Projects

### Overview: Interactive Deep Colorization is a software tool that leverages deep learning to colorize black and white images. This project introduces a method that allows users to colorize images with just a few clicks, making the process intuitive and accessible. ### Key Features: - **Real-Time User-Guided Image Colorization**: The software automatically colors black and white images based on user interactions. - **GUI and iPython Notebooks**: Provides both a "barebones" version in an iPython notebook and a full GUI demo. - **Supports Multiple Backends**: The system supports both Caffe and PyTorch models, offering flexibility and compatibility. ### Tech Stack: - Python ### Why It Matters: Interactive Deep Colorization simplifies the process of image colorization, making it accessible to a wide range of users. The ability to colorize images with minimal user interaction enhances the user experience and democratizes the use of deep learning techniques.

### Overview: Awesome Knowledge-Distillation is a repository that compiles a collection of papers on knowledge distillation, a technique that aims to transfer knowledge from a teacher model to a student model. This repository serves as a comprehensive resource for researchers and practitioners interested in knowledge distillation, offering insights into various forms of knowledge distillation, including knowledge from logits, intermediate layers, graph-based methods, and more. ### Key features: - **Knowledge from logits**: Techniques for distilling knowledge from the logits of a teacher model. - **Knowledge from intermediate layers**: Methods for distilling knowledge from the intermediate layers of a teacher model. - **Graph-based**: Approaches that leverage graph-based methods for knowledge distillation. - **Mutual Information & Online Learning**: Techniques that use mutual information and online learning for knowledge distillation. - **Self-KD**: Methods that distill knowledge from the student model itself. - **Structural Knowledge**: Techniques that focus on the structural aspects of knowledge distillation. - **Privileged Information**: Methods that incorporate privileged information for knowledge distillation. - **KD + GAN**: Knowledge distillation with Generative Adversarial Networks. - **KD + Meta-learning**: Knowledge distillation with meta-learning. - **Data-free KD**: Knowledge distillation without relying on labeled data. - **KD + AutoML**: Knowledge distillation in the context of Automated Machine Learning. - **KD + RL**: Knowledge distillation with Reinforcement Learning. - **KD + Self-supervised**: Knowledge distillation using self-supervised learning. - **Multi-teacher and Ensemble KD**:

Overview: This repository contains a collection of pen-and-paper exercises in machine learning, covering topics such as linear algebra, optimisation, graphical models, and inference techniques. Key features: - Detailed solutions for each exercise - Topics covered: linear algebra, optimisation, directed graphical models, undirected graphical models, expressive power of graphical models, factor graphs and message passing, inference for hidden Markov models, model-based learning, sampling and Monte-Carlo integration, variational inference - Available as a compiled PDF on arXiv Tech stack: TeX Why it matters: This collection provides practical, hands-on exercises that complement traditional machine learning textbooks, offering a deeper understanding through practical implementation. Repository: Michael Gutmann's ml-pen-and-paper-exercises GitHub stars: 2,690 GitHub description: Pen and paper exercises in machine learning Primary language: TeX Creative Commons Attribution 4.0 International License: CC BY 4.0 Contributing: Use GitHub issues for reporting mistakes or typos, and contribute to the exercises with detailed solutions.

Overview: LISA is a large language model-assisted segmentation assistant designed for image segmentation tasks. It provides online and offline capabilities, including online demos and local deployment options. Key features: - Online demo for real-time image segmentation - Offline model for local deployment - Segmentation mask output with explanations - Supports various image inputs and outputs Tech stack: Python Why it matters: LISA simplifies complex image segmentation tasks by leveraging large language models, making it accessible and efficient for both researchers and practitioners. Repository: JIA-Lab-research/LISA GitHub stars: 2689 GitHub description: Project Page for "LISA: Reasoning Segmentation via Large Language Model"

Overview: This repository contains a PyTorch implementation of the original Graph Attention Network (GAT) paper by Veličković et al., aimed at making it easy to start playing and learning about GAT and GNNs in general. Key features: - PyTorch implementation of GAT - Playground.py file for visualizing Cora dataset, GAT embeddings, attention mechanism, and entropy histograms - Support for both Cora (transductive) and PPI (inductive) examples Tech stack: PyTorch Why it matters: This repository provides a practical way to understand and experiment with Graph Neural Networks (GNNs) and Graph Attention Networks (GAT), making it accessible for beginners and intermediate users alike.

