AI Trending

Project directory

Projects

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

Jupyter Notebook 727 stars

moein-shariatnia/OpenAI-CLIP

Simple implementation of OpenAI CLIP model in PyTorch.

### Overview: This project is a simple implementation of the OpenAI CLIP model in PyTorch. It aims to provide a faithful replica of the paper's contrastive loss that remains stable even with multiple captions per image. ### Key Features: - CLIP model implementation in PyTorch. - Normalizes both image and text projections before computing logits. - Builds symmetric cross-entropy targets from the sample `id` column. - Updates the dataset, dataloaders, scripts, and notebook to prevent the model from inventing its own targets. ### Tech Stack: PyTorch ### Why it Matters: This implementation is valuable for developers and researchers interested in understanding and implementing CLIP models. It provides a straightforward way to experiment with and test the model's functionality, contributing to the field of multimodal learning and natural language processing.
Python 727 stars

swz30/MIRNet

[ECCV 2020] Learning Enriched Features for Real Image Restoration and Enhancement. SOTA results for image denoising, super-resolution, and image enhancement.

### Overview: MIRNet is a state-of-the-art architecture designed for real image restoration and enhancement. It achieves remarkable results on five real image benchmark datasets, including image denoising, super-resolution, and image enhancement. ### Key Features: - **Multi-scale Residual Block**: Captures rich contextual information from multiple scales. - **Parallel Multi-resolution Convolution Streams**: Extracts multi-scale features. - **Information Exchange Across Streams**: Facilitates the integration of low-resolution and high-resolution information. - **Spatial and Channel Attention Mechanisms**: Enhances both spatial and channel attention for better contextual understanding. - **Enriched Feature Learning**: Combines contextual information from multiple scales to preserve high-resolution spatial details. ### Tech Stack: Python ### Why it Matters: MIRNet is a significant advancement in the field of image restoration and enhancement, offering superior performance compared to existing methods. Its ability to maintain spatially precise high-resolution representations while incorporating strong contextual information makes it a valuable tool for various applications in surveillance, computational photography, medical imaging, and remote sensing.
Unknown 726 stars

SamBelkacem/AI-ML-cheatsheets

Cheatsheets for AI and Machine Learning

**Overview:** This repository contains cheatsheets and quick reference guides for Artificial Intelligence (AI) and Machine Learning (ML), organized into folders based on various topics. **Key features:** - Mathematics: Includes cheat sheets on calculus, linear algebra, probability, statistics, and more. - C++: Reference cards for C++ programming. - Python: Reference cards for Python programming, including beginner guides, OOP, NumPy, Pandas, Matplotlib, and Scikit-Learn. **Tech stack:** The repository uses Markdown for its content and GitHub Pages for hosting. **Why it matters:** These cheatsheets provide a quick and accessible way to learn essential concepts in AI and ML, making it easier for developers to dive into the field. **Repository:** SamBelkacem/AI-ML-cheatsheets **GitHub stars:** 725 **Primary language:** Unknown **GitHub description:** Cheatsheets for AI and Machine Learning **Google Drive version:** Available [here](https://drive.google.com/drive/folders/1yReFa2Af9icmdEaAYOUocepkQk_gGJJK?usp=sharing). **AI-ML-roadmap:** The global structure follows a roadmap for AI and Machine Learning. **Structure of the folder:** - 01- Mathematics - 02- C++ - 03- Python - 04- Computer architecture - 05- Data structures - 06- Automata theory - 06- Complexity theory -
Python 726 stars

aledipa/Free-GPT4-WEB-API

Not just GPT4! Easy to use, Self-Hosted, Unlimited and Free WEB API of the latest A.I. like Gemini, DeepSeek, Claude and GPT

Overview: This project is a self-hosted web API that provides access to modern language models like Gemini, DeepSeek, Claude, and GPT. It offers unlimited requests and is free to use, making it easy to integrate AI into web applications. Key features: - Free to use - Unlimited requests - Simple HTTP interface - Optional web GUI for configuration - Docker support - Smart timeout handling Tech stack: Python Why it matters: This project simplifies the integration of AI into web applications by providing a straightforward, self-hosted API that can be easily configured and managed. Repository: aledipa/Free-GPT4-WEB-API GitHub stars: 726 GitHub description: Not just GPT4! Easy to use, Self-Hosted, Unlimited and Free WEB API of the latest A.I. like Gemini, DeepSeek, Claude and GPT README excerpt: <img src="./img/FreeGPT4_Banner(Nicoladipa).png" style="width:100%; max-width:500px; min-width:200px; aspect-ratio: 5/2"/> [![Docker Image CI](https://github.com/aledipa/Free-GPT4-WEB-API/actions/workflows/docker-image.yml/badge.svg)](https://github.com/aledipa/Free-GPT4-WEB-API/actions/workflows/docker-image.yml) [![GPT4-API-PyApp](https://github.com/aledipa/Free-GPT4-WEB-API/actions/workflows
Rust 726 stars

pegainfer-project/pegainfer

Pure Rust + CUDA LLM inference engine — no PyTorch, OpenAI-compatible, serves Qwen3 to Kimi-K2

