Projects

### Overview: TurboQuant is an open-source implementation of Google's TurboQuant, designed to optimize the KV cache compression for large language models (LLMs) on GPUs. It achieves 5x compression with near-zero quality loss, making it ideal for scenarios where GPU memory is the bottleneck. ### Key features: - **KV Cache Compression**: TurboQuant reduces the size of the KV cache by up to 4×, significantly lowering memory usage. - **Near-Zero Quality Loss**: The compression method preserves the quality of the model output, ensuring minimal impact on inference performance. - **Efficient Memory Utilization**: The approach is particularly effective in scenarios where GPU memory is the limiting factor. ### Tech stack: TurboQuant is written in Python, leveraging the PyTorch framework for efficient computation. ### Why it matters: TurboQuant is crucial for developers working on LLMs, especially those aiming to optimize their models for deployment on constrained hardware. By reducing the size of the KV cache, TurboQuant allows models to fit into smaller memory constraints, making it easier to deploy models on edge devices or on devices with limited GPU memory.

### Overview: Open Voice Input Linux 是一个原生支持的 Linux 原生自适应语音输入工具，专为 Ubuntu 24.04 x86_64 和 IBus 设计。它通过 IBus 的预编辑功能在光标内实时显示和提交文字，提供术语纠错和用户自定义训练数据，旨在简化语音到文字的转换过程。 ### Key features: - 实时显示和提交文字 - 术语纠错 - 用户自定义训练数据 - 无需粘贴，直接在应用内完成识别 - 本地安装，轻量级 `.deb` 包约 404 KiB - 保留个人 ASR 数据，可选配置 - 支持远程桌面，可手动确认识别结果 ### Tech stack: Python ### Why it matters: Open Voice Input Linux 提供了一种简单且高效的方式来将语音转换为文字，同时保持了简洁和易用性。它通过 IBus 的预编辑功能实现了实时显示和提交文字，避免了粘贴和复制粘贴的繁琐步骤。此外，它还支持术语纠错和用户自定义训练数据，使得输入更加准确和个性化。由于其本地安装和轻量级 `.deb` 包，它在性能和安装方面表现出色，特别适合在远程桌面环境中使用。

Overview: This project provides an official API for Nano Banana 2, offering model IDs, pricing, parameters, and example usage for the OpenAI-compatible APIMart endpoint. Key features: - Model ID: Nano Banana 2 (Gemini 3.1 Flash Image) - Pricing: Token-billed, $0.4/M for text and image inputs, $2.4/M for text outputs, $48/M for image outputs - Parameters: Model ID, Prompt, Size, Resolution, Number of Images, Image URLs, NSFW Check - Pricing Details: Effective price per million tokens, tiered pricing based on resolution Tech stack: Python Why it matters: This API simplifies access to Nano Banana 2 through a familiar OpenAI-compatible endpoint, allowing teams to switch models seamlessly without changing their client code. It offers a cost-effective, pay-as-you-go model with real-time billing and support for asynchronous tasks, making it ideal for developers and API integrations. GitHub stars: 84

Overview: Batteries Included is a Kubernetes-based software platform designed for modern application development, offering production-ready capabilities, ease of use, and enterprise-grade features. Key features: - Automated deployment and scaling of databases, monitoring, and web services - Built-in security features including SSL/TLS certificate management and OAuth/SSO integration - Jupyter notebooks with pre-configured environments for AI/ML development - Automated PostgreSQL, Redis, and MongoDB deployment and management - Grafana dashboards and VictoriaMetrics for monitoring and metrics collection Tech stack: Elixir Why it matters: Batteries Included simplifies cloud infrastructure management, allowing developers to focus on building applications without the complexities of managing infrastructure, ensuring production-ready services and enterprise-grade features. Repository: batteries-included/batteries-included GitHub stars: 85 GitHub description: Batteries Included is a Kubernetes based software platform for database, ai, web, monitoring, and more.

