NVlabs/RADIO

Official repository for "AM-RADIO: Reduce All Domains Into One"

What it solves

RADIO is designed to replace multiple large vision foundation models with a single, unified backbone. Instead of running separate models for text grounding, segmentation, and general image recognition, RADIO distills the knowledge of several "teacher" models (such as CLIP, DINOv2, and SAM) into one model that preserves the unique strengths of each.

How it works

It uses an agglomerative distillation framework to combine features from various vision models. The system supports arbitrary input resolutions and non-square images. It also employs "adaptors" that allow the model to output features aligned with specific teachers (e.g., SigLIP or SAM), and some of these adaptors support text tokenization and encoding for vision-language tasks.

Who it’s for

This project is for computer vision researchers and developers who need a high-performance vision backbone that is versatile across different domains and more efficient than running multiple separate foundation models.

Highlights

  • Multi-Teacher Distillation: Integrates capabilities from CLIP, DINOv2, and SAM into a single model.
  • Flexible Input: Supports non-square images and arbitrary resolutions.
  • Efficiency: Includes an E-RADIO variant that is 6-10x faster than CLIP and DINOv2.
  • Commercial Availability: Offers C-RADIO model families under a commercially permissive license.
  • Variable Tokenization: The RADIO1D research variant allows tunable trade-offs between feature compactness and reconstruction fidelity via a 1D sequence of global tokens.

Related

  • Project
    chn-lee-yumi/MaterialSearch

    A natural language search tool for local photos and videos that enables text-based and image-based retrieval of visual assets.

  • Project
    yehengchen/Object-Detection-and-Tracking

    This project is a practical collection of object detection and tracking pipelines, featuring one-stage methods like YOLO and SSD, two-stage methods like Mask R-CNN, and ROS integrations for robotics. It helps developers deploy real-time detection and tracking on custom data, including pedestrian counting, social distance measurement, and robot perception.

  • Project
    zs1083339604/FaceWinUnlock-Tauri

    FaceWinUnlock‑Tauri is a Windows‑only tool that adds webcam‑based face unlock by injecting a custom Credential Provider DLL. It combines a Tauri‑Rust back‑end (registry edits, SQLite storage) with a Vue 3 front‑end and OpenCV face matching, offering multi‑account support, optional liveness detection, and logging. The core code is now closed‑source; use only on non‑critical PCs and test in a VM because registry changes can lock you out.

  • Project
    manisandro/gImageReader

    gImageReader is a graphical interface for the Tesseract OCR engine that allows users to easily extract text from images, PDFs, and scanned documents.