### Overview: LongLoRA and LongAlpaca are advanced models designed for long-context language understanding. These models are part of the LongLoRA research project, which aims to enhance the performance and efficiency of large language models. The models are optimized for handling long sequences and provide significant improvements in accuracy and speed. ### Key features: - **LongLoRA**: A model that leverages LongLoRA to enhance the performance of large language models. - **LongAlpaca**: A series of models fine-tuned on the LongAlpaca-12k dataset, specifically designed for long context understanding. - **StreamLLM**: A streaming inference method that allows for longer dialogue sequences in models like StreamingLLM. ### Tech stack: Python ### Why it matters: These models significantly improve the performance and efficiency of large language models, making them more suitable for applications requiring long context understanding. They are particularly useful in scenarios where the input or output can be very long, such as in chatbots, summarization, or any task where the context is extensive. ### Repository: JIA-Lab-research/LongLoRA ### Primary language: Python ### GitHub stars: 2687 ### GitHub description: Code and documents of LongLoRA and LongAlpaca (ICLR 2024 Oral) ### README excerpt: <p align="center" width="100%"> <img src="imgs/LongAlpaca.png" alt="Stanford-Alpaca" style="width: 100%; min-width: 30

### Overview: InfiniteYou: Flexible Photo Recrafting While Preserving Your Identity is a project that leverages advanced Diffusion Transformers (DiTs) to achieve flexible and high-fidelity identity-preserved image generation. The project addresses significant issues in existing methods, such as insufficient identity similarity, poor text-image alignment, and low generation quality and aesthetics. The key feature is InfuseNet, which injects identity features into the DiT base model via residual connections, enhancing identity similarity while maintaining generation capabilities. ### Key Features: - InfuseNet: A component that injects identity features into the DiT base model via residual connections, enhancing identity similarity while maintaining generation capabilities. - Multi-stage training strategy: Includes pretraining and supervised fine-tuning (SFT) with synthetic single-person-multiple-sample (SPMS) data, improving text-image alignment, image quality, and alleviating face copy-pasting issues. ### Tech Stack: Python ### Why It Matters: InfiniteYou demonstrates significant progress in achieving flexible and high-fidelity identity-preserved image generation, particularly with advanced DiTs like FLUX. The project's design and implementation offer a valuable contribution to the broader community, making it a valuable resource for developers and researchers in the field of image generation and identity preservation.

### Overview: Conquer Any Code in VSCode: One-Click Comments, Conversions, UI-to-Code, and AI Batch Processing of Files! 💪 ### Key Features: - 🔄 **Code Convert**: Transform code between any programming languages with one click. - 📖 **Code Viewer Helper**: Add detailed comments to enhance code readability. - 🔧 **Expert Code Enhancer**: Hand your code over to AI for optimization and see how an expert would write it. - 📋 **Smart Paste**: Intelligently convert clipboard content (code or images) when pasting. - 🗂️ **AI Batch Processor**: Process multiple files using AI according to custom requirements. - 🏷 **Rename Variable**: Get AI-suggested variable names with explanations. - 💬 **Ask AI**: Execute custom AI commands on selected files or folders. - 📝 **Copy as AI Prompt**: Easily copy files/folders content as AI prompts. ### Tech Stack: TypeScript ### Why It Matters: This tool simplifies the process of working with code by providing quick and easy solutions for common tasks, making it a valuable addition to any developer's workflow.

Overview: Bright Data MCP provides a powerful Model Context Protocol (MCP) server that enables real-time access to public web data through 32 tools, including web search, page scraping, structured data extraction, browser automation, and LLM response collection. Key features: - Web search - Page scraping - Structured data (Web Scraper API) - Browser automation - LLM response collection Tech stack: JavaScript Why it matters: Bright Data MCP simplifies the process of integrating AI agents and language models with public web data, making it easier to extract structured information and automate tasks across various platforms.

### Overview: The **FMA** dataset, introduced at the International Society for Music Information Retrieval Conference (ISMIR) 2017, is a comprehensive and open-access dataset designed for evaluating various tasks in Music Information Retrieval (MIR). It comprises 917 GiB of high-quality audio from 16,341 artists and 14,854 albums across 161 genres, providing full-length and high-quality audio, pre-computed features, and detailed metadata. ### Key Features: - **Tracks**: Detailed metadata including ID, title, artist, genres, tags, and play counts for 106,574 tracks. - **Genres**: Hierarchical taxonomy of 161 genres with parent genres, aiding in genre inference and top-level genre identification. - **Features**: Common features extracted using **librosa** for 13,129 tracks, pre-computed for efficiency. - **Audio Data**: Available in various sizes: 8,000 tracks of 30s, 8 balanced genres (GTZAN-like) for 7.2 GiB, 25,000 tracks of 30s, 16 unbalanced genres for 22 GiB, and 106,574 tracks of 30s, 161 unbalanced genres for 93 GiB, and 106,574 untrimmed tracks for 879 GiB. ### Tech Stack: -

