value is 0.5. Positive value means foreground closer, background further. Negative value means foreground flatter, background closer. 0 is no change of foreground and background strength.

14. Display Mode
    It determines how the left and right eye scenes are arranged in the output. Default is Half-SBS for most VR devices, TAB is also an alternative; Full-SBS is mainly for AR glasses.

    • Full-SBS (Full Side-by-Side, 32:9)
      Two full-resolution images are placed side by side: one for the left eye, one for the right. Requires a display capable of handling double-width input. Offers higher image quality but demands more bandwidth and processing.

    • Half-SBS (Side-by-Side Half, 16:9)
      Two images are placed side by side, but each is compressed horizontally to fit into a single frame. More compatible with standard displays and media players. Slightly lower image quality due to reduced resolution per eye.

    • TAB (Top-and-Bottom, 16:9)
      Left and right eye images are stacked vertically: one on top, one on bottom. Each image is compressed vertically to fit the frame. Common in streaming and broadcast formats; quality similar to Half-SBS.

    • Depth Map (16:9)
      Visualizes the depth estimation output as a color-mapped image (blue = far, red = near). Press D key to toggle between depth map and original RGB view. Useful for debugging depth model accuracy.

15. IPD (Interpupillary Distance)
    IPD is the distance between the centers of your pupils, it affects how your brain interprets stereoscopic 3D. The default IPD is 0.064 meter (m), which is the average human IPD value.

16. Convergence Point
    The convergence point refers to the point where the lines of sight from both eyes intersect, affecting the depth perception in stereoscopic vision. A smaller convergence point enhances the "out-of-screen" effect but may cause blurriness or ghosting at the screen edges.
    The default convergence point is set to 0.5 for a balanced 3D effect and screen-edge clarity.

17. Controller Model (OpenXR Link Only)
    Select the VR controller 3D model to display in-headset. Default is PICO. Supported models are auto-detected from the controllers/ folder.

18. Stereo Output (Local Viewer / RTMP Streamer / 3D Monitor)
    When multiple monitors are connected, select which monitor the Stereo Viewer window should appear on. Default is the first external monitor (not the capture input). Set to Viewer Window to use the default window placement.

19. Stream Protocol (RTMP Streamer Only)
    Default is HLS for best compatibility, and HLS M3U8 can be used in mobile VLC Player. RTMP, RTSP, HLS, HLS M3U8, WebRTC are provided. You can toggle the protocol to show the target URL, all URLs are ready to use when the RTMP Streamer is working.

20. Streamer URL (RTMP Streamer, MJPEG Streamer, Legacy Streamer Only)
    Read only, dynamically determined by the streaming protocol and your local IP.

21. Stream Key (RTMP Streamer Only)
    The private key string set for RTMP Streamer, which will be applied in the Streamer URL.

22. Stream Quality (MJPEG Streamer / Legacy Streamer Only)
    JPEG encoding quality for MJPEG streaming. Range 50–100, default is 100.

23. CRF (RTMP Streamer Only)
    Default is 20, you can set it in the range of 16~27. It refers to Constant Rate Factor that controls the video bitrate. A lower value is a higher quality.

24. Stereo Mix (RTMP Streamer Only)
    This is the Stereo Mix device to capture the system playback audio.

    • On Windows, Stereo Mix device is mostly Stereo Mix (Realtek(R)) to be used with Realtek(R) HD Audio as the output device in Windows audio settings. Or use the virtual audio device from Screen Capture Recorder.
    • On MacOS, you can choose Stereo Mix device BlackHole or Virtual Desktop Speakers or Loopback or other virtual audio devices. Please use the same audio output device in MacOS audio settings.

25. Audio Delay (RTMP Streamer Only)
    Default is -0.15 seconds, which is used to align the processed audio and video timestamp. A negative value will make the audio earlier than the video, whereas a positive value will make the audio later than the video.

26. Lossless Scaling Support (RTMP Streamer Only, Windows Only)
    When enabled, uses precise window-crop capture via gfxcapture to frame only the Stereo Viewer window instead of the full monitor. Useful when using Lossless Scaling for frame generation.

27. Download Path
    Default download path is the models folder under the working directory.

28. Depth Model
    Modify the depth model id from HuggingFace, the model id under depth_model mostly shall end with -hf. Large model can cause higher GPU usage and latency. Default depth model: depth-anything/Depth-Anything-V2-Small-hf. You can also manually add the hugging face models in the settings.yaml which include model.safetensors, config.json, preprocessor_config.json files on HuggingFace.