### Overview This project is an example of a Retrieval-Augmented Generation (RAG) application built with Haystack, designed to demonstrate the creation of a functional search and generative question-answering system with a user-friendly interface. ### Key Features - **Backend**: Built with FastAPI and Haystack 2, providing a RAG pipeline using OpenAIGenerator. - **Frontend**: Leveraging React and Bootstrap, offering an intuitive interface for interacting with the RAG system. - **Tech Stack**: Python, FastAPI, Haystack 2, OpenAIGenerator, OpenSearch, and Nginx proxy. - **Why it Matters**: This project serves as a practical example for developers looking to implement RAG systems, showcasing the integration of Haystack, OpenSearch, and OpenAI within a user-friendly frontend. ### Tech Stack: Python The project is developed using Python, leveraging FastAPI for the backend and Haystack 2 for the RAG pipeline. ### Why it Matters: Demonstrates RAG System Implementation This project provides a clear example of how to integrate Haystack, OpenSearch, and OpenAI within a user-friendly frontend, making it valuable for developers looking to implement similar systems.

Overview: This repository is a Unity plugin for Android Speech Recognition based on a Java implementation, providing basic functionality and additional features like setting language, maximum number of results, and prompt questions. Key features: - Set Recognition Language - Set the max number of results - Set the prompt question Tech stack: Java Why it matters: This plugin allows developers to integrate speech recognition functionality into their Unity projects, making it easier to build voice-based applications without needing to write custom code.

Overview: PyOMlx is a local server designed to serve MLX models, inspired by Ollama, for Apple Silicon. Key features: - Revamped http server to use `mlx_lm.server` module, supporting dynamic model information and automatic download from HuggingFace Hub. - Added OpenAI API-compatible chat completions and list models endpoint. - Added `/download` endpoint for direct model downloads from HuggingFace Hub. - `/swagger.json` endpoint to serve OpenAPI Spec. - Automatically discover and serve MLX models from MLX Huggingface community. - Easy startup/shutdown via MacOS App. - System tray indication. Tech stack: Python, Flask, MLX models, HuggingFace Hub. Why it matters: PyOMlx simplifies interacting with MLX models, making it easier to integrate Apple Silicon MLX models into workflows, potentially benefiting developers and researchers.

Overview: CLIP is a model that connects text and images, pre-trained from natural language supervision, achieving state-of-the-art performance across various datasets. Key features: - Pre-trained on natural language supervision - Achieves high accuracy on image classification tasks - Supports multiple datasets including CIFAR100 and CIFAR10 Tech stack: Python Why it matters: CLIP is a groundbreaking model that opens up new possibilities for image understanding and text-to-image generation, making it a valuable asset in the field of computer vision and natural language processing. Repository: leaderj1001/CLIP Primary language: Python GitHub stars: 85 GitHub description: CLIP: Connecting Text and Image (Learning Transferable Visual Models From Natural Language Supervision)

Overview: This project focuses on improving flow-based generative super-resolution models by introducing a learned prior, specifically for the "Boosting Flow-based Generative Super-Resolution Models via Learned Prior" paper presented at CVPR 2024. The repository includes the training and evaluation codes for the proposed method. Key features: - Introduces a learned prior to flow-based SR models - Addresses grid artifacts and exploding inverses - Enhances the quality of synthesized images without modifying the original models - Effective, flexible, and generalizable to both fixed-scale and arbitrary-scale SR frameworks Tech stack: Python Why it matters: This framework addresses critical issues in flow-based SR methods and improves the quality of synthesized images, making it valuable for applications requiring high-resolution super-resolution. Repository: liyuantsao/BFSR GitHub stars: 85 GitHub description: The official repository of "Boosting Flow-based Generative Super-Resolution Models via Learned Prior" [CVPR 2024]

Overview: ManySpeech is a comprehensive voice processing suite developed by the manyeyes community, built on C# and utilizing Microsoft.ML.OnnxRuntime for ONNX model decoding. It addresses compatibility, model adaptation, and integration challenges in cross-platform deployment, real-time/offline scenarios, and multi-language support. Key features: - Voice recognition - Endpoint detection - Punctuation restoration - Audio separation and enhancement Tech stack: C# Why it matters: ManySpeech simplifies the development process, improves deployment flexibility, and provides strong support for voice processing in the .NET ecosystem, enhancing development efficiency and versatility. Repository: manyeyes/ManySpeech GitHub stars: 85