Overview: PegaInfer is a pure Rust-based LLM inference engine designed for high-performance inference. It leverages CUDA for optimal performance and serves as an OpenAI-compatible alternative, eliminating the need for PyTorch or Python runtime. Key features: - Pure Rust implementation - CUDA-accelerated inference - No PyTorch or Python runtime - Supports multiple models Tech stack: Pure Rust Why it matters: PegaInfer offers a high-performance, production-ready solution for LLM inference, ensuring optimal performance and reducing the need for complex Python or PyTorch setups. Repository: pegainfer-project/pegainfer GitHub stars: 724 GitHub description: Pure Rust + CUDA LLM inference engine — no PyTorch, OpenAI-compatible, serves Qwen3 to Kimi-K2 Documentation: pegainfer.org Slack community: join.slack.com License: Apache 2.0 </details>
Python 726 stars

tanjeffreyz/auto-maple

Artificial intelligence for MapleStory that uses machine learning and computer vision to navigate challenging in-game environments

### Overview: Auto Maple is an intelligent AI designed to play MapleStory using machine learning and computer vision techniques. It navigates challenging in-game environments by simulating key presses and utilizing TensorFlow and OpenCV to track the player's position and execute routines efficiently. ### Key Features: - **Command Books**: Created by community members, these books contain detailed instructions for each class and map, allowing the AI to perform specific actions effectively. - **Routines**: User-created CSV files that define the path and commands to be executed at each location, enhancing the AI's ability to perform specific tasks efficiently. - **Minimap**: Uses OpenCV template matching to accurately track the minimap and player's position, optimizing routine execution by calculating the shortest path using A* search algorithm. ### Tech Stack: Auto Maple is built using Python, TensorFlow, OpenCV, and other machine learning and computer vision libraries. ### Why It Matters: Auto Maple offers a powerful tool for MapleStory players looking to enhance their gaming experience by automating repetitive tasks and improving the accuracy of their actions. Its modular design allows for customization and modularity, making it suitable for various characters and maps.
Unknown 726 stars

ymcui/MacBERT

Revisiting Pre-trained Models for Chinese Natural Language Processing (MacBERT)

Overview: MacBERT is a pre-trained model that addresses the inconsistency between pre-training and downstream tasks by introducing a masked language model (MLM as correction) pre-training task. This model uses similar words to replace the [MASK] token, which is not present in downstream tasks. The algorithm is based on Word2Vec and WordNet. Key features: - Introduction of a masked language model (MLM as correction) pre-training task. - Use of similar words to replace [MASK] tokens. - Whole Word Masking (wwm) and N-gram masking techniques. - Seamless transition from existing code without modifications. Tech stack: Python, PyTorch, Hugging Face Transformers. Why it matters: MacBERT significantly improves performance on various NLP tasks by addressing the inconsistency between pre-training and downstream tasks.
Python 726 stars

yzhao062/agent-style

21 writing rules for AI coding and writing agents. Drop-in for Claude Code, Codex, Copilot, Cursor, and Aider, so their output reads like a tech pro.

### Overview: This project provides a set of 21 writing rules designed for AI agents to emulate the style of technical professionals. The rules are based on classic writing guides and observations from field studies, aiming to enhance the readability and professionalism of AI-generated content. ### Key Features: - 21 Writing Rules: Each rule is designed to improve the clarity and professionalism of technical writing. - Rules from Classic Guides: 12 rules from classic writing guides. - Field Observations: 9 rules based on observations of large language model outputs. - 2022-2026: Rules developed during this period. - House Style: A personal style guide derived from the author's previous research papers, which the agent follows to maintain consistency. - Focus on Generation Time: The rules are applied during the generation process, not as post-hoc linters. ### Tech Stack: Python ### Why It Matters: These rules help AI agents produce content that reads like it was written by a seasoned technical professional, enhancing the credibility and professionalism of the output.
Python 725 stars