    Currently supported models include (partial list):

    • depth-anything/Depth-Anything-V2-Small-hf
    • depth-anything/Depth-Anything-V2-Base-hf
    • depth-anything/Depth-Anything-V2-Large-hf
    • depth-anything/Video-Depth-Anything-Small
    • depth-anything/Video-Depth-Anything-Base
    • depth-anything/Video-Depth-Anything-Large
    • depth-anything/DA3-SMALL
    • depth-anything/DA3-BASE
    • depth-anything/DA3-LARGE-1.1
    • depth-anything/DA3-GIANT-1.1
    • depth-anything/DA3METRIC-LARGE
    • depth-anything/DA3NESTED-GIANT-LARGE-1.1
    • depth-anything/Depth-Anything-V2-Metric-Outdoor-Small-hf
    • depth-anything/Depth-Anything-V2-Metric-Outdoor-Base-hf
    • depth-anything/Depth-Anything-V2-Metric-Outdoor-Large-hf
    • depth-anything/Depth-Anything-V2-Metric-Indoor-Small-hf
    • depth-anything/Depth-Anything-V2-Metric-Indoor-Base-hf
    • depth-anything/Depth-Anything-V2-Metric-Indoor-Large-hf
    • depth-anything/Metric-Video-Depth-Anything-Small
    • depth-anything/Metric-Video-Depth-Anything-Base
    • depth-anything/Metric-Video-Depth-Anything-Large
    • LiheYoung/depth-anything-small-hf
    • LiheYoung/depth-anything-base-hf
    • LiheYoung/depth-anything-large-hf
    • xingyang1/Distill-Any-Depth-Small-hf
    • lc700x/Distill-Any-Depth-Base-hf
    • xingyang1/Distill-Any-Depth-Large-hf
    • facebook/dpt-dinov2-small-kitti
    • lc700x/dpt-dinov2-base-kitti-hf
    • lc700x/dpt-dinov2-large-kitti-hf
    • lc700x/dpt-dinov2-giant-kitti-hf
    • lc700x/dpt-dinov2-small-nyu-hf
    • lc700x/dpt-dinov2-base-nyu-hf
    • lc700x/dpt-dinov2-large-nyu-hf
    • facebook/dpt-dinov2-giant-nyu
    • lc700x/depth-ai-hf
    • lc700x/dpt-hybrid-midas-hf
    • Intel/dpt-beit-base-384
    • Intel/dpt-beit-large-512
    • Intel/dpt-large
    • lc700x/dpt-large-redesign-hf
    • Intel/zoedepth-nyu-kitti
    • Intel/zoedepth-nyu
    • Intel/zoedepth-kitti
    • apple/DepthPro-hf # Slow, NOT recommended

29. HF Endpoint (Hugging Face)
    HF-Mirror is a mirror site of the original Hugging Face site hosting AI models. The depth model will automatically be downloaded to Download Path from Hugging Face at the first run.

30. Inference Optimizer (Windows/Ubuntu Only)
    These optimizers can typically increase the output FPS by 30%~50%. However, not all models support Inference Optimizer, if the optimization fails, the inference process will fall back to PyTorch.
    Use the Recompile checkbox to force regeneration of cached optimized models.

    AMD GPUs (ROCm7):

    • torch.compile: leverages Triton under the hood to generate optimized kernels automatically, and provides slight to moderate speedups by fusing operations and reducing overhead.

    NVIDIA GPUs:

    • torch.compile: leverages Triton under the hood to generate optimized kernels automatically, and provides slight to moderate speedups by fusing operations and reducing overhead.
    • TensorRT: NVIDIA’s high-performance deep learning inference SDK. It optimizes trained models for deployment, especially on NVIDIA GPUs, and provides significant speedups and high inference efficiency. Check Recompile TensorRT to force engine rebuild.

    Apple Silicon (MPS):

    • CoreML: CoreML is optimized to leverage Apple silicon’s CPU, GPU, and Neural Engine for fast, private, and offline predictions. Check Recompile CoreML to force model reconversion.

    Intel GPUs:

    • OpenVINO: This is Intel’s deep learning inference optimization toolkit designed to improve inference performance on Intel hardware. By optimizing and accelerating models, it can significantly increase inference speed. Check Recompile OpenVINO to force cache rebuild.

    AMD GPUs, etc. DirectML devices:

    • Unlock Threads (Legacy Streamer): unlock the multithreads for Legacy Streamer mode.

Warning: Unlock Threads (Legacy Streamer) sometimes fails with UTF-8 error under Python3.11 due to the limitations of torch-directml libraries. You may try stop and run multiple times for a successful streaming process. Warning: torch.compile currently is not compatible with AMD RX6000 Series GPU.