### Overview: MOTFM is a medical image generation framework that accelerates the process while maintaining high quality, suitable for both 2D/3D and class/mask-conditional setups. ### Key Features: - **Flow Matching for Medical Image Synthesis**: MOTFM accelerates medical image generation while preserving and often improving quality. - **2D/3D Support**: The framework supports 2D and 3D medical image synthesis. - **Class/Mask Conditional**: It can generate images conditioned on class labels or masks. ### Tech Stack: Python ### Why it Matters: MOTFM is designed to bridge the gap between speed and quality in medical image synthesis, making it a valuable tool for researchers and practitioners in the field. ### Repository: milad1378yz/MOTFM ### Primary Language: Python ### GitHub Stars: 85 ### GitHub Description: Flow Matching for Medical Image Synthesis: Bridging the Gap Between Speed and Quality ### Data Preparation: Your training data must be stored in a single `.pkl` file, following a specific structure. Ensure your dataset adheres to the described data structure before running the training or inference pipelines. ### Configuration Files: You must either create or modify a YAML configuration file to suit your dataset paths, model parameters, and hyperparameters. Sample configuration files are provided in the `configs/` folder. ### Training: To train the model, run: ```bash python trainer.py --config_path configs/default.yaml ``` or (after installation): ```bash motfm-train --config_path configs

### Overview: This project presents an unsupervised approach to discover generative concepts from a collection of images, enabling the disentanglement of different art styles, objects, and lighting in paintings, objects, and kitchen scenes. The discovered concepts accurately represent the content of images and can be recombined and composed to generate new artistic and hybrid images. These concepts are further utilized as a representation for downstream classification tasks. ### Key features: - Discover generative concepts from images. - Disentangle different art styles, objects, and lighting. - Represent image content accurately. - Generate new artistic and hybrid images. - Use as a representation for downstream classification tasks. ### Tech stack: Python ### Why it matters: This project addresses the inverse problem of discovering generative concepts from images, enabling the disentanglement of complex visual information and providing a versatile representation for various downstream tasks.

Overview: This TypeScript library simplifies the use of various LLM provider APIs, offering a unified interface for real-time responses, function calling, and structured output. Key features: - Universal API - Streaming responses - Function invocation - Vision capabilities - Type-safe JSON responses - Multi-turn conversations - Graceful operation cancellation - Usage tracking Tech stack: TypeScript Why it matters: This library streamlines the integration of different LLM providers, making it easier for developers to build applications that leverage advanced AI capabilities without the need for extensive configuration. Repository: nbonamy/multi-llm-ts Primary language: TypeScript GitHub stars: 85 GitHub description: A TypeScript library to use LLM provider APIs in a unified way.

Overview: tgdb 是一个基于 C++23 的第三方 Telegram 群组消息搜索机器人，支持 OCR 和语义搜索，提供快速便捷的搜索体验。 Key features: - 支持 OCR 和语义搜索 - 使用 Inline Query 进行搜索 - 基于 tdlib 开发，采用 cpp20 coroutine 包装 Tech stack: C++23 Why it matters: 提供了快速、强大的 Telegram 群组消息搜索功能，解决了 Telegram 搜索体验差的问题，并且支持基于阿里 GME 的图文语义搜索，功能丰富且易于扩展。

### Overview: Dub Anything is a cutting-edge tool that allows you to dub videos into multiple languages using the original actor's voices. It preserves the original music and sound effects while automatically translating and dubbing the video line by line, making the process seamless and natural. ### Key features: - **Voice Matching and Translation:** The AI agent matches the voice of each actor from video frames and dialogue, translating lines accurately and acting them out as the original actor would. - **Line-by-Line Dubbing:** The process is line-by-line, ensuring that each line is acted out as the original actor would, maintaining the authenticity of the dubbing. - **Custom Accents:** You can specify a different accent for each actor, ensuring a natural dubbing experience. - **Transcripts and Subtitles:** The AI provides a transcript and subtitles in both languages, with translations timed for reading, and subtitles that appear on screen as they do in the original video. ### Tech stack: The tool is built using Python and relies on AI coding agents such as Claude Pro or Max, ChatGPT Plus or Pro, and OpenCode with an API key. ### Why it matters: Dub Anything simplifies the dubbing process by automating the translation, voice matching, and line-by-line dubbing, making it accessible to anyone with a computer and a basic understanding of AI. It ensures a natural and authentic dubbing experience, preserving the original music and sound effects while maintaining the integrity of the original actor's voices.