AGI-Arena/MARS

The official implementation of MARS: Unleashing the Power of Variance Reduction for Training Large Models

### Overview: MARS is a unified optimization framework designed to address the challenges of training large models. It combines variance reduction techniques with scaled stochastic gradient methods to accelerate the training process. ### Key Features: - **MARS-AdamW**: Uses a scaled stochastic gradient with gradient correction and a preconditioned update for better gradient complexity. - **MARS-Lion**: Approximates the Hessian matrix using the squared gradients, leading to a simpler update. - **MARS-Shampoo**: Provides a preconditioned update that approximates second-order Newton's method for improved per-iteration complexity. ### Tech Stack: Python ### Why It Matters: MARS offers a more efficient and effective way to train large models by reducing stochastic gradient variance and accelerating optimization. This makes it particularly useful for training models that require significant computational resources, such as those used in fine-tuning pre-trained models or large language models. Repository: AGI-Arena/MARS Primary Language: Python GitHub Stars: 724 GitHub Description: The official implementation of MARS: Unleashing the Power of Variance Reduction for Training Large Models
Python 725 stars

FireRedTeam/FireRedASR2S

A SOTA Industrial-Grade All-in-One ASR system with ASR, VAD, LID, and Punc modules. FireRedASR2 supports Chinese (Mandarin, 20+ dialects/accents), English, code-switching, and both speech and singing ASR. FireRedVAD supports speech/singing/music in 100+ langs. FireRedLID supports 100+ langs and 20+ zh dialects. FireRedPunc supports zh and en.

Overview: FireRedASR2S is a state-of-the-art industrial-grade all-in-one ASR system featuring ASR, VAD, LID, and Punc modules. It excels in speech and singing transcription for Chinese, English, and code-switching, achieving exceptional performance on various benchmarks. Key features: - FireRedASR2: Supports Chinese (Mandarin and 20+ dialects/accents), English, and code-switching with a CER of 2.89% on 4 public Mandarin benchmarks and 11.55% on 19 Chinese dialects and accents benchmarks, surpassing Doubao-ASR, Qwen3-ASR-1.7B, Fun-ASR, and Fun-ASR-Nano-2512. - FireRedVAD: Detects speech/singing/music in 100+ languages with a F1 score of 97.57%, outperforming Silero-VAD, TEN-VAD, FunASR-VAD, and WebRTC-VAD. - FireRedLID: Identifies spoken language in 100+ languages and 20+ Chinese dialects/accents with an accuracy of 97.18%, surpassing Whisper and SpeechBrain. - FireRedPunc: Predicts punctuation for Chinese and English with an average F1 score of 78.90%, outperforming FunASR-Punc (62.77%). Tech stack: Python Why it
Rust 725 stars

Flaminis/Dalaran

Dalaran — Apache-2.0, robotics-first visualization and data infrastructure for multimodal time-series. ROS 2 native, reads existing .rrd recordings. A hard fork of Rerun.

### Overview: Dalaran is an observability and visualization stack designed for robotics-first applications, providing a unified timeline for multimodal time-series data such as robot states, sensor sweeps, and more. It leverages ROS 2 and Arrow-backed chunk stores to efficiently manage and visualize large datasets. ### Key Features: - **Unified Data Management**: Stores data in an Arrow-backed columnar chunk store, allowing for efficient querying and visualization. - **ROS 2 Integration**: Supports ROS 2 topics, LiDAR sweeps, camera streams, IMU, and TF transforms, making it compatible with existing robotics pipelines. - **Real-time Visualization**: Renders data in a 3D/2D viewer that stays in sync across all sensors, providing a seamless experience. - **Data Streaming and Replay**: Records data for later playback, allowing for easy analysis and replay of past events. - **Python, Rust, and C++ SDKs**: Provides a robust set of APIs for developers to integrate Dalaran into their projects. ### Tech Stack: - **Rust**: The primary language used for the core functionality and the CLI. - **Python**: The main entry point for Python developers, providing a simple and intuitive interface. - **C++**: Used for the native C++ SDK, offering a low-level but powerful interface for developers. - **ROS 2**: The robotics operating system, enabling seamless integration with existing robotics infrastructure. - **Arrow**: A columnar storage engine for efficient data management. - **Rosbag2**: A ROS 2 bag file format for replaying data. ###
Python 725 stars

