yuanze-lin/Olympus

[CVPR 2025 Highlight] Official code for "Olympus: A Universal Task Router for Computer Vision Tasks"

What it solves

Olympus is a universal task router for computer vision. It allows users to provide a single natural-language instruction that can span multiple vision tasks—such as generating an image, editing it, and then creating a 3D model from that edited image—without having to manually chain different AI models together.

How it works

Olympus acts as a central dispatcher. It parses a user's prompt into "routing tokens," which identify the specific vision tasks required. It then dispatches these tasks to a set of specialist models (e.g., Qwen-Image for generation, TRELLIS.2-4B for 3D generation) and automatically resolves dependencies between them, ensuring the output of one step becomes the input for the next.

Who it’s for

It is designed for researchers and developers who need to execute complex, multi-step computer vision pipelines (covering 20 different tasks) through a single interface using natural language.

Highlights

  • Multi-Task Routing: Supports 20 different computer vision tasks from a single prompt.
  • Asset Generation: Produces finished files including .png images, .mp4 videos, and .glb 3D models.
  • Automated Chaining: Automatically handles the dataflow between specialist models so users don't have to wire steps manually.
  • Memory Efficient: Specialists are loaded and freed one at a time to keep peak GPU memory usage low.

Related

  • Project
    ultralytics/ultralytics

    A high-performance computer vision framework implementing YOLO models for real-time object detection, segmentation, classification, and pose estimation.

  • Project
    bytedance/Lance

    Lance is a 3B parameter unified multimodal model that integrates image and video understanding, generation, and editing into a single framework.

  • Project
    LibreYOLO/libreyolo

    An MIT-licensed computer vision library that provides a single API for detection, segmentation, pose, depth, OCR, and other vision tasks, including training and export tools.

  • Project
    NVIDIA/cosmos

    An open platform of omnimodal world models and tools for building Physical AI, enabling the joint processing and generation of text, vision, audio, and action sequences for robotics and autonomous systems.

  • Project
    img2threejs/img2threejs

    img2threejs is a Python‑backed skill that lets an LLM turn a single reference image into a procedural Three.js model. It builds a spec, runs a gated vision‑review loop, and emits TypeScript code that recreates the object from primitives and shaders, ready for animation. The system is extensible via domain plugins (e.g., CS2 weapon skins) and includes a live demo gallery.