Projects

### Overview: Mojito is an open-source project designed for generating videos by controlling motion trajectories and intensity. It leverages advanced motion trajectory and intensity control techniques to create realistic video sequences. ### Key features: - **Motion Trajectory Control**: Mojito allows precise control over the motion of objects in videos, ensuring smooth and realistic movement. - **Intensity Control**: It enables the adjustment of video intensity, enhancing the realism and impact of the generated videos. - **Gradio Demo**: A user-friendly interface for testing and experimenting with Mojito's functionalities. ### Tech stack: - Python ### Why it matters: Mojito provides a powerful tool for researchers and developers interested in video generation and motion control, offering a flexible and efficient solution for creating realistic video sequences.

### Overview: TexFusion AI is an AI art generation tool created with the Flutter framework. It can convert text prompts into stunning images, including variations and 3D renderings. The app supports multiple themes, languages, and modes, making it versatile and user-friendly. ### Key Features: - **Generate Images from Text**: Converts text prompts into high-quality images. - **High Quality Images**: Produces HD images with minimal prompt descriptions. - **Unlimited Response**: Supports generating multiple variations of the same image. - **12 Styles Including 3D Render**: Offers 12 different styles, including 3D renderings. - **Download Image**: Allows users to download the generated image. - **Resolution Ratio 16:9**: Ensures the image is suitable for widescreen displays. - **Custom Theme and Modes**: Supports multiple themes and modes for personalization. - **Multi Languages**: Supports multiple languages for international users. - **Lightweight & User-friendly**: Easy to use with a simple interface. ### Tech Stack: Flutter Framework ### Why It Matters: TexFusion AI is a powerful tool for anyone looking to create unique and visually stunning images from text prompts. Its ability to generate high-quality 3D renderings and support for multiple styles makes it a versatile solution for artists, designers, and anyone interested in AI-generated art. ### Repository: Repository: VikramadityaDev/text_to_image_gen GitHub Stars: 62 GitHub Description: The TexFusion AI app was created by the Flutter framework and is a new AI art generation

Overview: pq-vector is a vector search implementation that utilizes only Parquet and DataFusion, providing an embedded index, standard compatibility, and zero-copy indexing. Key features: - Embedded Index: Index stored within the Parquet file itself, no separate index files. - Standard Compatible: Indexed files remain valid Parquet, allowing them to be read by DuckDB, Pandas, etc. - DataFusion Integration: Ergonomic vector search with just SQL. - Zero-copy: In-place Parquet indexing. Tech stack: Rust Why it matters: pq-vector simplifies vector search by eliminating the need for separate index files, ensuring compatibility with various data processing tools, and providing efficient in-place indexing for better performance and reduced overhead.

Overview: This is the official Nano Banana API (`gemini-2.5-flash-image-preview-official`) on the APIMart platform, providing text and image generation services at a token-billed rate. Key features: - Supports text and image generation - Token-billed pricing model - OpenAI-compatible endpoint - Supports multiple resolutions and image inputs Tech stack: Python Why it matters: This API is useful for developers looking to integrate OpenAI-compatible image generation into their applications, offering a flexible and cost-effective solution.

Overview: This project provides an official API for Nano Banana, offering model id, pricing, parameters, and example usage for the OpenAI-compatible APIMart endpoint. Key features: - Model ID: Nano Banana (Gemini 2.5 Flash Image) - Pricing: Token-billed, with $0.24/M for text_input and image_input - Parameters: Model, prompt, size, resolution, n, image_urls, nsfw_check - Example Usage: A POST request to the APIMart endpoint with specific parameters Tech stack: Python Why it matters: This API simplifies access to Nano Banana's image generation capabilities, making it easier for developers to integrate and utilize the model's features without the complexities of setting up a subscription or monthly plan. GitHub stars: 62 Repository: apimartyr1ot/gemini-2.5-flash-image-preview-official-api-cn

Overview: SMART is a PyTorch implementation of a regularization technique designed to fine-tune pretrained language models, focusing on robust and efficient fine-tuning for pre-trained natural language models. Key features: - SMART Loss Function - Robust Fine-Tuning - Efficient Fine-Tuning Tech stack: PyTorch Why it matters: This method helps in fine-tuning pre-trained models more effectively, ensuring robustness and efficiency, making it valuable for applications requiring high accuracy and stability.

