Projects

Overview: Jecq is an open-source C++ library designed for efficient similarity search of dense vectors, offering advanced dimension-aware compression techniques that reduce memory footprint while maintaining high search accuracy. Key features: - Dimension-Aware Compression - Faiss Compatibility - Hyper-Parameter Optimization - CPU Implementation Tech stack: C++ Why it matters: Jecq provides a significant reduction in storage while maintaining high search accuracy, making it an ideal solution for applications requiring fast and efficient vector similarity search. Repository: JaneaSystems/jecq GitHub stars: 58 GitHub description: Faiss-based library for efficient similarity search

### Overview: GigaGAN is an implementation of [GigaGAN: Scaling up GANs for Text-to-Image Synthesis](https://arxiv.org/pdf/2303.05511), a method designed to enhance the scalability of Generative Adversarial Networks (GANs) for generating high-quality images from text inputs. ### Key Features: - **GigaGAN Generator**: A generator model that scales up the architecture of GANs, enabling it to process larger datasets and produce more detailed images. - **GigaGAN Discriminator**: A discriminator model that is also scaled up, improving the quality of generated images by better distinguishing real from fake samples. - **Adaptive Kernel**: The generator and discriminator use adaptive kernels, which dynamically adjust their size based on the input size, enhancing the model's efficiency and performance. ### Tech Stack: - Python ### Why it Matters: GigaGAN is significant because it addresses the scalability challenges of GANs, particularly in the context of text-to-image synthesis. By scaling up the architecture, it can handle larger datasets and produce higher-quality images, making it a valuable tool for applications in image generation and text-to-image synthesis.

Overview: StageMate is a smart presentation assistant designed to enhance your pitch by covering all aspects of your presentation, from skipping slides to reminding you of major points. Key features: - Tracks your time - Reminds you of missed major points - Assists with audio transcription Tech stack: HTML Why it matters: StageMate aims to automate the process of presenting, making it more efficient and less reliant on human intervention, thereby freeing up time for more creative aspects of the presentation. Repository: Langhalsdino/StageMate GitHub stars: 58 GitHub description: StageMate is the smart assistant for your presentation. It will cover all aspects of your pitch from skipping slides to reminding you if you miss some major point.

Overview: This repository provides five production-grade pipelines for modern multimodal vision systems, focusing on efficient VLM adaptation, embedding-space drift detection, edge inference, robustness, and video/temporal reasoning. Key features: - Efficient VLM adaptation - Embedding-space drift detection - Edge inference - Robustness & safety - Video/temporal reasoning Tech stack: Python Why it matters: This repository offers five runnable pipelines that implement current (June 2026) frontier techniques with production constraints built in from the start. It includes CPU-only fallback paths, honest benchmark tables, before/after metrics with working code, and inline citations to specific 2025–2026 papers behind each technique. GitHub stars: 58

### Overview: Xamarin.Cognitive.Speech is a client library that simplifies working with the Microsoft Cognitive Services Speech Services Speech to Text API on Xamarin.iOS, Xamarin.Android, UWP, and Xamarin.Forms/.NET Standard libraries. It includes a sample app for Xamarin.Forms, supporting interactive, conversation, and dictation modes, profanity modes, and efficient audio streaming. ### Key Features: - **100% `HttpClient`-based, managed code (.NET Standard 2.0)** - **Works with the latest version of the Speech Services Speech to Text API** - **Advanced features**: Chunked transfer encoding, multiple recognition modes (Interactive, Conversation, Dictation), simple vs. detailed output mode, profanity modes (Masked, Removed, Raw), recognition language support, and multiple authentication modes. - **Seamless integration with AudioRecorder Plugin** for seamless audio streaming. - **Easy to use API** ### Tech Stack: C# ### Why It Matters: - **Updated and Mobile-Specific**: The official C# SDK is designed for Windows, not Xamarin, making it easier to use with mobile platforms. - **Feature Support**: Includes all advanced features of the Speech Services Speech to Text API. - **Cross-Platform**: Works with various platforms, including Xamarin.Forms, UWP, and Android. - **No Overhead**: Minimal code and no need for older APIs or complex configurations. ### GitHub Stars: 58 ### Setup: - Add the NuGet package to your Xamarin.iOS, Xamarin.Android, UWP, or other .NET Standard

Overview: DriveSceneGen is a novel method for generating diverse and realistic driving scenarios from scratch, addressing the need for large and diverse data in autonomous driving research. Key features: - Real-world driving dataset learning - Generation of novel driving scenarios with high fidelity and diversity - Two-stage method outperforming existing methods on specific tasks Tech stack: Python Why it matters: DriveSceneGen provides a practical solution for generating realistic driving scenarios, which is essential for the development and validation of autonomous driving systems.

