Projects

Overview: ReAlign is a reformatted alignment method designed to enhance the quality of instruction data, making it better aligned with human values and preferences. Key features: - Reformatting instruction data to improve alignment with pre-established criteria and evidence. - Boosts mathematical reasoning, factuality, and readability of language models. - Enhances general alignment ability without additional data or advanced training techniques. Tech stack: JavaScript Why it matters: ReAlign offers a simple yet effective solution for improving the alignment of language models, highlighting the importance of human-LLM collaboration in the alignment process. Repository: GAIR-NLP/ReAlign Primary language: JavaScript GitHub stars: 111 GitHub description: Reformatted Alignment

Overview: This is the official PyTorch implementation of "TFMQ-DM: Temporal Feature Maintenance Quantization for Diffusion Models," a novel training-free framework that achieves state-of-the-art results in precision-4 weight quantization of diffusion models, particularly on CelebA-HQ. It significantly reduces the FID score by 6.7 and accelerates the quantization process. Key features: - TFMQ-DM framework - Temporal Information Block (TIB) - Temporal Information Aware Reconstruction (TIAR) - Finite Set Calibration (FSC) - Achieves 4-bit weight quantization - Reduces FID score by 6.7 on CelebA-HQ - Accelerates quantization time Tech stack: PyTorch, Jupyter Notebook Why it matters: TFMQ-DM provides a practical solution for quantizing diffusion models without the need for training, making it particularly useful for accelerating the quantization process and improving the performance on specific tasks like CelebA-HQ.

Overview: NVIDIA Elements is a developer toolkit designed for AI factories, robotics, and autonomous vehicles, providing agent-ready tooling, framework agnostic support, and stable API contracts. Key features: - Agent-ready CLI and MCP for component APIs, tokens, examples, imports, validation, and setup. - Framework agnostic support for Web Components across React, Angular, Vue, Svelte, Lit, plain HTML, server-rendered templates, and mixed stacks. - Built for AI infrastructure, offering operational UI for AI/ML workloads, autonomous vehicle tools, and robotics consoles. - Stable API contracts with Skills and lint guide for authoring best practices, common UI patterns, and automated static analysis. Tech stack: TypeScript Why it matters: NVIDIA Elements simplifies the development process for AI and robotics projects, ensuring a consistent and efficient workflow. Repository: NVIDIA/elements GitHub stars: 106 GitHub description: The UI SDK and agent-ready developer toolkit for AI/ML Factories, Robotics, and Autonomous Vehicles.

### Overview: This repository is a curated map of agents that perceive across modalities, reason in a loop, and act on the digital or physical world. It serves as an evidence-gated research and implementation index for multimodal agents, spanning end-to-end systems, enabling research, evaluation resources, open infrastructure, and reusable skills. ### Key Features: - **Agent Models and Policies**: Omni-Modal and Interactive Models, Action Models, Spatial, Memory, and Tool-Use Models. - **Learning and Planning**: Tool Use, Workflow, Program Synthesis, Agentic Learning, Multi-Agent Coordination. - **Cross-Section Reading Paths**: General Multimodal Agency, Embodied Agents and World-Model-Based Planning, Multi-Agent and Human-Agent Collaboration. - **End-to-End Systems**: Information Seeking and Media Understanding, Creative and Multimodal Workflows, Conversation and Social Interaction, Digital and Software Environments, Virtual, 3D, and Open Worlds, Physical Robotics and Multi-Robot Systems. ### Tech Stack: JavaScript ### Why It Matters: This repository provides a comprehensive catalog of multimodal agents, helping researchers, developers, and practitioners to find and evaluate relevant projects, tools, and resources for their work.

### Overview: Blurt is a Gnome shell extension designed for accurate offline speech-to-text input in Linux. It leverages whisper.cpp for transcribing speech into text, providing a user-friendly interface with customizable key bindings and preferences. ### Key Features: - **Local whisper.cpp Installation**: Supports whisper.cpp installed locally. - **Whisper.cpp Server**: Transcribes speech using a whisper.cpp server for improved speed and reliability. - **Preferences and Setup**: Right-click for preferences and setup options. - **Key Bindings**: Start/Stop speech-to-text input with <CTRL+ALT+a> and <MB1 (left) click>. - **Icon Status**: Icon color indicates the status during operation. - **Whisperfile Inference**: Option to transcribe with a whisperfile for even faster processing. ### Tech Stack: Blurt is built on JavaScript, leveraging Gnome shell extensions and whisper.cpp for speech-to-text transcription. ### Why it Matters: Blurt offers a seamless and efficient way to input text from speech, enhancing productivity and accessibility on Linux systems.