### Overview: ComfyUI is a powerful and modular Stable Diffusion GUI, API, and backend with a graph/nodes interface. This Docker image simplifies the deployment of ComfyUI, making it accessible and easy to use. ### Key Features: - ComfyUI v0.43.0 - Jupyter Lab - Code-server - runpodctl - OhMyRunPod - RunPod File Uploader - Croc - Rclone - Application Manager 2.0.2 - CivitAI Downloader ### Tech Stack: Docker ### Why It Matters: This image streamlines the setup and deployment of ComfyUI, making it more accessible and easier to manage. The inclusion of various tools ensures a comprehensive and efficient development environment.

Overview: Z-Image Turbo API is a fast text-to-image route on APIMart, priced per image, designed for developers and creators looking to integrate text-to-image functionality into their projects without the need for a subscription. Key features: - Supports 300+ models, including text-to-image, image-to-image, and more. - Charges per image, with a default price of $0.0375 per image. - Supports various resolution tiers and customizable prompts. - Provides detailed model parameters and pricing information. - Offers an easy-to-use API key integration and real-time pricing. Tech stack: Python Why it matters: Z-Image Turbo API simplifies the integration of text-to-image functionality into projects, offering a fast and cost-effective solution without the need for a subscription.

Overview: Voyant is a fact-verified travel assistant that uses a single meta-agent to plan, execute, store receipts, and verify travel requests. Key features: - One-agent orchestration - Receipts and verification for answers - Resilient I/O and explicit timeouts Tech stack: TypeScript Why it matters: Voyant simplifies travel planning by providing a clear, reliable, and verifiable travel workflow. Repository: chernistry/voyant Primary language: TypeScript GitHub stars: 63

### Overview: Mettis is a distributed vector database implemented in Go, designed for efficient full-text and semantic search. It supports fault tolerance using Raft for replication and provides a lightweight text embedding service for integration with Python applications. ### Key Features: - **Full-Text Search**: Uses proximity ranking for efficient text retrieval. - **Semantic Search**: Leverages Hierarchical Navigable Small Worlds (HNSW) for accurate semantic search. - **Basic Text Embedding Service**: Provides a Python HTTP API for embedding services. - **Reciprocal Rank Fusion**: Merges full-text and semantic search results for enhanced relevance. - **In-Memory Serving**: Fast in-memory access for quick results. - **Disk Persistence**: Supports disk persistence for reliable data storage. - **Fault-Tolerance**: Raft-based replication ensures data reliability and availability. ### Tech Stack: - **Go**: The primary programming language for the implementation. - **Raft**: Used for fault-tolerant replication. ### Why it Matters: Mettis is a lightweight, efficient vector database that can be easily integrated into existing applications, providing a robust solution for search and retrieval tasks. ### Repository: - **GitHub Repository**: farouqzaib/mettis - **Stars**: 62 - **Description**: A distributed vector database built from scratch, supporting full-text and semantic search with Raft for fault tolerance. ### Architecture: - **Architecture Diagram**: Assets/architecture.png ### Getting Started: - **Dependencies**: Install dependencies in `third_party` using `pip`. - **Start the

### Overview: This repository contains experiments designed to test and evaluate various speech recognition systems for the SEPIA Framework. The primary focus is on integrating these systems into the SEPIA Speech-to-Text (STT) Server. ### Key Features: - **ASR Engines:** - **Whisper org:** The original Whisper version by Open-AI, known for high accuracy but lacking streaming support. - **Whisper TFlite:** A Tensorflow Lite compatible Whisper port, optimized for performance. - **Whisper Cpp:** A small C++ port of Whisper, efficient and fast. - **Whisper CT2:** An efficient and fast CTranslate2 port of Whisper, suitable for Arm64 systems. - **Sherpa ncnn:** A next-gen Kaldi implementation for streaming ASR, fast and supports streaming audio. - **Nvidia NeMo:** A toolkit for various end-to-end ASR models and languages, with a focus on performance and simplicity. - **Vosk:** A fast, small, and accurate speech recognition engine, easy to customize, and works well with classic Kaldi models. - **Wake-Word Detection:** - **OpenWakeWord:** A robust, NN-based open-source wake-word detection framework. ### Tech Stack: - Python ### Why It Matters: This repository serves as a platform for developers to experiment with different speech recognition systems, ensuring seamless integration into the SEPIA Framework. It provides a comprehensive set of tools and scripts to evaluate the performance and integration of various ASR and wake