SqueezeAILab/SqueezeLLM

[ICML 2024] SqueezeLLM: Dense-and-Sparse Quantization

### Overview: SqueezeLLM is a post-training quantization framework that enhances the efficiency of large language models (LLMs) by incorporating Dense-and-Sparse Quantization, enabling them to be served with reduced memory footprint, lower latency, and improved accuracy. ### Key Features: - **Dense-and-Sparse Quantization**: Splits weight matrices into a dense component and a sparse component to optimize model performance and memory usage. - **Reduced Precision Quantization**: Enables efficient LLM serving with lower memory footprint and improved accuracy. - **Improved Performance**: Squeezed models can be served with larger memory footprints and higher accuracy, such as the Squeeze variant of Vicuna models serving within 6 GB of memory with 2% higher MMLU than the baseline model in FP16. ### Tech Stack: Python ### Why it Matters: By addressing the challenges of deploying large LLMs with reduced precision quantization, SqueezeLLM offers a more efficient and accurate solution, making it suitable for various applications requiring high performance and low memory usage.
Rust 725 stars

afshinm/zerobox

Lightweight, cross-platform process sandboxing powered by OpenAI Codex's runtime. Sandbox any command with file, network, and credential controls.

Overview: Zerobox is a lightweight, cross-platform process sandboxing tool that allows you to run any command with file, network, and credential controls using OpenAI Codex's sandbox runtime. Key features: - Deny by default: Writes, network, and environment variables are blocked unless explicitly allowed. - Credential injection: Passes API keys that the process never sees, with real values injected only for approved hosts. - File access control: Allows or denies reads and writes to specific paths. - Network filtering: Allows or denies outbound traffic by domain. - Clean environment: Only essential environment variables (PATH, HOME, etc.) are inherited by default. - SDKs for Rust, TypeScript, and Python with a consistent API across languages. - Cross-platform: macOS and Linux. Windows support planned. - Single binary: No Docker, no VMs, approximately 10ms overhead. Why it matters: Zerobox provides a secure and efficient way to sandbox commands, ensuring that sensitive information and writes are restricted, while still allowing necessary network and file access. This makes it ideal for developers working on projects that require isolated environments, such as testing, development, or CI/CD pipelines. Repository: afshinm/zerobox Primary language: Rust GitHub stars: 724 GitHub description: Lightweight, cross-platform process sandboxing powered by OpenAI Codex's runtime.
Rust 725 stars

elastic/mcp-server-elasticsearch

Overview: The Elasticsearch MCP Server is a deprecated tool that connects AI agents to Elasticsearch data using the Model Context Protocol (MCP). It enables natural language interactions without custom APIs. Key features: - Connects AI agents to Elasticsearch indices - Enables querying, analyzing, and retrieving data - Supports stdio and streamable-HTTP protocols Tech stack: Rust Why it matters: The MCP Server is no longer recommended and is being superseded by the Elastic Agent Builder MCP endpoint.
Python 725 stars

flymin/MagicDrive-V2

[ICCV 2025] Official implementation of the paper “MagicDrive-V2: High-Resolution Long Video Generation for Autonomous Driving with Adaptive Control”

Overview: MagicDrive-V2 is an official implementation of the paper "MagicDrive-V2: High-Resolution Long Video Generation for Autonomous Driving with Adaptive Control," accepted at ICCV 2025. This project aims to address the scalability and control challenges in autonomous driving video generation by introducing a novel approach based on the DiT architecture. Key features: - High-resolution and long video generation - Diverse 3D geometry control - Multiview consistency in video generation Tech stack: Python Why it matters: MagicDrive-V2 provides a scalable and flexible solution for generating high-resolution and long videos for autonomous driving applications, addressing the limitations of existing methods in terms of scalability and control conditions.
Python 725 stars

lucidrains/x-clip

A concise but complete implementation of CLIP with various experimental improvements from recent papers

Overview: x-clip is a concise implementation of CLIP, featuring various experimental improvements from recent papers. Key features: - CLIP implementation with additional experimental enhancements - Supports external visual transformers for image encoding - External text transformer for text encoding Tech stack: Python Why it matters: x-clip provides a practical and efficient way to integrate CLIP into projects, leveraging recent advancements in the field. Repository: lucidrains/x-clip GitHub stars: 725 GitHub description: A concise but complete implementation of CLIP with various experimental improvements from recent papers README excerpt: <img src="./clip.png" width="600px"></img>
TypeScript 725 stars

markfulton/NanoBananaEditor

The most advanced Nano Banana image generator and editor application. Your central hub for AI image generation and revisions. Intuitive UI features reference images, editing with image masks, version history, and more. Powered by Gemini Flash images API.