### Overview: Scene Card Studio is an AI-driven tool that transforms personal photos into structured, editable visual narratives. ### Key Features: - **Photo to Narrative:** Converts single photos or photo sequences into structured narrative prompts. - **Bilingual Briefs:** Supports both English and Chinese briefs for better communication. - **Expression Profiles:** Offers 23 customizable Expression Profiles for diverse photo styles. - **Prompt Compilation:** Generates versioned prompts for multiple photos or sequences. - **Aesthetic Review:** Ensures the aesthetic quality of the generated images. - **Retry and Accept:** Allows users to retry or accept the generated images. ### Tech Stack: Python ### Why It Matters: Scene Card Studio simplifies the process of turning personal photos into structured narratives, making it easier for anyone to create compelling visual stories.

### Overview: 3D Paintbrush is a technique for automatically texturing local semantic regions on meshes using text descriptions. The method operates directly on meshes, producing texture maps that seamlessly integrate into standard graphics pipelines. It leverages multiple stages of a cascaded diffusion model to supervise its local editing technique with generative priors learned from images at different resolutions, improving both the quality of localization and stylization. ### Key features: - **3D Texturing**: Automatically textures local semantic regions on meshes. - **Local Editing**: Operates directly on meshes, producing texture maps that integrate seamlessly into standard graphics pipelines. - **Cascaded Score Distillation (CSD)**: Supervises the local editing technique with generative priors from multiple stages of a cascaded diffusion model, enabling control over both the granularity and global understanding of the supervision. ### Tech stack: Jupyter Notebook ### Why it matters: 3D Paintbrush offers a new approach to texturing local semantic regions on meshes, improving the quality of both localization and stylization. Its ability to leverage multiple stages of a cascaded diffusion model to supervise its technique with generative priors from different resolutions makes it a significant advancement in the field of 3D texturing.

Overview: This project integrates PyTorch with JetStream, enabling efficient inference of large language models (LLMs) using the JetStream engine. Key features: - PyTorch/XLA integration for LLM inference - Support for various LLM models, including Llama-2, Meta-Llama, and Mixtral - Option to run models in 8-bit quantization for improved performance - Ability to download and use pre-trained weights from HuggingFace or custom checkpoints Tech stack: Python Why it matters: This project provides a practical solution for running large language models efficiently on JetStream, leveraging the strengths of both PyTorch and JetStream's optimized inference capabilities. GitHub stars: 84

Overview: DuckBuddy 是一个开源的具身智能机器人，能够行走、听觉、视觉，并具有自主行为和生命层，所有功能都由机器人自己实现。 Key features: - 能够行走、听觉、视觉和自主行为。 - 采用强化学习策略训练，具备摔倒保护功能。 - 配备了离线语音识别和双麦克风声音定位功能。 - 可以执行语音指令并按照指令完成动作。 - 有自主行为和生命层，包括需求和情绪管理。 Tech stack: Rust, Orange Pi Zero 3W, Feetech HL-2915 servos, Pollen Robotics' Microduck (unofficial). Why it matters: DuckBuddy 是一个开源的具身智能机器人，能够行走、听觉、视觉并具有自主行为和生命层，所有功能都由机器人自己实现。这使得它成为研究和开发具身智能机器人系统的一个重要工具，同时也为用户提供了可复刻和改进的开源系统。

### Overview: This project is a Pytorch implementation of the paper "Key-Locked Rank One Editing for Text-to-Image Personalization", focusing on personalizing text-to-image models for better image quality and diversity. ### Key Features: - **Multiple Concepts Inference**: Supports multiple concepts for personalized image generation. - **Support SDXL-1.0**: Enables training and generation using Stable Diffusion XL v1.0. - **CLIP Metrics**: Evaluates the quality and diversity of generated images using CLIP metrics. - **Advanced Sampler**: Supports advanced samplers from Stable Diffusion, such as EulerEDMSampler. ### Tech Stack: Python ### Why it Matters: This project offers a powerful tool for personalizing text-to-image models, enhancing the quality and diversity of generated images, and improving the overall user experience by leveraging advanced techniques and metrics. ### Repository: ChenDarYen/Key-Locked-Rank-One-Editing-for-Text-to-Image-Personalization ### GitHub Stars: 84 ### Environment: Create and activate the conda environment: ```bash conda env create -f environment.yaml conda activate perfusion ``` ### Training: Download the Stable Diffusion models (`v1.5` or `v2.1`) and place them in `./ckpt/`. Then run: ```bash python main.py \ --name teddy \ --base ./configs/perfusion_teddy.yaml \ --basedir ./ckpt \ -t True \ --gpus