Projects

### Overview: SynCD is a pipeline for generating multi-image synthetic data for text-to-image customization. It uses either explicit 3D object assets or implicitly through masked shared attention across different views. The dataset is available for download and can be fine-tuned with the FLUX.1-dev model. ### Key Features: - Multi-image synthetic data generation - Consistent object identity for deformable categories - 3D multiview consistency for rigid object categories - Aesthetic score and DINOv2 filtering for quality enhancement - Masked Shared Attention (MSA) for foreground object features in images ### Tech Stack: Python ### Why it Matters: SynCD offers a practical solution for generating high-quality synthetic data for text-to-image customization, which can be used in various applications such as image editing, personalization, and augmented reality. The ability to generate consistent and visually appealing images from text prompts makes it particularly useful in the domain of text-to-image customization.

Overview: Context-Keeper 是一个基于LLM驱动的智能记忆与上下文管理系统，旨在解决AI工具在记忆与上下文理解方面的局限，提供更智能、高效的工作体验。 Key features: - 多维宽召回：通过语义检索、时间线检索和知识图谱检索，实现高覆盖率和高准确率。 - LLM精排序：结合LLM智能分析、质量评估和相关性排序，提供精准结果。 - 多维融合：通过语义、时间、知识维度的深度融合，实现个性化输出。 Tech stack: Go Why it matters: Context-Keeper 是一个智能解决方案，旨在打破传统记忆工具的边界，提供更智能、高效的上下文理解和记忆能力，从而提升开发团队的效率和决策质量。

### Overview: 这是一个简单的图片搜索引擎项目，通过三步构建，帮助开发者掌握向量数据库和以图搜图、文本搜索图片的技术。 ### Key features: - 使用FFmpeg生成模型，利用图片数据集。 - 构建轻量级向量数据库应用，实现图片和文本搜索。 - 三步构建，快速掌握向量数据库和图像搜索技术。 ### Tech stack: Python ### Why it matters: 掌握向量数据库和以图搜图、文本搜索图片的技术，是现代图像处理和搜索引擎开发中的重要技能。

Overview: This repository is a landscape of the infrastructure powering the generative AI ecosystem. Key features: - Open source projects - Contribution guidelines - Data format for projects - Build and serve site instructions Tech stack: HTML Why it matters: This repository serves as a resource for developers looking to understand and contribute to the infrastructure supporting generative AI projects.

### Overview: GLM-5.2-NVFP4-REAP-469B is a turnkey Docker setup designed for serving the **0xSero/GLM-5.2-NVFP4-REAP-469B** model on **4× NVIDIA RTX PRO 6000 Blackwell (SM120)** GPUs. This model is optimized with REAP pruning, NVFP4, and DeepSeek Sparse Attention, providing a high-performance vLLM launch recipe with 250K context, 60 tokens per second decode speed, and FP8 KV cache. The model is validated for 313 GB disk usage and 250k context, making it ideal for applications requiring quick, high-quality responses. ### Key Features: - **Model**: 0xSero/GLM-5.2-NVFP4-REAP-469B - **Context Size**: 250K - **Decoding Speed**: ~60 tokens/s - **Max Tokens**: ~313 GB - **GPU Configuration**: 4× RTX PRO 6000 Blackwell (SM120), 96 GB each - **Interconnect**: PCIe - **Model Pruning**: REAP-pruned - **MoE Loaders**: NVFP4 - **Sparse Attention**: DeepSeek Sparse Attention - **MTP Speculative Decode**: Enabled - **Reasoning**: On by default - **Max Tokens

### Overview: Anchoring AI is an open-source no-code tool designed for teams to collaborate on building, evaluating, and hosting applications leveraging GPT and other large language models. It simplifies the process of creating, managing, and sharing AI-powered applications, including budgeting and batch job handling. ### Key Features: - **No-Code Interface**: Quick and easy app creation using language models. - **Modular Design**: Easy addition of custom models, datasets, and extensions. - **Drag-and-Drop**: Chain components to create powerful applications. - **Batch Processing**: Efficient evaluation and repetitive task handling. - **Prompt Management**: Easy management of prompts and chains. - **Easy Sharing**: Streamline collaboration and sharing. - **Secure Access**: Customizable authentication for team management. - **Langchain Integration**: Seamless compatibility with Langchain (Python). - **Optimized Caching**: Cost-effective and performance-boosting caching. ### Tech Stack: JavaScript ### Why it Matters: Anchoring AI is a valuable tool for teams looking to transform their operations into an AI-centric powerhouse. It simplifies the process of building, managing, and sharing AI-powered applications, making it easier for teams to leverage the power of large language models without the need for advanced coding skills. The tool's modular design, secure access, and efficient batch processing capabilities make it a versatile solution for various projects. ### GitHub Stars: 155 ### Repository: AnchoringAI/anchoring-ai

