Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 460 stars

LHRLAB/HyperGraphRAG

[NeurIPS 2025] Official resources of "HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation".

Overview: HyperGraphRAG is an official resource for the NeurIPS 2025 paper "HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation". It provides comprehensive guidance and evaluation metrics for developers interested in enhancing language models with knowledge representation. Key features: - Knowledge HyperGraph Construction - Knowledge HyperGraph Query - Evaluation Metrics Tech stack: Python Why it matters: HyperGraphRAG offers a practical approach to integrating knowledge representation into language models, enhancing their ability to generate accurate and contextually relevant responses. Repository: LHRLAB/HyperGraphRAG GitHub stars: 459 GitHub description: Official resources of "HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representation".

C++ 460 stars

NVlabs/intrinsic3d

Intrinsic3D - High-Quality 3D Reconstruction by Joint Appearance and Geometry Optimization with Spatially-Varying Lighting (ICCV 2017)

### Overview: Intrinsic3D is a method for high-quality 3D reconstruction using low-cost RGB-D sensors. The algorithm optimizes geometry, surface albedos, camera poses, and scene lighting to achieve fine-scale geometric details and sharp surface textures. ### Key Features: - **Joint Appearance and Geometry Optimization**: Recovers fine-scale geometric details and sharp surface textures by simultaneously optimizing for geometry, surface albedo, camera poses, and scene lighting. - **Spatially-Varying Lighting**: Accounts for spatially-varying lighting conditions, improving reconstruction accuracy. - **High-Quality Reconstruction**: Provides high-quality 3D reconstructions from low-cost sensors, making it useful for applications like augmented reality and robotics. ### Tech Stack: - **Programming Languages**: C++ - **Dependencies**: CMake, Eigen, OpenCV 4, Boost, Ceres Solver, CXSparse (on Windows) ### Why It Matters: Intrinsic3D offers a significant improvement in 3D reconstruction quality compared to traditional methods, making it particularly useful for applications requiring high accuracy and detail, such as augmented reality and robotics. The method's ability to handle spatially-varying lighting conditions makes it robust against environmental changes, enhancing its practical applicability.

Unknown 460 stars

Zheng-Chong/Awesome-Try-On-Models

A repository for organizing papers, codes and other resources related to Virtual Try-on Models

### Overview: Virtual Try-on (VTON) refers to the technology that allows users to try on clothing or accessories in a virtual environment, typically using generative models in AI. This repository focuses on image, video, and 3D-based try-on models published after 2023, providing insights into the latest technological trends. ### Key Features: - **Image-based Models:** - [2026-07-24] [Layering Virtual Try-On](https://arxiv.org/abs/2607.22924) **(ECCV 2026)** [![Star](https://img.shields.io/github/stars/ChuenFung/Layering-Virtual-Try-On.svg?style=social&label=Star)](https://github.com/ChuenFung/Layering-Virtual-Try-On) - [2026-07-23] [Oxygen-TryOn: Fashion-Native Foundation Model for Any-item Virtual Try-On](https://arxiv.org/abs/2607.21694) **(arXiv)** [![Star](https://img.shields.io/github/stars/OxygenVision/Oxygen-Try-On.svg?style=social&label=Star)](https://github.com/OxygenVision/Oxygen-Try-On) - [2026-06-26] [OrthoTryOn: Geometric Orthogonalization for Conflict-Free Unified Fashion Generation](

Unknown 460 stars

awesome-rag/awesome-rag

Awesome-RAG: Collect typical RAG papers and systems.

### Overview: Awesome-RAG is a comprehensive collection of typical RAG (Reinforcement Learning with Graphs) papers and systems. It serves as a valuable resource for researchers and developers looking to stay updated with the latest advancements in the field. ### Key Features: - **Survey-2022** - **Survey-2023** - **Survey-2024** - **RAG-2024** - **Benchmark-2024** ### Tech Stack: Python, Markdown, and Git. ### Why It Matters: This repository provides a curated overview of RAG research, helping users stay informed about the latest developments and systems in the field.

Python 460 stars

encord-team/encord-active

The toolkit to test, validate, and evaluate your models and surface, curate, and prioritize the most valuable data for labeling.