Overview: Nano Banana Editor is a powerful image generator and editor application that allows users to generate and edit images using AI models like Nano Banana 2 and Nano Banana Pro. The intuitive UI features reference images, editing with image masks, version history, and more. Key features: - Generate images from text with reference images and aspect ratios - Edit images with a sentence and follow-up edits carry context - Mask regions with a brush and eraser for local changes - Search for prompts with real-world facts - Control thinking on Flash models for different levels of depth - Compare before and after with a slider and 4K re-render - Save history and per-item settings, download results, and copy to clipboard - Support for light and dark themes and keyboard control Tech stack: React + TypeScript Why it matters: Nano Banana Editor provides a centralized platform for generating and editing images using AI models, making it a valuable tool for developers and designers who need to create and refine visual content quickly and efficiently. Repository: markfulton/NanoBananaEditor GitHub stars: 722 GitHub description: The most advanced Nano Banana image generator and editor application. Your central hub for AI image generation and revisions. Intuitive UI features reference images, editing with image masks, version history, and more. Powered by Gemini Flash images API.
Python 724 stars

Algolzw/image-restoration-sde

Image Restoration with Mean-Reverting Stochastic Differential Equations, ICML 2023. Winning solution of the NTIRE 2023 Image Shadow Removal Challenge.

### Overview: This project introduces Image Restoration with Mean-Reverting Stochastic Differential Equations (IR-SDE), a winning solution to the NTIRE 2023 Image Shadow Removal Challenge. The method uses mean-reverting stochastic differential equations to restore images, achieving state-of-the-art results on synthetic and real-world datasets. ### Key Features: - **IR-SDE**: A winning solution to the NTIRE 2023 Image Shadow Removal Challenge. - **Refusion**: Enables large-size realistic image restoration using latent-space diffusion models. - **Universal Image Restoration Framework**: Controls vision-language models to achieve better results. ### Tech Stack: Python ### Why It Matters: IR-SDE and Refusion are groundbreaking methods that significantly improve image restoration, particularly in challenging scenarios such as shadow removal and large-scale image restoration. Their applications are widely applicable across various domains, including synthetic and real-world datasets. ### Repository: Repository: Algolzw/image-restoration-sde GitHub Stars: 724 GitHub Description: Image Restoration with Mean-Reverting Stochastic Differential Equations, ICML 2023. Winning solution of the NTIRE 2023 Image Shadow Removal Challenge. ### Dependencies: - OS: Ubuntu 20.04 - NVIDIA: CUDA 11.7, cudnn 8.5.0 - Python 3 - PyTorch >= 1.13.0 - Python packages: `pip install -r requirements.txt` ### How to Use Our Code: - **Example for Image Deraining**:
Python 724 stars

ThibaultGROUEIX/AtlasNet

This repository contains the source codes for the paper "AtlasNet: A Papier-Mâché Approach to Learning 3D Surface Generation ". The network is able to synthesize a mesh (point cloud + connectivity) from a low-resolution point cloud, or from an image.

Overview: AtlasNet is a Papier-Mâché approach designed for generating 3D surface meshes from low-resolution point clouds or images. This project was presented at CVPR 2018 and utilizes PyTorch, PyMesh, and CUDA 10.1 for efficient implementation. Key features: - Synthesizes meshes from low-resolution point clouds - Converts images into 3D surface representations - Utilizes Papier-Mâché technology for efficient learning Tech stack: Python, PyTorch, PyMesh, CUDA 10.1 Why it matters: AtlasNet offers a novel method for generating 3D surfaces from point clouds and images, demonstrating the potential of Papier-Mâché techniques in computer vision and 3D reconstruction.
TypeScript 724 stars

comet-ml/opik-openclaw

🦞 Official plugin for OpenClaw that exports agent traces to Opik. See and monitor agent behaviour, cost, tokens, errors and more.

Overview: This plugin allows developers to monitor and observe agent behavior, cost, tokens, and errors in OpenClaw runs, providing insights into LLM request/response spans, sub-agent interactions, tool calls, and more through the Opik observability platform. Key features: - Agent traces and metrics - OpenClaw integration - Opik integration - Observability and monitoring Tech stack: TypeScript Why it matters: This plugin enables developers to easily integrate and monitor OpenClaw runs, providing valuable insights into the performance and behavior of their models, allowing for better optimization and troubleshooting.