Comp0.2

Comp 0.2 / Optional AI tools

Local AI and Worker setup.

Comp includes the local AI runtime. Download compatible model bundles for the tools you want to use, select their folders in Settings and enable the models.

Model downloads

Download the AI model bundle for Comp 0.2 and AI Worker 0.1.2 as a ZIP archive.

Download AI models

Choose the tool

ToolPurposeProcessing route
AutoMatte with SAM3Select a subject with prompts.Local runtime or external Worker.
AutoMatte with MatAnyone2Refine mattes across frames.Local runtime or external Worker.
CCkeyGreen-screen keying based on Corridor Crew Keyer 1.Integrated local runtime.
CCdespillSpill treatment and luminance restoration.Ordinary image processing; no model required.

Enable local processing

  1. Extract comp_ai_bundles.zip, keeping each model folder's manifest and required assets together.
  2. In Local AI settings, select the extracted mat2 and sam3 folders for the models you want to use.
  3. Enable the models and save. Enabled models load together; resolve any initialization error before using them.
  4. For CCkey, choose a 1024 or 2048 folder inside cckey_resolutions_mps on macOS or cckey_resolutions_cuda on Windows or Linux. Match that folder to the inference resolution. Output keeps the source dimensions.

Local processing uses Apple Metal (MPS) on macOS and NVIDIA CUDA on Windows or Linux. Precision options depend on the model and device; start with the default.

Some settings take effect after restarting Comp; the dialog marks them.

Local or Worker processing

AutoMatte uses the Worker when its local model is disabled. If an enabled local model fails, Comp reports the error instead of switching to the Worker. CCkey runs locally.

Store analysis results

Set AutoMatte's result-cache folder in Settings. Large sequences can use disk storage. A new folder applies only to new results; existing files stay in their original location.

Use an external Worker when needed

The Worker runs AutoMatte in a separate process or on another machine. Worker 0.1.2 uses the mat2 and sam3 folders from the same model archive.

Replace <models-folder> with the folder where you extracted the archive. Pass these paths to the Worker's --family-bundle options:

  • mat2=<models-folder>/comp_ai_bundles/mat2
  • sam3=<models-folder>/comp_ai_bundles/sam3

Set --target to mps for an Apple GPU, cuda for an NVIDIA GPU or cpu for CPU processing. Enter the Worker's host and port in Comp; the default local address is localhost:19282.

Include frame-transfer time when comparing local and remote processing.

Model files and terms

AI models are downloaded separately. Follow the access and licence terms for each bundle and its upstream model.

Download Comp

Get CompBrowse nodes