### Overview: Encord Active is an open-source toolkit designed to test, validate, and evaluate models, surface and prioritize valuable data, and enhance model performance. It provides advanced error analysis, model explainability reports, natural language search capabilities, and features to detect and fix dataset errors and biases. ### Key Features: - **Model Testing and Validation**: Advanced error analysis and model evaluation. - **Data Surface and Prioritize**: Surface and prioritize valuable data for labeling. - **Natural Language Search**: Beta feature for searching through data using natural language. - **Error Detection and Fixing**: Detect and fix duplicates, outliers, and label errors. ### Tech Stack: Python ### Why It Matters: Encord Active is essential for developers looking to improve model performance by testing, validating, and evaluating models. It helps in identifying and fixing errors in datasets, leading to more accurate and reliable models. The toolkit also aids in surface and prioritizing valuable data, which is crucial for effective labeling and model training.

Python 460 stars

himovo/movo

Turn DeepSeek Harness into a self-hosted enterprise Agent platform with knowledge, deep research, content generation, vibe coding, browser automation, governance, and admin controls.

### Overview: MOVO is a self-hosted enterprise Agent platform built on DeepSeek Harness (DSH). It combines the DSH Runtime, Skills, Tools, and MCP ecosystem with a deployable user workspace, enterprise knowledge, identity and access control, administration, governance, and file delivery. ### Key features: - **MOVO Community Edition**: A self-hosted enterprise Agent platform built on DeepSeek Harness. - **Enterprise Knowledge**: Integrates enterprise-level knowledge management. - **Identity and Access Control**: Secure user authentication and access management. - **Admin Controls**: Full administrative capabilities for managing the platform. - **Governance**: Ensures compliance and adherence to enterprise standards. - **File Delivery**: Efficient file transfer and sharing within the enterprise. ### Tech stack: Python ### Why it matters: MOVO simplifies the process of deploying and managing DSH Agents in an enterprise environment, providing a robust and secure platform without requiring developers to handle the complexities of authentication, knowledge management, administration, governance, and file delivery themselves. This makes it ideal for teams looking to transition from development experiments to production-level use of DSH Agents.

Python 460 stars

jingh-ai/ultralytics-YOLO-DeepSort-ByteTrack-PyQt-GUI

a GUI application, which uses YOLOs (YOLOv8, YOLO11, YOLOv13) for Object Detection/Tracking, Human Pose Estimation/Tracking from images, videos or camera

Overview: This is a GUI application that utilizes ultralytics YOLO for Object Detection/Tracking, Human Pose Estimation/Tracking, and Segmentation from images, videos, or cameras. It supports YOLO models such as YOLO26, YOLOv13, and YOLO11, as well as DeepSort and ByteTrack trackers. Key features: - Object Detection - Pose Estimation - Segmentation - Oriented Bounding Boxes Tech stack: Python Why it matters: This application provides a user-friendly interface for performing advanced computer vision tasks such as object detection, pose estimation, and segmentation, making it ideal for developers and researchers looking to integrate these capabilities into their projects.

Unknown 459 stars

FrancescoSaverioZuppichini/ViT

Implementing Vi(sion)T(transformer)

Overview: The repository `FrancescoSaverioZuppichini/ViT` contains the implementation of the ViT (Vision Transformer) model, a groundbreaking architecture for image recognition. This project provides a step-by-step guide on how to build and understand the ViT model, including data preprocessing, patch embedding, and the transformer architecture. Key features: - Data preprocessing using PyTorch - Implementation of ViT architecture - Detailed explanation of each component of the transformer - Code structure and documentation Tech stack: The project uses PyTorch for the implementation and includes basic Python packages for image processing and plotting. Why it matters: This project is valuable for developers and researchers interested in understanding and implementing advanced image recognition models, particularly the ViT architecture, which has significantly improved image classification and recognition capabilities. Repository: The repository is hosted on GitHub and has received 459 stars, indicating a high level of interest and engagement from the developer community. GitHub stars: 459