### Overview: A Comprehensive RAG App on AWS This repository provides a complete Terraform infrastructure for deploying a Retrieval-Augmented Generation (RAG) application on AWS. It integrates with Google's free-tier Gemini Pro and uses a Remote MCP server for real-time web search. The project includes a Streamlit UI for interaction and a PostgreSQL RDS with the `pgvector` extension for efficient text embedding storage. The backend uses AWS Lambda for serverless compute, and the frontend is built with Streamlit for a user-friendly interface. The project is designed with CI/CD workflows, multi-environment support, and comprehensive testing, making it suitable for both development and production environments. ### Key Features: - **IaC with Terraform**: Consistent and repeatable deployments across environments. - **Serverless Compute**: AWS Lambda for backend logic. - **Vector Storage**: PostgreSQL RDS with `pgvector` extension. - **AI Integration**: Gemini Pro for generation and Gemini Embedding models for text embeddings. - **Authentication**: Secure user management with AWS Cognito. - **CI/CD Workflows**: GitHub Actions for automated deployment, testing, and cleanup. - **Multi-Environment Support**: `dev`, `staging`, and `production` environments. - **Comprehensive Testing**: Unit and integration tests for backend Lambda functions. - **Streamlit UI**: Includes a login page, document upload, query interface, and RAG evaluation dashboard. ### Tech Stack: - **Backend**: AWS Lambda, PostgreSQL RDS, Terraform, Streamlit - **Frontend**: Streamlit

Overview: Geniusrise is a framework designed to enable developers to run OpenAI models, including LLMs, STTs, and TTS, on their own infrastructure. It simplifies the process by providing a single binary and YAML configuration file, along with an OpenAI-compatible endpoint that autoscales and scales to zero. Key features: - A single binary and YAML file for easy setup. - Tested, optimized runtimes for various models (vLLM, llama.cpp, Speaches). - Autoscaling with a hard monthly budget cap. - API keys with per-key rate and usage limits. - Nodes with zero inbound ports, allowing seamless integration behind NAT and firewalls. Tech stack: Go. Why it matters: Geniusrise streamlines the process of deploying and running AI models, making it accessible and cost-effective for developers without requiring complex cloud infrastructure setup or expensive hosting services. Repository: geniusrise/geniusrise GitHub stars: 62 GitHub description: Geniusrise: Framework for building geniuses

Overview: This is a 7-day AI/ML fundamentals workshop designed to help beginners get started with AI and machine learning. It offers a supportive community and a strong foundation for understanding advanced concepts. Key features: - Comprehensive course content covering AI and machine learning fundamentals - Interactive sessions and practical exercises - Access to a community of learners with similar goals - Strong grasp of AI and machine learning fundamentals Tech stack: HTML Why it matters: This workshop provides a solid foundation for anyone interested in AI and machine learning, making advanced concepts easier to understand and apply in real-world scenarios. Repository: hemansnation/AI-ML-14-Hours-FREE-Course Primary language: HTML GitHub stars: 62 GitHub description: This FREE workshop aims to help beginners get started with AI and machine learning, offering a supportive community and a strong foundation for understanding advanced concepts.

### Overview: This tutorial provides a comprehensive introduction to diffusion models, focusing on their theoretical foundation through stochastic differential equations (SDE). It covers the basics of SDEs, the concept of diffusion models, and how they are used to generate images and other data samples. ### Key Features: - **Introduction to SDEs**: Understanding the basics of SDEs and how they relate to diffusion models. - **Diffusion Models Overview**: A brief introduction to diffusion models, including their applications and importance. - **Stochastic Differential Equations (SDEs)**: Detailed explanation of SDEs and their role in diffusion models. - **Image Generation**: Practical implementation of diffusion models for generating images, including forward and reverse processes, training, and sampling. - **Evaluation**: Methods for evaluating the performance of diffusion models. ### Tech Stack: - **Jupyter Notebook**: The primary language used for this tutorial. ### Why It Matters: - Diffusion models are a powerful technique for generating realistic images and other data samples by simulating the process of diffusion through a series of steps. - Understanding the theory behind diffusion models is crucial for effectively implementing and applying these models in various applications, such as image generation, denoising, and more. ### Repository: - **Repository**: `hieuristic/Tutorial_4` - **GitHub Stars**: 62 - **GitHub Description**: A Tutorial for Diffusion Models ### Setup: - **Requirements**: Install the required packages using `pip install -r requirements.txt`. - **Code Structure**: The repository contains detailed code