Overview: Wan 2.7 Video Edit API (`wan2.7-videoedit`) is an OpenAI-compatible API endpoint that provides high-quality video editing services for a low cost, making it ideal for developers looking to integrate AI into their projects without the need for a monthly subscription. Key features: - Supports multiple video editing tasks - Offers professional-grade performance - Compatible with various resolutions and aspect ratios - Supports image references for target styles - Charges based on the cost of the generated video, with a minimum of $1 per call - Asynchronous processing with task tracking Tech stack: Python Why it matters: This API is essential for developers who need to integrate AI video editing into their applications without the financial burden of a monthly subscription. It provides a cost-effective solution for video editing tasks, making it a valuable tool for developers looking to enhance their applications with AI capabilities. Repository: a1bvc/wan-2-7-videoedit-api-es Primary language: Python GitHub stars: 49 GitHub description: Wan 2.7 Video Edit API (`wan2.7-videoedit`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

### Overview: 本地部署大语言模型的中文教学 ### Key features: - **本地部署**：无需联网，无需按月付费，适合隐私保护和成本控制。 - **快速入门**：十分钟内跑起第一个模型，简单易懂。 - **详细解释**：每个新词解释清楚，硬核技术细节标注。 ### Tech stack: Python, PyTorch, Hugging Face Transformers, Docker, Kubernetes ### Why it matters: - **灵活部署**：适合个人、企业、团队等不同规模的使用场景。 - **成本效益**：节省了网络费用和按月付费的开销。 - **隐私保护**：避免了模型暴露在公共网络中，保护用户隐私。 ### Repository: anseryuer/Local_LLM_Deployment_Guide_Chinese ### GitHub stars: 57 ### Primary language: Unknown ### GitHub description: 本地部署大语言模型的中文教学

Overview: This project provides an official API for Nano Banana 2, offering pricing, parameters, and a copy-paste example for the OpenAI-compatible APIMart endpoint. Key features: - Pricing information - Model id and routes - Pricing model - Request parameters Tech stack: Python Why it matters: This API allows developers to easily integrate Nano Banana 2 into their projects using the OpenAI-compatible APIMart, providing a cost-effective and scalable solution for image generation tasks.

Overview: Convert text into beautiful artistic images using a generative adversarial network. Key features: - Text-to-image conversion - High-quality artistic output - Multiple stages for improved quality Tech stack: Python Why it matters: This project offers a unique way to transform text into visually appealing art, potentially opening up new avenues in creative design and art generation. Repository: azmiozgen/text2painting Primary language: Python GitHub stars: 58 GitHub description: Convert text into beautiful artistic images README excerpt:

### Overview: Paris is the world's first decentralized trained open-weight diffusion model, designed to achieve superior parallelism efficiency using 14× less data and 16× less compute compared to traditional methods. ### Key Features: - **8 Independently Trained Expert Diffusion Models**: Each model has 605M parameters, totaling 4.84B parameters. - **No Gradient, Parameter, or Activation Synchronization**: Training is conducted in isolation with no need for gradient, parameter, or activation exchange. - **Lightweight Transformer Router**: A router with ~129M parameters for dynamic expert selection. - **11M LAION-Aesthetic Images**: Across 120 A40 GPU-days, ensuring diverse and high-quality training data. - **14× Less Training Data**: Paris uses significantly less training data than prior decentralized baselines. - **16× Less Compute**: Paris requires 16× less compute than prior decentralized baselines. - **Competitive Generation Quality**: Achieves a FID score of 12.45 on DiTExpert XL/2, demonstrating high-quality generation. ### Tech Stack: The project is implemented using TensorFlow and Hugging Face Transformers. ### Why It Matters: Paris represents a significant advancement in decentralized training, offering superior efficiency and reduced resource requirements compared to traditional parallelization strategies. The model's ability to achieve competitive generation quality while using minimal data and compute is particularly noteworthy, making it a valuable tool for researchers and developers looking to improve the efficiency and scalability of diffusion models.