References

@article{depthanything3,
  title={Depth Anything 3: Recovering the visual space from any views},
  author={Haotong Lin and Sili Chen and Jun Hao Liew and Donny Y. Chen and Zhenyu Li and Guang Shi and Jiashi Feng and Bingyi Kang},
  journal={arXiv preprint arXiv:2511.10647},
  year={2025}
}

@article{video_depth_anything,
  title={Video Depth Anything: Consistent Depth Estimation for Super-Long Videos},
  author={Chen, Sili and Guo, Hengkai and Zhu, Shengnan and Zhang, Feihu and Huang, Zilong and Feng, Jiashi and Kang, Bingyi},
  journal={arXiv:2501.12375},
  year={2025}
}

@article{depth_anything_v2,
  title={Depth Anything V2},
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Zhao, Zhen and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  journal={arXiv:2406.09414},
  year={2024}
}

@inproceedings{depth_anything_v1,
  title={Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data}, 
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  booktitle={CVPR},
  year={2024}
}

@article{li2024amodaldepthanything,
  title={Amodal Depth Anything: Amodal Depth Estimation in the Wild}, 
  author={Li, Zhenyu and Lavreniuk, Mykola and Shi, Jian and Bhat, Shariq Farooq and Wonka, Peter},
  year={2024},
  journal={arXiv preprint arXiv:x},
  primaryClass={cs.CV}}

@article{he2025distill,
  title   = {Distill Any Depth: Distillation Creates a Stronger Monocular Depth Estimator},
  author  = {Xiankang He and Dongyan Guo and Hongji Li and Ruibo Li and Ying Cui and Chi Zhang},
  year    = {2025},
  journal = {arXiv preprint arXiv: 2502.19204}
}

@article {Ranftl2022,
    author  = "Ren\'{e} Ranftl and Katrin Lasinger and David Hafner and Konrad Schindler and Vladlen Koltun",
    title   = "Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer",
    journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
    year    = "2022",
    volume  = "44",
    number  = "3"
}

@article{birkl2023midas,
      title={MiDaS v3.1 -- A Model Zoo for Robust Monocular Relative Depth Estimation},
      author={Reiner Birkl and Diana Wofk and Matthias M{\"u}ller},
      journal={arXiv preprint arXiv:2307.14460},
      year={2023}
}

@article{bhat2023zoedepth,
  title={Zoedepth: Zero-shot transfer by combining relative and metric depth},
  author={Bhat, Shariq Farooq and Birkl, Reiner and Wofk, Diana and Wonka, Peter and M{\"u}ller, Matthias},
  journal={arXiv preprint arXiv:2302.12288},
  year={2023}
}

@inproceedings{Bochkovskii2024:arxiv,
  author     = {Aleksei Bochkovskii and Ama\"{e}l Delaunoy and Hugo Germain and Marcel Santos and Yichao Zhou and Stephan R. Richter and Vladlen Koltun},
  title      = {Depth Pro: Sharp Monocular Metric Depth in Less Than a Second},
  booktitle  = {International Conference on Learning Representations},
  year       = {2025},
  url        = {https://arxiv.org/abs/2410.02073},
}

@article{Ranftl2020,
	author    = {Ren\'{e} Ranftl and Katrin Lasinger and David Hafner and Konrad Schindler and Vladlen Koltun},
	title     = {Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer},
	journal   = {IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
	year      = {2020},
}

@misc{oquab2023dinov2,
      title={DINOv2: Learning Robust Visual Features without Supervision}, 
      author={Maxime Oquab and Timothée Darcet and Théo Moutakanni and Huy Vo and Marc Szafraniec and Vasil Khalidov and Pierre Fernandez and Daniel Haziza and Francisco Massa and Alaaeldin El-Nouby and Mahmoud Assran and Nicolas Ballas and Wojciech Galuba and Russell Howes and Po-Yao Huang and Shang-Wen Li and Ishan Misra and Michael Rabbat and Vasu Sharma and Gabriel Synnaeve and Hu Xu and Hervé Jegou and Julien Mairal and Patrick Labatut and Armand Joulin and Piotr Bojanowski},
      year={2023},
      eprint={2304.07193},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Credits

• lovettchris/wincam
• NiiightmareXD/windows-capture
• BoboTiG/python-mss
• nagadomi/nunif
• VirtualDrivers/Virtual-Display-Driver
• waydabber/BetterDisplay
• guinmoon/rocm7_builds
• woct0rdho/triton-windows
• Other related tools and libraries
• All feedback from the users