### Overview: Diffusion models are a class of generative models that generate images by gradually inverting a diffusion process. This tutorial provides a comprehensive introduction to diffusion models, starting with the theory of stochastic differential equations (SDEs). It covers the implementation of diffusion models from scratch, including the forward and backward SDEs, VPSDE, VESDE, and the training and sampling processes. The tutorial is designed to be accessible to developers and researchers interested in deep learning and generative models. ### Key Features: - **Stochastic Differential Equations (SDEs):** The tutorial begins with a detailed explanation of SDEs, providing the theoretical foundation for diffusion models. - **Implementation of Diffusion Models:** The tutorial covers the implementation of diffusion models from scratch, including the forward and backward SDEs, VPSDE, VESDE, and the training and sampling processes. - **Training and Sampling:** The tutorial provides a step-by-step guide to training and sampling diffusion models, including the implementation of the loss function, training loop, and sampling loop. - **Additional Features:** The tutorial includes optional tasks such as implementing EMA training, ISM loss, ODE sampling, and Schrodinger Bridge, which can be added as desired. ### Tech Stack: - **Primary Language:** Jupyter Notebook - **GitHub Stars:** 62 ### Why It Matters: Understanding diffusion models is crucial for anyone working in the field of generative models, as it provides a deep insight into the theoretical foundations of these models. The tutorial is designed to be accessible to

### Overview: A local Qwen-Image-2.1 studio for Apple Silicon, featuring text-to-image and multi-image editing capabilities. The CLI and web UI are both included, with the CLI being the primary interface for command-line editing. ### Key features: - **CLI (Command Line Interface)**: Directly interacts with the Qwen-Image-2.1 pipeline, allowing for text-to-image and multi-image editing. - **Web UI (Helmstudio)**: A web-based interface for the studio, providing a more user-friendly experience with a browser-based rendering process. - **Prompt Enhancer**: Utilizes Qwen's 9B prompt enhancer to enhance the quality of generated images, improving the final output. - **Device**: Supports both CPU and GPU devices, allowing for efficient rendering on Apple Silicon. - **Flags and Environment Variables**: Provides flexibility through flags and environment variables for customization and optimization. ### Tech stack: Python ### Why it matters: This local studio streamlines the process of generating high-quality images and enhancing prompts, making it a valuable tool for developers and content creators looking to leverage Qwen's capabilities on Apple Silicon without the need for a cloud-based solution.

### Overview: EarthSynth is an innovative project that leverages diffusion models to generate informative Earth observation images. This tool aims to enhance the accuracy and detail of satellite imagery, making it easier to analyze and interpret. ### Key Features: - **Diffusion Model Generation**: EarthSynth utilizes advanced diffusion models to create high-quality, realistic Earth observation images. - **Incorporation of Satellite Data**: The project integrates satellite imagery data to provide a comprehensive view of the Earth's surface. - **User-Friendly Interface**: The tool is designed to be user-friendly, making it accessible to a wide range of users. ### Tech Stack: Python ### Why It Matters: EarthSynth offers a significant improvement in the quality and detail of Earth observation images, making it invaluable for researchers, environmentalists, and anyone interested in understanding the Earth's surface. The project's ability to generate realistic images from satellite data is particularly impactful, offering a more accurate and detailed analysis of the Earth's surface features. ### Repository: - **Repository**: jaychempan/EarthSynth - **Stars**: 62 - **Description**: Official Code for "EarthSynth: Generating Informative Earth Observation with Diffusion Models" ### Primary Language: Python ### GitHub Description: Official Code for "EarthSynth: Generating Informative Earth Observation with Diffusion Models" ### Dataset: - **ModelScope**: EarthSynth-website - **HuggingFace**: Dataset - **HuggingFace**: Model - **GitHub**: Dataset - **GitHub**: Model ### License: -

Overview: SEED is a diffusion-based method designed to enhance speaker embeddings robustly, improving recognition accuracy under adverse acoustic conditions. Key features: - Easily applied to various speaker representation models. - No need for speaker labels or modification to existing pipelines. Tech stack: Python, PyTorch, CUDA. Why it matters: SEED can significantly improve recognition accuracy in real-world applications by mitigating performance degradation due to environmental mismatch. Repository: kaistmm/seed-pytorch GitHub stars: 62

Overview: This project is a fastText model serving service designed for developers, offering both HTTP REST and gRPC API endpoints for serving fastText models. Key features: - HTTP REST API for serving fastText models - gRPC API for serving fastText models Tech stack: Rust Why it matters: This project provides a convenient way to serve fastText models, making it easier for developers to integrate fastText functionalities into their applications.