Overview: This repository is a curated list of resources on graph-based retrieval-augmented generation (GraphRAG), including surveys, papers, benchmarks, and open-source projects. It is continuously updated to stay relevant and useful for researchers and developers in the field of large language models. Key features: - GraphRAG Survey: A curated list of resources on GraphRAG. - MemGraphRAG: A memory-enhanced RAG accepted by KDD'26. - ProbeRAG: A RAG faithfulness accepted by ACL'26. - BAPO: A reliable agentic search accepted by ACL'26. - LegalGraphRAG: A reliable legal reasoning accepted by ACL'26. - LogicPoison: A GraphRAG attack model accepted by ACL'26. - LinearRAG: An efficient GraphRAG accepted by ICLR'26. - GraphRAG Benchmark: An accepted benchmark for evaluating GraphRAG models. - GraphRAG Benchmark Dataset: A dataset for evaluating GraphRAG models. - GraphRAG Survey: A survey accepted by AAAI'26. - LinearRAG: A relation-free graph construction method accepted by ICLR'26. - GraphRAG Benchmark: An accepted benchmark for evaluating GraphRAG models. Tech stack: The repository uses Markdown, GitHub, and various open-source tools for documentation and collaboration. Why it matters: GraphRAG is a significant advancement in the field of large language models, offering a new approach to customized LLM applications. It addresses limitations of traditional RAG

### Overview: UX/UI Agent Skills is a comprehensive kit designed to transform Claude into a senior design architect. It includes structured instructions, design tokens, runnable skills, and 138 design systems that ensure consistent, accessible, and token-driven design outputs across any framework and design system. ### Key features: - **52 Objective Gates**: Measure the performance of the gates and ensure they pass. - **WCAG 2.2 AA→AAA Compliance**: Ensure accessibility standards are met. - **52 Component Specs**: Detailed specifications for 52 components. - **25 Runnable Skills**: Skills that can be run directly in any project. - **14 DTCG Token Files**: Token-driven design for Claude. - **16 Framework Adapters**: Support for various frameworks. - **138 Design Systems**: Consistent design across different systems. ### Tech stack: JavaScript ### Why it matters: UX/UI Agent Skills provides a robust and accessible toolkit for developers, ensuring consistent and high-quality design outputs. It is particularly useful for teams looking to improve their design systems and accessibility standards.

Overview: Sentiment is a Node.js module designed for performing sentiment analysis on text input using the AFINN-165 wordlist and Emoji Sentiment Ranking. It provides performance, flexibility, and the ability to easily add and overwrite words and define custom strategies for negation and emphasis. Key features: - Performance - Ability to append and overwrite word / value pairs from the AFINN wordlist - Easy addition and support for new languages - Custom strategies for negation, emphasis, etc. on a per-language basis Tech stack: JavaScript Why it matters: Sentiment offers a robust and customizable sentiment analysis tool for developers, enhancing their ability to analyze text sentiment in various applications and projects.

Overview: Cloud Annotations is a fast, easy, and collaborative image annotation tool designed for teams and individuals. Key features: - Real-time collaboration - Automatic image classification - Multi-user support - Automatic annotation suggestions Tech stack: TypeScript Why it matters: Cloud Annotations simplifies the process of image annotation for teams and individuals, making it accessible and efficient for both beginners and experienced users. Repository: cloud-annotations/cloud-annotations Primary language: TypeScript GitHub stars: 2677

Overview: MimicMotion is a high-quality human motion video generation framework that addresses the challenges of video length, richness of details, and temporal smoothness in generative artificial intelligence for video generation. It introduces confidence-aware pose guidance, regional loss amplification, and a progressive latent fusion strategy to generate videos of any length with acceptable resource consumption. Key features: - Confidence-aware pose guidance for robust model performance - Regional loss amplification for improved image quality - Progressive latent fusion for generating long, smooth videos Tech stack: Python Why it matters: MimicMotion offers a solution for generating high-quality videos of any length with rich details and smooth temporal transitions, making it a significant advancement in the field of generative artificial intelligence for video generation.

Overview: xTuring is a powerful tool for fine-tuning and evaluating open-source language models (LLMs) like GPT-OSS, LLaMA/LLaMA 2, Qwen3, MiniMax M2, GPT-J, GPT-2, DistilGPT-2, and Mamba. It simplifies the process of personalizing and running these models on your own data, making it easy and efficient. Key features: - Easy data preprocessing and fine-tuning - Personalized models with customizable reasoning levels - Support for various LLMs and models - Private and cost-effective deployment options - Built-in evaluation metrics like perplexity Tech stack: Python Why it matters: xTuring streamlines the process of fine-tuning and evaluating LLMs, making it accessible and cost-efficient for developers and researchers. Its simple API and private deployment options make it ideal for both personal use and enterprise-level projects. Repository: stochasticai/xTuring GitHub stars: 2674 GitHub description: Build, personalize and control your own LLMs. From data pre-processing to fine-tuning, xTuring provides an easy way to personalize open-source LLMs. Join our discord community: https://discord.gg/TgHXuSJEk6