Overview: This workshop focuses on building Agentic Workflows on AWS, enabling large language models to autonomously perform tasks or provide assistance. Key features: - Leverage capabilities in Amazon Bedrock to create both chatbot and non-chatbot workflows - Incorporate features such as code interpreter, guardrails, knowledge bases, and more! - Use Kiro to assist in coding and debugging Tech stack: Python Why it matters: This workshop provides a practical guide for developers interested in creating autonomous workflows using AWS, enhancing the capabilities of large language models for various tasks and user assistance.

Overview: LocalSTT is an Android application that provides automatic speech recognition services without requiring an internet connection. It utilizes the Vosk and Mozilla DeepSpeech libraries, and is currently a proof of concept based on Kõnele and Vosk Android Demo. Key features: - Automatic speech recognition without internet connection - Supports multiple languages (catalan in this case) - Built on top of Kõnele and Vosk Android Demo libraries - Pre-built binary available for download with Vosk and DeepSpeech models Tech stack: Java Why it matters: This application demonstrates the potential of local speech recognition, showcasing how it can be integrated into various Android applications without the need for an internet connection, making it suitable for offline use cases.

### Overview: OLIV is a Thai and English voice typing tool for macOS that converts spoken tech words into English, making it easier for Thai speakers to work in tech fields. ### Key Features: - **Thai and English Support:** OLIV supports Thai and English, allowing users to type Thai words in English scripts. - **Offline Dictation:** Dictation works offline, ensuring that users can type Thai words even when no internet connection is available. - **Fast Response Time:** Median response time is approximately 1.0 second, making it suitable for quick typing tasks. - **Compatibility with Multiple Apps:** OLIV works seamlessly in various apps like Slack, VS Code, and Terminal. - **Customizable Dictionary:** Users can add names and jargon to the dictionary, allowing OLIV to spell them correctly even if the speech model gets it slightly wrong. - **Automatic Formatting:** OLIV automatically formats Thai numbers, dates, and other common Thai words. - **Music Ignoring:** OLIV ignores the music playing in the background, ensuring that users can type without being distracted by background noise. - **Free and Open Source:** OLIV is free, MIT-licensed, and doesn't require an account or subscription. ### Tech Stack: - **Typhoon Whisper Turbo:** A Thai fine-tune of Whisper, used for transcribing Thai speech. - **Gemma E2B:** Used for cleaning up the transcribed text, restoring English tech words. - **Rust Core:** The core of the application, responsible for transcribing and cleaning up the speech. ### Why It Matters: OL

Overview: This project is an example of verbal communication using ChatterBot, showcasing text-to-speech (TTS) and speech recognition functionalities. Key features: - Text-to-speech (TTS) using ChatterBot - Speech recognition using Anthony Zhang's SpeechRecognition library - Speech synthesis using Festival Tech stack: Python Why it matters: This project provides a practical demonstration of integrating ChatterBot with speech technologies, demonstrating how text can be converted into spoken language and vice versa, which can be useful for applications requiring conversational AI. Repository: gunthercox/chatterbot-voice Primary language: Python GitHub stars: 111 GitHub description: A example of verbal communication using ChatterBot

Overview: This repository contains the code for Davidsonian Scene Graph (DSG), a new framework for fine-grained text-to-image evaluation introduced in the paper "Davidsonian Scene Graph: Improving Reliability in Fine-grained Evaluation for Text-to-Image Generation" at ICLR 2024. Key features: - DSG-1k score calculation script - Human correlation calculation script on TIFA-160 - Answers to DSG-1k questions by VQA models (Instruct-BLIP/mPLUG-Large/PaLI 17B) on images generated with SDv2.1 Tech stack: Jupyter Notebook Why it matters: DSG provides a robust and reliable evaluation framework for text-to-image models, addressing key challenges in existing QG/A approaches and offering a comprehensive benchmark with 1,060 prompts for fine-grained semantic categories.