Python 459 stars

Zhen-Dong/Magic-Me

Codes for ID-Specific Video Customized Diffusion

### Overview: Magic-Me is a cutting-edge project that leverages advanced video generation techniques to create personalized videos featuring your loved ones. By training an ID token with pre-existing images, it allows users to generate videos with specific identities, enhancing the authenticity and emotional impact of their memories. ### Key Features: - **Customized Diffusion (VCD)**: A novel framework that generates videos with specified identities using pre-trained ID tokens. - **ID Module**: Trained to disentangle ID information from background noise, ensuring accurate and detailed video generation. - **Text-to-Video (T2V) Module**: Enhances inter-frame consistency and quality through 3D Gaussian Noise Prior. - **Face Deblurring and Upscaling**: Deblurs faces and upscals videos for higher resolution, improving overall video quality. ### Tech Stack: Python ### Why It Matters: Magic-Me offers a unique solution for personalized video generation, making it a valuable tool for anyone looking to create meaningful and emotionally resonant videos. Its ability to preserve and enhance the identity of the subjects ensures that the videos remain true to the original memories and loved ones. Repository: Zhen-Dong/Magic-Me GitHub Stars: 459 GitHub Description: Codes for ID-Specific Video Customized Diffusion News Update: Deployed on Hugging Face's GPU platform, available for immediate use. Check it out on [Hugging Face Spaces](https://huggingface.co/spaces/visionMaze/Magic-Me).

Python 459 stars

afiaka87/clip-guided-diffusion

A CLI tool/python module for generating images from text using guided diffusion and CLIP from OpenAI.

### Overview: This project is a CLI tool and Python module designed for generating images from text using guided diffusion and CLIP from OpenAI. It allows users to create high-quality images by combining text descriptions with pre-trained models. ### Key features: - **Text to Image Generation**: Quickly generate images from text prompts. - **Multiple Prompts and Weights**: Combine multiple prompts with different weights for enhanced results. - **GPU and CPU Support**: Utilizes GPU for faster performance and CPU for slower but more accurate results. - **Timestep Resampling**: Adjusts the number of timesteps to balance between accuracy and speed. - **Blending with Existing Images**: Blend an image with the diffusion process for a specified number of steps. ### Tech stack: Python ### Why it matters: This tool is essential for artists, designers, and anyone looking to create visually interesting images from text descriptions. It simplifies the process of generating high-quality images, making it accessible to a wide range of users. ### Repository: afiaka87/clip-guided-diffusion ### GitHub Stars: 459 ### Install: ```sh git clone https://github.com/afiaka87/clip-guided-diffusion.git cd clip-guided-diffusion uv sync ``` ### Run: ```sh uv run cgd --image_size 256 --prompts "32K HUHD Mushroom" ``` ### Usage - CLI: - **Text to Image Generation**: `uv run cgd --image_size

Python 459 stars

erreurt/MahjongAI

Japanese Riichi Mahjong AI agent. (Feel free to extend this agent or develop your own agent)

Overview: This project is a Mahjong AI agent for Japanese Riichi Mahjong, designed as part of a master's thesis. The agent aims to improve upon existing AI systems by using bagging of multilayer perceptrons and an enhanced opponents model. Key features: - Implementing a Mahjong AI agent - Using bagging of multilayer perceptrons for opponents model - Enhancing opponents model through LSTM training Tech stack: Python Why it matters: This project aims to enhance the performance of Mahjong AI agents by addressing the complexities of the game and improving the opponents model, potentially leading to better strategic decisions in the game.

Lua 459 stars

gaohuang/MSDNet

Multi-Scale Dense Networks for Resource Efficient Image Classification （ICLR 2018 Oral）

Overview: This repository contains the code for the paper "Multi-Scale Dense Networks for Resource Efficient Image Classification" presented at ICLR 2018. Key features: - Multi-Scale Dense Networks (MSDNet) - Anytime Prediction - Batch Computational Budget Setting Tech stack: Lua Why it matters: This project showcases a novel network architecture that efficiently classifies images with limited computational resources, making it particularly useful in resource-constrained environments or real-time applications. Repository: gaohuang/MSDNet GitHub stars: 459

JavaScript 459 stars

genielabs/HomeGenie

HomeGenie: The Programmable Intelligence with 100% Local Agentic AI.