### Overview: EvaDB is a database system designed for AI-powered applications. It simplifies the process of integrating AI models into traditional database applications, enabling developers to build AI apps in a few lines of code. ### Key Features: - **Easy Data Source Integration**: Connects the EvaDB query engine with PostgreSQL, SQLite, MySQL, MariaDB, Clickhouse, Snowflake, and local filesystems. - **Pre-trained AI Models**: Queries connected data with pre-trained AI models from Hugging Face, OpenAI, YOLO, and more. - **Structured and Unstructured Data Support**: Supports structured data sources like PostgreSQL and unstructured data sources such as local filesystems and AWS S3 buckets. ### Tech Stack: Python ### Why It Matters: EvaDB streamlines the integration of AI models into traditional databases, making it easier and more efficient for developers to build AI-powered applications. Its powerful SQL API simplifies the process of developing AI apps, reducing the need for complex coding and improving overall application performance. ### GitHub Stars: 2,674 ### Follow EvaDB: - Join the Slack Community: [Join EvaDB Slack Community](https://evadb.ai/community) - Follow EvaDB on Twitter: [@evadb_ai](https://twitter.com/evadb_ai) - Visit EvaDB on Medium: [EvaDB on Medium](https://evadb.ai/) - Visit EvaDB Website: [EvaDB Website](https://evadb.ai/) - Explore Roadmap: [Roadmap](https://github.com/orgs/georgia-tech-db

### Overview: The FreedomGPT project is a React and Electron-based application that allows users to run LLM models locally on their Mac and Windows machines. It provides a chat-based interface for executing the FreedomGPT LLM, offering offline and private access to the model. ### Key Features: - **Offline and Private Access**: Users can run the LLM model locally without internet connectivity. - **Chat-Based Interface**: A user-friendly chat interface for interacting with the LLM. - **Electron and React Integration**: Built using Electron for the frontend and React for the backend. ### Tech Stack: - **Electron**: The frontend framework for building the user interface. - **React**: The backend framework for handling the chat and model execution. ### Why It Matters: This project is designed to provide a practical solution for users who need to run large language models offline and privately. It offers a convenient way to interact with the LLM without the need for an internet connection, making it ideal for remote work, travel, or environments where internet access is restricted. ### GitHub Stars: The project has received 2,674 GitHub stars, indicating a high level of interest and engagement from the developer community. ### Repository: The repository is hosted on GitHub under the name `ohmplatform/FreedomGPT`. ### Primary Language: The project is primarily written in TypeScript, leveraging its strong typing and type safety features. ### Join Our Discord Community: Join our Discord server to stay updated on the latest developments and interact with the community. ### How to Contribute: - **Prerequisites**: Install

### Overview: This repository serves as a benchmark for evaluating graph neural networks (GNNs), providing a comprehensive framework for researchers to test and compare the performance of different GNN architectures across various datasets. ### Key features: - **Benchmarking Framework**: A structured environment for running and comparing GNN models on diverse datasets. - **Datasets**: Includes datasets like ZINC, AQSOL, GraphTheoryProp, and CYCLES for testing GNNs on specific graph properties. - **Technical Report**: Updated technical report on the framework, available on [ArXiv](https://arxiv.org/pdf/2003.00982.pdf). - **Dependencies**: Updated to DGL 0.6.1 and higher, with environments for CPU and GPU configurations. - **Updates**: Multiple updates since its release, including new datasets and improved experimental pipelines. ### Tech stack: - Jupyter Notebook ### Why it matters: This benchmarking tool is crucial for researchers and developers in the field of graph neural networks, providing a standardized way to evaluate and compare the performance of different GNN architectures across various datasets and configurations.

Overview: OmAgent is a Python library designed for building multimodal language agents, offering a simple and easy-to-use interface for developers and researchers. It simplifies complex engineering tasks like worker orchestration and provides abstractions for reusable agent components, enabling the creation of complex agents from basic components. Key features: - Flexible agent architecture for graph-based workflow orchestration. - Native multimodal interaction support for VLM models, real-time APIs, computer vision models, and mobile device connections. - Suite of state-of-the-art unimodal and multimodal agent algorithms. - Local deployment of models with Ollama or LocalAI. - Fully distributed architecture with custom scaling support. - Lite mode for eliminating middleware deployment. Tech stack: Python Why it matters: OmAgent simplifies the process of building multimodal language agents, making it accessible and efficient for developers and researchers without the need for extensive engineering knowledge. Repository: om-ai-lab/OmAgent GitHub stars: 2669