Overview: This repository contains all the code and materials for the TwoSetAI YouTube channel, offering a comprehensive guide to AI-related topics. Key features: - Code and tutorials for the channel's videos - AI-related products and trends - In-depth engineering insights - Career advice for AI enthusiasts - Entrepreneurial side of AI Tech stack: Jupyter Notebook Why it matters: This repository serves as a valuable resource for anyone interested in learning about AI, providing a blend of technical knowledge and practical applications.

Overview: This Python Series of tutorials is designed to introduce and build upon AI concepts using Python, starting from the basics and progressing to full-fledged case studies. Key features: - Introduction to Python programming - Basics of Python data types - Understanding inheritance - Building full-fledged AI case studies Tech stack: Jupyter Notebook Why it matters: This series provides a comprehensive learning experience for AI concepts, starting with Python fundamentals and culminating in practical application through real-world data and problem-solving scenarios. Repository: rjrahul24/ai-with-python-series Primary language: Jupyter Notebook GitHub stars: 112

### Overview: Dynamic Prompt Learning (DPL) is a novel approach to address cross-attention leakage in text-based image editing. This method aims to improve the quality and effectiveness of text-based image manipulation by leveraging dynamic prompts and cross-attention mechanisms. ### Key Features: - **Dynamic Prompt Learning (DPL)**: A method for text-based image editing that addresses cross-attention leakage. - **Cross-Attention Leakage**: A common issue in text-based image editing where attention mechanisms can leak information from one part of the image to another, leading to poor results. - **Improved Editing Quality**: DPL aims to enhance the quality of text-based image editing by mitigating cross-attention leakage. - **Cross-Attention Maps**: The method generates cross-attention maps to visualize and analyze attention patterns, aiding in the understanding and improvement of the editing process. ### Tech Stack: - **Python**: The primary programming language used for implementing the DPL method. ### Why It Matters: - **Enhanced Text-Based Image Editing**: DPL addresses a significant problem in text-based image editing, leading to better quality and effectiveness in the editing process. - **Cross-Attention Leakage Mitigation**: By mitigating cross-attention leakage, DPL ensures that the attention mechanisms do not leak information, resulting in more accurate and effective text-based image editing. ### Repository: wangkai930418/DPL - **Primary Language**: Python - **GitHub Stars**: 111 - **GitHub Description**: [NeurIPS 2023] Dynamic Prompt Learning: Address

### Overview: IllumiCraft is an official implementation of the NeurIPS 2025 paper "IllumiCraft: Unified Geometry and Illumination Diffusion for Controllable Video Generation". This project provides the training code, IllumiCraft dataset, model, and inference scripts for generating controllable videos. ### Key Features: - **Training Code**: Provides the training script to generate controllable videos. - **Dataset**: Includes a dataset for training and inference. - **Model**: Offers a base model and a pretrained checkpoint. - **Inference**: Provides scripts for running the model for inference. ### Tech Stack: Python ### Why It Matters: IllumiCraft is a significant contribution to the field of video generation, offering a unified approach to geometry and illumination diffusion that can be used to create controllable videos with high quality and flexibility. The availability of the code and dataset makes it accessible for researchers and developers to experiment and improve upon the method. ### Repository: - **Repository**: [yuanze-lin/IllumiCraft](https://github.com/yuanze-lin/IllumiCraft) - **GitHub Stars**: 112 - **GitHub Description**: [NeurIPS 2025] The official code for "IllumiCraft: Unified Geometry and Illumination Diffusion for Controllable Video Generation" - **Primary Language**: Python ### Installation: ```bash git clone https://github.com/yuanze-lin/IllumiCraft.git cd IllumiCraft conda create -n illumicraft python=3

### Overview: Auto Data is a library designed to automatically generate realistic datasets for fine-tuning Large Language Models (LLMs), addressing the scarcity and imbalance of data by enabling the creation of comprehensive datasets across various topics. ### Key features: - Automatic dataset generation - Support for multiple formats (json, parquet) - Flexible topic selection - Lightweight and efficient implementation ### Tech stack: Python ### Why it matters: Auto Data simplifies the process of fine-tuning models by providing a solution to the common issue of insufficient data, ensuring more accurate and robust models.

