Comp releases

Welcome to the release page for Comp. Thank you for your interest.

Warning

Comp is currently available as an early public alpha for evaluation and testing. Please keep backups of important files and use the application at your own discretion.

If you would like to try the project and help shape its future direction, I would be very grateful for your feedback.

I have been working on Comp for more than a year and a half. This release started as a simple personal side project, but after seeing some interest from the VFX community, I decided to package it into a more coherent form so that anyone interested can test it and share feedback.

Your feedback will help define the direction of future development, as well as the pace at which the project can move forward.

Community and Support

• LinkedIn: https://www.linkedin.com/in/yuri-radaev
• Buy Me a Coffee: https://buymeacoffee.com/yuri.r
• Discord: https://discord.gg/caJSDWwXX

Downloads

The current public alpha is Comp v0.1.1, released on July 14, 2026.

Comp application

Platform	Recommended download	Alternative package	SHA-256 checksums
Windows x64	Comp v0.1.1 — ZIP, 50.9 MB	—	Windows checksums
macOS arm64	Comp v0.1.1 — DMG, 77.4 MB	ZIP, 67.8 MB	macOS checksums

• Windows: extract the ZIP before running Comp. The current build is unsigned, so Windows may display a security warning. Verify the downloaded file using the published checksum before allowing it to run.
• macOS: the DMG is recommended for normal installation. The ZIP is provided as an alternative package.
• AI features: Comp itself does not require the AI Worker. Download the Worker only if you specifically want to test the AI modules described below.

View the Comp v0.1.1 release page
View all releases

Optional: Comp AI Worker Service

Comp can be used as a standalone compositor. You do not need the AI Worker for standard compositing operations.

The Comp AI Worker Service is a separate command-line component used by the optional AI modules currently available in Comp:

• SAM3 — rough image segmentation;
• MatAnyone2 — mask-edge refinement.

For the simplest setup, run Comp and the Worker on the same machine. The Worker can also accept requests over a local network, but transferring image frames between machines may add significant overhead.

Worker downloads

The current Worker release is v0.1.0.

Platform	Recommended download	Alternative package	SHA-256 checksums
Windows, CPU/CUDA	Worker v0.1.0 — ZIP, 1.71 GB	—	Windows checksums
macOS, CPU/MPS	Worker v0.1.0 — DMG, 47.8 MB	ZIP, 39.9 MB	macOS checksums

View the Worker v0.1.0 release page

Important

Neural-network model weights are not included in the Worker packages. Download the required files separately and place them in the folders shown below.

Required model files

Module	Download	Destination	Final filename
MatAnyone2	Download mat2_bundle.pt	<worker-installation-root>/family_bundles/mat2/models/	mat2_bundle.pt
SAM3 tokenizer	Download bpe_simple_vocab_16e6.txt.gz	<worker-installation-root>/family_bundles/sam3/assets/	bpe_simple_vocab_16e6.txt.gz
SAM3.1 model	Download sam3_bundle.pt	<worker-installation-root>/family_bundles/sam3/models/	sam3_bundle.pt

The resulting directory structure should look like this:

<worker-installation-root>/
└── family_bundles/
    ├── mat2/
    │   └── models/
    │       └── mat2_bundle.pt
    └── sam3/
        ├── assets/
        │   └── bpe_simple_vocab_16e6.txt.gz
        └── models/
            └── sam3_bundle.pt

If Hugging Face asks you to sign in, request access, or accept model repository conditions, complete that step before downloading the file.

Third-party model files have their own licenses and access conditions. Review those conditions before downloading or using the models.

Starting the Worker

Replace the placeholder paths and TARGET in the following example:

"<path-to-worker-executable>" serve \
  --family-bundle "mat2=<worker-installation-root>/family_bundles/mat2" \
  --family-bundle "sam3=<worker-installation-root>/family_bundles/sam3" \
  --target TARGET \
  --bind-host 127.0.0.1 \
  --port 19282 \
  --temp-dir "<path-to-temp-directory>/comp-worker" \
  --cpu-threads 0 \
  --precision fp32 \
  --verbose

Use one of the following target values:

• cuda — NVIDIA GPU;
• mps — Apple GPU on macOS;
• cpu — CPU processing.

The example uses 127.0.0.1, which makes the Worker accessible only from the same machine.

For a reduced-memory test, add:

--max-propagation-history-depth 2
--max-mem-frames 2

The command above uses Unix-style line continuation. In PowerShell, either put the command on one line or replace each trailing \ with a PowerShell backtick.

Supported platforms

At the moment, Comp is available for Windows x64 and macOS arm64

A Linux build is also planned. Before starting the Linux port, I would like to reach a stable MVP on the current platforms and use that version as a baseline for parity testing.

The Windows build may work through Wine, but I have not tested this.

Development is currently done mostly on macOS. From time to time, I also use my old Windows PC to build and test the Windows version.

GPU rendering

At the moment, Comp does not use GPU acceleration for compositing operations. Only image drawing in the viewer is accelerated.

I have not yet chosen the backend for GPU compositing. The cross-platform Qt RHI looks like the simplest solution, but it would introduce Qt dependencies into the headless renderer.

One possible option would be to simply not support GPU acceleration in headless mode, but this is still open for discussion.

A more hardcore option would be to write three separate implementations for Metal, Vulkan and CUDA, this would provide speed, flexibility, and fewer limitations. However, before doing that, the current feature set needs to be stabilized so that rendering is predictable and can be used as a reference for comparison with the GPU implementation.

Approximate roadmap for Comp

• Release the first alpha version.
• Collect feedback and fix bugs.
• Release the second alpha version.
• If no critical bugs are found, freeze this version as the reference for the Linux port.
• Hold a vote and choose a Linux distribution for development and testing.
• Release the second alpha version for Linux.
• Fix bugs and freeze that version as the reference.
• Develop and integrate GPU rendering identical to CPU rendering.
• Release the third alpha version.

Approximate roadmap for AI Worker Service

• Finish implementing two new modules:
  • Corridor Crew keyer
  • DVD depth estimator

Related projects, papers, and dependencies

• MatAnyone2 (project page)
• Facebook/SAM3 (Meta research page, SAM3.1 model weights)
• OpenCV (GitHub)
• V-BM3D / VBM3D paper
• VBM3D implementation paper
• JSON (ECMA-404)
• libtorch / PyTorch C++ API
• C++17 / ISO/IEC 14882:2017 (cppreference C++17 overview)
• Qt RHI
• Qt Advanced Docking System — ADS
• nlohmann/json
• ExprTk
• Tcl/Tk
• OpenEXR / Imath
• OpenMP