Overview: HomeGenie is an open-source, cloud-independent ecosystem designed for smart automation, providing a complete, privacy-focused, local-first experience. It includes firmware for DIY devices, an efficient backend service, and dedicated mobile and tablet apps. Key features: - 100% Local Agentic AI: Run state-of-the-art LLMs directly on your HomeGenie server for autonomous reasoning, context awareness, and natural language control. - AI Vision Suite: Full integration of YOLO (Object Detection, Instance Segmentation, Pose Estimation) directly on server, ESP32-CAM modules, and generic IP cameras. - Universal API & Programmability: Automate anything with a fluent API programmable in C#, JavaScript, and Python. Create custom programs and widgets with full developer tools. - Advanced Scheduler with Genie Commands: Supports extended cron expressions, variables, conditions, and AI-driven natural language tasks for flexible automations. - Customizable Dashboard & Widgets: A modern, responsive UI with a powerful built-in editor to create and customize your own widgets using HTML, JavaScript, and CSS. - Multi-protocol Support: Integrated drivers for X10, Z-Wave, ZigBee, GPIO, SPI, I2C, IR/RF, and more. - Smart Display & FPV Car Integration: Transform ESP32 hardware into interactive smart displays or AI-powered FPV robotic platforms. - Visual Programming: Create complex scenarios intuitively with the Visual Program editor—no coding skills required. - Local Speech Recognition & Synthesis: Use Whisper recognition and Piper voices in

Dockerfile 459 stars

jim60105/docker-whisperX

Dockerfile for WhisperX: Automatic Speech Recognition with Word-Level Timestamps and Speaker Diarization (Dockerfile, CI image build and test)

Overview: This project provides a Docker image for the WhisperX Automatic Speech Recognition tool, focusing on efficient Docker build and CI workflow management. Key features: - WhisperX Automatic Speech Recognition - Word-Level Timestamps - Speaker Diarization - Continuous Integration (CI) Docker Build Tech stack: Dockerfile Why it matters: This project streamlines the CI workflow for WhisperX, ensuring smooth and efficient Docker image building, particularly for large models and high parallel processing requirements.

TypeScript 458 stars

crosswk/SayIt

Open-source voice typing for Windows — a Wispr Flow / Superwhisper alternative. Press a shortcut, speak, and AI-polished text lands at your cursor. Local models, your own API keys, or a self-hosted backend.

### Overview: SayIt is an open-source voice typing tool designed for Windows, allowing users to dictate text and have it appear at their cursor, making the process of writing more efficient and customizable. ### Key Features: - **Voice Typing Anywhere**: Dictate into any Windows application, including editors, chat apps, and web browsers. - **Editable AI Cleanup**: Remove filler words, correct recognition errors, and format ideas. Every prompt is yours to change. - **Context-Aware Writing**: Reads the text around your cursor to match its tone and terminology. Select text first, and your speech becomes an editing instruction. - **Flexible Speech Recognition**: Supports cloud ASR providers, local GGUF models, and self-hosted backends. - **English and Chinese Interface**: The UI adapts to your system language and can be switched at any time. - **Hotwords and Per-App Rules**: Improve names and technical terms, then change cleanup behavior for different apps. - **Overlay Feedback**: A waveform overlay shows recording state and elapsed time, with optional live captions while you speak. - **Transparent Data Flow**: The app shows which mode is active and where audio and text are processed. - **Local History and Diagnostics**: Review recordings, re-transcribe them, and collect useful troubleshooting details without guesswork. ### Tech Stack: TypeScript ### Why It Matters: SayIt simplifies the process of using AI by turning speech into text that can be used immediately, while allowing users to retain control over important choices such as the tone and terminology of their writing. It supports various speech recognition

Python 458 stars

inclusionAI/Ming-UniAudio

Ming-UniAudio: Speech LLM for Joint Understanding, Generation and Editing with Unified Representation

### Overview Ming-UniAudio is a unified framework that integrates speech understanding, generation, and editing. It utilizes a unified continuous speech tokenizer to achieve balanced performance in both understanding and generation tasks. The model is designed to handle complex semantic and acoustic modifications without manual region specification, making it the first to enable universal free-form speech editing. ### Key Features - **Unified Continuous Speech Tokenizer**: Ming-UniAudio introduces a unified continuous speech tokenizer based on a VAE framework with a causal Transformer architecture, the first continuous speech tokenizer to effectively integrate semantic and acoustic features. This tokenizer is essential for a closed-loop system with LLMs. - **Unified Speech Language Model**: We pretrain an end-to-end unified speech language model with a single LLM backbone, enhanced with a Diffusion Head for high-quality speech synthesis. - **Instruction-Guided Free-Form Speech Editing**: Ming-UniAudio is the first to enable universal, free-form speech editing guided solely by natural language instructions, handling complex semantic and acoustic modifications without manual region specification. ### Tech Stack - Python ### Why It Matters Ming-UniAudio represents a significant advancement in the field of speech processing and language modeling, offering a unified approach to understanding, generating, and editing speech. Its ability to handle complex semantic and acoustic modifications without manual region specification makes it a game-changer in the field of speech synthesis and editing.