Overview: UniVST is a unified framework for training-free localized video style transfer based on diffusion models, designed for efficient video stylization tasks. Key features: - Applies DDIM inversion to the original video and style image to obtain their initial noise. - Integrates Point-Matching Mask Propagation to generate masks for object regions. - Performs AdaIN-Guided Localized Video Stylization with a three-branch architecture. - Incorporates Sliding-Window Consistent Smoothing into the denoising process. - Supports five backbones: SD-v1.5, SD-v2.1, Animatediff-v2, SD-v3.0, and SD-v3.5. Tech stack: Python Why it matters: UniVST simplifies the process of video style transfer by eliminating the need for training, making it more accessible and efficient for various applications. Repository: QuanjianSong/UniVST Primary language: Python GitHub stars: 111 GitHub description: [TPAMI 2025] Official Pytorch Code of the Paper "UniVST: A Unified Framework for Training-free Localized Video Style Transfer"

### Overview: SoraFlows is an open-source, cross-platform web application for AI-powered video creation and editing using the latest OpenAI Sora model. It allows users to generate, edit, and share videos from text or images, with integrated support for voice-to-text, text-to-speech, voice cloning, and multi-language interfaces. ### Key Features: - **Video Generation**: Easily create videos from text or images. - **Voice-to-Text**: Convert speech to text. - **Text-to-Speech**: Convert text to speech. - **Voice Cloning**: Enhance voice quality. - **Multi-Language Support**: Support multiple languages. - **Integrated Tools**: Prompt Editor, Voice to Text, Text to Speech, Voice Cloning. - **Cross-Platform**: Works on any browser. - **OpenAI Sora Model**: Uses the latest OpenAI Sora model for advanced video creation. - **Stable and Interactive Interface**: Designed for a smooth and intuitive user experience. ### Tech Stack: TypeScript ### Why it Matters: SoraFlows is a powerful tool for anyone interested in AI-driven video creation and editing. Its ability to generate high-quality videos from text or images, combined with the integration of advanced AI features like voice-to-text and voice cloning, makes it a valuable asset for creators and enthusiasts alike. The platform's stable and interactive interface ensures a seamless user experience, making it easy to use even for those new to AI-driven video creation. ### Repository: TornadoInsight/Sora-FullStack ### Primary Language: TypeScript ### GitHub Stars:

### Overview: Vector Inference simplifies inference on Slurm clusters, making it accessible and efficient for researchers and developers. ### Key features: - **Easy Deployment**: Runs inference servers on Slurm clusters using open-source engines. - **Native Support**: Works seamlessly on Vector Institute cluster environments. - **Customization Options**: Supports customization through environment configuration files. - **Documentation**: Comprehensive documentation for both installation and usage. ### Tech stack: Python ### Why it matters: Vector Inference streamlines the process of running inference servers on Slurm clusters, enabling researchers and developers to focus on their research without worrying about the underlying infrastructure.

### Overview: This repository is a curated survey-style collection of papers, codebases, and benchmarks on weight space learning, focusing on the study and operation of neural network parameters. It aims to provide a systematic foundation for reasoning about, operating on, and reusing neural network parameters. ### Key features: - Papers, codebases, and benchmarks on weight space learning. - Papers, codebases, and benchmarks on weight space understanding. - Papers, codebases, and benchmarks on weight space representation. - Papers, codebases, and benchmarks on weight space generation. - Papers, codebases, and benchmarks on applications to related domains. ### Tech stack: Python, Markdown, Git. ### Why it matters: Weight space learning is a research perspective that shifts focus from studying neural networks only through their input–output functions to directly analyzing and leveraging their parameters. It aims to build a systematic foundation for reasoning about, operating on, and reusing neural network parameters, making it distinct from function-space or purely optimization-centric views. Repository: Zehong-Wang/Awesome-Weight-Space-Learning GitHub stars: 99

### Overview: This project benchmarks and evaluates generative text-to-image models, providing a unified way to measure and compare their quality. It includes standardized FID calculations, precomputed metrics for the MS-COCO validation set, and implementations of popular models to ensure reproducible results. ### Key features: - Unified metrics and datasets for all text-to-image models. - Reproducible results. - User-friendly interface for popular metrics: FID and CLIP-score. ### Tech stack: Jupyter Notebook ### Why it matters: This benchmark provides a standardized way to evaluate generative models, ensuring fair and comparable results across different models and datasets. It helps researchers and practitioners to objectively compare the quality of text-to-image models and contribute to the field of generative AI. ### Repository: boomb0om/text2image-benchmark ### GitHub stars: 110 ### GitHub description: Benchmark for generative image models