scope-yolo-mask

A Scope plugin that segments objects in real-time video using YOLO26, producing binary masks for VACE inpainting and conditioning.

Features

• Object Segmentation — Real-time detection and segmentation of 19 COCO object classes (person, car, dog, etc.) using YOLO26-seg (video mode)
• Mask Output — Binary segmentation mask displayed as a grayscale image, suitable for debugging and mask preview
• Overlay Output — Mask blended on the original frame with a green tint for intuitive visualization
• Invert Mask — Option to invert the mask to segment the background instead of the detected object
• VACE Integration — Outputs masks in VACE format ([1, 1, F, H, W]) for downstream inpainting pipelines
• Multiple Model Sizes — Choose from nano, small, medium, large, or xlarge for speed vs. accuracy tradeoffs

Install

Follow the Scope plugins guide to install this plugin using the URL:

https://github.com/daydreamlive/scope_yolo_mask.git

Upgrade

Follow the Scope plugins guide to upgrade this plugin to the latest version.

Architecture

This plugin registers one pipeline via the register_pipelines hook in plugin.py:

YOLO Mask (yolo_mask)

A video-mode preprocessor pipeline that segments objects in each input frame using YOLO26-seg. It accepts a model_size load parameter (nano/small/medium/large/xlarge, default nano) that determines the YOLO model variant. At runtime, target_class selects which COCO class to detect (default person), confidence_threshold (0.0–1.0, default 0.5) filters low-confidence detections, output_mode (mask or overlay) controls the display output, and invert_mask flips the segmented region.

The pipeline outputs three tensors:

• video — display frames (T, H, W, 3) in [0, 1] range (mask visualization or overlay)
• vace_input_frames — masked video in VACE format [1, C, F, H, W] in [-1, 1] range
• vace_input_masks — binary masks [1, 1, F, H, W] where 1 indicates regions to inpaint

Models are downloaded automatically by Ultralytics to ~/.daydream-scope/models/ultralytics/.