Overview: This project is an official implementation of "Adversarial Supervision Makes Layout-to-Image Diffusion Models Thrive" (ICLR 2024), focusing on the development of a layout-to-image diffusion model. Key features: - Adversarial Supervision - Layout-to-Image Diffusion Models - Segmentation Network UperNet Tech stack: Jupyter Notebook Why it matters: This project provides a comprehensive solution for generating high-quality images from layouts, showcasing the effectiveness of adversarial supervision in improving diffusion models. Repository: boschresearch/ALDM GitHub stars: 58 GitHub description: Official implementation of "Adversarial Supervision Makes Layout-to-Image Diffusion Models Thrive" (ICLR 2024)

### Overview: Snippets is an intelligent code repository system that automatically extracts, processes, and indexes code snippets from GitHub repositories into a searchable vector database. It leverages AI-powered code analysis and semantic search capabilities to help developers quickly find relevant code examples and patterns. ### Key Features: - **🚀 Use Your Claude Subscription**: Access this tool with your Claude subscription plan. - **🤖 Automated Repository Processing**: Extract meaningful code snippets from any GitHub repository. - **🔍 Semantic Search**: Find code by meaning, not just keywords, using vector embeddings. - **🔧 MCP Integration**: Seamless integration with Claude Code through Model Context Protocol. - **⚡ Background Processing**: Efficient queue-based processing for large repositories. - **🐳 Docker Ready**: Complete containerized setup for easy deployment. ### Tech Stack: Python ### Why It Matters: Snippets provides a powerful tool for developers looking to quickly find and use code examples, enhancing their development process by providing context-aware insights directly within their repositories.

Overview: This is the official .NET SDK for Deepgram, providing comprehensive tools and features for developers to integrate speech and language AI into their applications. Key features: - Pre-Recorded (Synchronous) and (Asynchronous / Callbacks) pre-recorded audio processing - Streaming audio processing - Conversational speech recognition (Flux) - Text to speech (Flux) - Voice agent integration - Text to speech REST and streaming APIs - Text intelligence capabilities - Authentication and key management - Project management and key management - Member and scope management - Invitations and project management - Usage tracking and billing - Self-hosted distribution credentials - Logging and backcompatibility - Development and contributing resources Tech stack: C# Why it matters: This SDK simplifies the integration of Deepgram's powerful speech and language AI models into .NET applications, making it easier for developers to build applications that can understand and respond to human speech and language, enhancing user experience and improving communication.

### Overview: FunASR.NET 是一个基于 .NET 10 的纯 C# 项目，旨在实现普通话 ASR、TTS 和电话音频处理。它使用固定 F16 编码器和 Q8_0 解码器，支持 CPU SIMD 和 Vulkan GPU 混合后端。项目包含详细的文档、示例代码和测试数据，旨在提供一个独立的社区实现，与官方 SDK 和工具箱不兼容。 ### Key features: - **ASR (Automatic Speech Recognition)**: 支持固定 F16 编码器和 Q8_0 解码器，实现普通话转写。 - **实时转写 (Real-time Transcription)**: 支持 WebSocket、VAD 切句和可修订预览功能。 - **中文 TTS (Text-to-Speech)**: 提供 Huayan 1 位和 AISHHELL3 174 位发音人，共 175 位。 - **浏览器 Demo**: 实现麦克风实时转写、上传 WAV、TTS 播放/下载功能。 - **电话音频 (Phone Audio)**: 支持 PCM16、G.711 PCMA/PCMU、重采样处理，可连接自研 SIP/PJSIP 媒体层。 - **并发 (Concurrency)**: 独立会话状态、只读权重共享、队列/配额/取消和缓冲所有权管理。 - **可选 GPU**: Vulkan FP32

### Overview: VisionTrim is an official repository for the paper "VisionTrim: Unified Vision Token Compression for Training-Free MLLM Acceleration" presented at ICLR 2026. This project aims to address the high computational costs of large language models due to excessive visual tokens by proposing a unified framework that integrates two effective modules: Dominant Vision Token Selection (DVTS) and Text-Guided Vision Complement (TGVC). ### Key Features: 1. **Dominant Vision Token Selection (DVTS)**: This module preserves essential visual tokens via a global-local view, ensuring that the model focuses on the most important visual information. 2. **Text-Guided Vision Complement (TGVC)**: This module facilitates context-aware token merging guided by textual cues, enhancing the model's ability to handle complex visual and textual inputs. 3. **Training-Free MLLM Acceleration**: The project provides a unified framework that accelerates the training of multimodal large language models without the need for extensive fine-tuning. ### Tech Stack: - Python ### Why It Matters: VisionTrim offers a significant improvement in the training-free acceleration of large language models by addressing the limitations of existing token reduction methods. By integrating DVTS and TGVC, the project aims to enhance the performance and efficiency of MLLMs in various applications, particularly in high-resolution and video-based scenarios. ### Repository: - **Repository**: hanxunyu/VisionTrim - **GitHub Stars**: 58 - **GitHub Description**: Official code repository for "⚡️VisionTrim: Unified Vision Token