Jupyter Notebook 458 stars

ray-project/ray-educational-materials

This is suite of the hands-on training materials that shows how to scale CV, NLP, time-series forecasting workloads with Ray.

### Overview: Welcome to a collection of education materials focused on Ray, a distributed compute framework for scaling your Python and machine learning workloads from a laptop to a cluster. ### Key features: - Hands-on training materials - Covers scaling CV, NLP, and time-series forecasting workloads - Demonstrates the Ray AI Runtime - Includes introductory notebooks, core notebooks, and observability notebooks ### Tech stack: Jupyter Notebook ### Why it matters: These materials provide practical, hands-on experience with Ray, enabling developers to efficiently scale their Python and machine learning workloads across multiple nodes, enhancing performance and scalability.

Python 457 stars

Joyce94/LLM-RLHF-Tuning

LLM Tuning with PEFT (SFT+RM+PPO+DPO with LoRA)

Overview: This project implements RLHF training in three stages and provides detailed implementation details. It supports instruction tuning for the Alpaca model, training a reward model, and training RL models using PPO, DPO, and RM with LoRA. Key features: - Instruction tuning for the Alpaca model - Training a reward model - Support for PPO algorithm training RL models - Supports two base models and two LoRA adapters, loading RM, SFT, Actor, and Critic models - Supports one base model and two LoRA adapters, loading RM, SFT, Actor, and Critic models - Supports one base model and one LoRA adapter, loading RM, SFT, Actor, and Critic models - Support for DPO algorithm training models Tech stack: Python Why it matters: This project provides a comprehensive solution for RLHF training, including support for different models, training algorithms, and LoRA adapters, making it a valuable resource for developers looking to improve their models' performance. Repository: Joyce94/LLM-RLHF-Tuning Primary language: Python GitHub stars: 453 GitHub description: LLM Tuning with PEFT (SFT+RM+PPO+DPO with LoRA) README excerpt: # LLM-RLHF-Tuning 本项目从零实现了RLHF三阶段训练，并在文档中详细写了实现细节，欢迎大家交流讨论[WeChat](assets/RLHF讨论群.png) ### 主要内容： - 支持指令

C++ 457 stars

NLPOptimize/flash-tokenizer

EFFICIENT AND OPTIMIZED TOKENIZER ENGINE FOR LLM INFERENCE SERVING

### Overview: FlashTokenizer is an efficient and optimized tokenizer library for language modeling inference serving, providing a significant speed and accuracy advantage over existing implementations. ### Key features: - High-performance tokenizer implementation for language models. - 10x faster than `BertTokenizerFast` in transformers. - Built-in support for FlashAttention and FlashInfer. - Easy installation via `pip`. ### Tech stack: C++ ### Why it matters: FlashTokenizer is the fastest and most accurate tokenizer for language models, making it essential for efficient inference serving in AI applications.

Python 457 stars

QwenAudio/FunCineForge

### Overview: Fun-CineForge is a unified dataset pipeline and model designed for zero-shot movie dubbing in diverse cinematic scenes. It includes a large-scale Chinese television dubbing dataset, CineDub-CN, and an MLLM-based dubbing model that consistently outperforms state-of-the-art methods in audio quality, lip-sync, timbre transition, and instruction following. ### Key features: - Large-scale Chinese television dubbing dataset, CineDub-CN - Unified dataset pipeline for diverse cinematic scenes - End-to-end model for zero-shot movie dubbing - Consistent performance in audio quality, lip-sync, timbre transition, and instruction following ### Tech stack: Python ### Why it matters: Fun-CineForge provides a comprehensive solution for producing high-quality movie dubbing datasets and models, making it a valuable resource for researchers and developers in the field of audio and visual processing.