### Overview: FlashVoice 是一款适用于 macOS 和 Windows 的本地优先语音输入工具，支持端侧 SenseVoice 识别、全局语音输入、文件转写及可选 AI 校对。它提供高效、私密的语音输入体验，无需上传任何录音数据。 ### Key features: - **本地识别**：基于 SenseVoice 在设备端实时将语音转换为文字。 - **两种输入方式**：支持按住说话和免按键连续语音输入。 - **文件转写**：将本地音频或视频文件转换为文字或字幕。 - **多种录音来源**：支持麦克风和系统音频。 - **AI 校对**：可选用云端大模型服务或本地模型优化识别结果。 - **实用工具**：支持自定义词典、转写历史和多语言输入。 ### Tech stack: Vue ### Why it matters: FlashVoice 提供了高效、私密的语音输入体验，适用于各种应用场景，如文档编辑、聊天记录和网页浏览等。其本地识别功能确保了数据安全，无需上传任何录音数据，从而保护用户的隐私。

### Overview: This project introduces a framework called B^2-DiffuRL, which addresses the sparse reward problem in the training of diffusion models. The framework, B^2-DiffuRL, combines backward progressive training and branch-based sampling to improve the alignment between prompts and generated images, while maintaining diversity in the generated outputs. ### Key features: - **B^2-DiffuRL** is a reinforcement learning-based framework designed for fine-tuning diffusion models. - It employs backward progressive training to focus on final denoising steps, easing the learning process from sparse rewards. - Branch-based sampling is used to evaluate the contributions of policies to the final image, helping to learn effective rather than unnecessary policies. ### Tech stack: - Python ### Why it matters: B^2-DiffuRL is significant because it provides a solution to the sparse reward problem in diffusion model training, leading to better alignment between prompts and generated images while maintaining diversity in the outputs. ### Repository: - **Repository**: hu-zijing/B2-DiffuRL - **GitHub stars**: 58 - **Primary language**: Python - **GitHub description**: Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards ### Run: - **Command**: `bash run_process.sh > log/exp_B2DiffuRL_b5_p3` - **Description**: Starts fine-tuning and stores results under `model/`. The pipeline includes sampling, evaluation, and training. ### Acknowledgment: - This project was built with reference to several repositories, including:

### Overview: Gemini Web2API is a blazing fast Go port of the original Python gemini-web2api. It converts Google Gemini's web interface into an OpenAI-compatible API, offering zero cost and a single static binary for cross-platform use. ### Key features: - **Optional API Keys**: No auth when `api_keys` is empty, OpenAI-style Bearer auth when configured. - **OpenAI Compatible**: Drop-in replacement for `/v1/chat/completions`, `/v1/models`, and `/v1/responses` (Codex CLI). - **Multimodal (Vision)**: Fully supports sending images (`image_url` / base64) to Gemini using Scotty Resumable Upload protocol natively with built-in image compression. - **TLS Impersonation**: Built-in support to bypass Cloudflare/WAF by mimicking Chrome/Edge TLS fingerprints using `tls-client` 🔥. - **Tool Calling**: Full function calling support (OpenAI format). - **Multiple Models**: Flash (3.6), Extended Thinking (20k+ char output), Pro, Auto, Lite. - **Thinking Depth**: Adjustable reasoning via `@think=N` suffix (0=deepest, 4=shallowest). - **Web Search**: Built-in internet access (Gemini's native search). - **High Performance**: Pure Go static binary with high concurrent SSE streaming throughput and minimal memory footprint. - **Cross-Platform**: Zero runtime dependencies, single static binary. - **Codex CLI**: Responses API (`/v1/responses`)

### Overview: The MCP server allows AI assistants like Claude, GPT, Cursor, and Copilot to control Home Assistant through the Model Context Protocol, providing a seamless integration for 50+ tools across three transports. This project is designed to be a versatile solution for automating your smart home setup, making it easier to manage and control your devices. ### Key Features: - **Model Context Protocol Integration**: Enables direct communication between AI assistants and Home Assistant. - **50+ Tools**: Supports a wide range of Home Assistant integrations. - **Three Transports**: Compatible with MQTT, Websockets, and HTTP transports. - **One `bunx` Command**: Simplifies the setup process with a single command. ### Tech Stack: JavaScript ### Why it Matters: This project is crucial for anyone looking to automate their smart home setup with AI-driven control, making it easier to manage and control your devices. The integration with Home Assistant via the Model Context Protocol ensures seamless communication, enhancing the user experience. ### Repository: Repository: jango-blockchained/advanced-homeassistant-mcp GitHub Stars: 58 GitHub Description: MCP server that gives AI assistants control over your Home Assistant. Including 50+ tools over 3 transports. Batteries included 🔋