Overview: CrawlAI RAG is an AI-powered website intelligence platform that transforms static websites into queryable knowledge bases, allowing users to crawl entire websites, index their content, and ask natural-language questions using Retrieval-Augmented Generation (RAG). Key features: - Crawling entire websites - Indexing content - Natural-language question answering using RAG - Multi-website indexing - Fast and scalable backend - Simple frontend - Secure configuration Tech stack: Python, FastAPI, Streamlit, ChromaDB, LangChain, Sentence-Transformers, Groq (LLaMA 3.3 70B), BeautifulSoup4 & Playwright, python-dotenv Why it matters: CrawlAI RAG provides a powerful tool for website intelligence, enabling users to quickly and accurately answer questions about websites, making it invaluable for businesses and researchers looking to extract valuable information from their online presence. Repository: AnkitNayak-dev/CrawlAI-RAG GitHub stars: 155

Overview: Transcriber is an open-source, privacy-first web application powered by OpenAI Whisper and FastAPI. It allows you to transcribe YouTube videos or local audio/video files instantly on your machine, without any data leaks or subscription fees. Key features: - YouTube & Local Support: Transcribe directly from YouTube URLs or upload `.mp3`, `.mp4`, `.wav`, and more. - Real-Time Progress: Watch the transcription status live via WebSockets — no more guessing when it finishes. - Smart Library: Manage your media and transcriptions in one place. Auto-correlates audio files with their transcripts and allows one-click cleanup of paired files. - Filename Sanitization: Automatic handling of special characters, accents, and spaces in filenames for robust cross-platform compatibility. - Dual AI Engine: `openai-whisper` (default) and `faster-whisper` (fast fallback). - GPU Acceleration: Automatic CUDA support for NVIDIA GPUs, with robust CPU fallback. - Zero-Config Setup: Auto-FFmpeg, yt-dlp Python API, and auto-installation. - Production Ready: Optimized Docker image (multi-stage build, non-root user, secure). - Modern UI: Dark/light mode, responsive design, history management, and file library. Tech stack: Python 3.13+, FastAPI, WebSockets, Alpine.js, Tailwind CSS. Why it matters: Transcriber provides a fast, privacy-focused solution for transcribing audio and video content, making it ideal for developers and content creators who value data security and control over

### Overview: RIVAL is an official implementation of the NeurIPS 2023 Spotlight paper "Real-World Image Variation by Aligning Diffusion Inversion Chain". It provides comprehensive tools for generating real-world image variations, text-to-image, image editing, and more, with a focus on aligning diffusion inversion chains. ### Key Features: - **Image Variations**: Generates images with the same semantic contents and style as the reference image. - **Editing-based Applications**: Allows users to modify images based on semantic concepts, with the option to customize object and style concepts. - **Customized Concept Editing**: Enables the creation of customized concepts for both object and style. - **Example-Based Inpainting**: Provides an example-based inpainting feature, although its application scope is limited as per the paper. - **Text-Driven Image Generation**: Generates images based on text prompts. - **ControlNet Integration**: Supports ControlNet for more advanced image manipulation. ### Tech Stack: Python ### Why it Matters: RIVAL offers a powerful toolset for researchers and developers interested in image generation and manipulation, making it a valuable resource for creating realistic and contextually relevant images.

### Overview: This repository is a curated list of papers and codes related to Efficient Diffusion Language Models (dLLMs) presented at the EMNLP 2026 conference. It includes a comprehensive survey of papers focusing on training efficiency, inference acceleration, and context scalability. ### Key Features: - **Training Efficiency**: Papers that explore methods to scale diffusion language models (dLLMs) from autoregressive models. - **Inference Acceleration**: Techniques to speed up inference, such as parallel decoding and cache management. - **Inference Acceleration**: Methods to optimize inference, including compression techniques. - **Contextual Scalability**: Papers that address the scalability of models with variable and long context lengths. - **System Framework**: Papers that discuss the architecture and frameworks used in dLLMs. - **Multimodal DLM**: Research on multimodal diffusion language models. - **Citations**: A collection of relevant citations and references. ### Tech Stack: The primary programming language used in the repository is unknown. ### Why it Matters: This repository provides a valuable resource for researchers and developers interested in the latest advancements in dLLMs, focusing on their training efficiency, inference acceleration, and scalability. It serves as a comprehensive guide to understanding and implementing efficient dLLMs in various applications.

### Overview: DiDi-Instruct is a training-based method that utilizes a pre-trained diffusion large language model (dLLM) to distill a few-step student for fast language generation. It matches or surpasses the performance of the dLLM teacher and the GPT-2 baseline while accelerating the generation process by up to 64×. The theoretical foundation of DiDi-Instruct is based on integral KL-divergence minimization, leading to practical training algorithms. Additionally, the method includes grouped reward normalization, intermediate-state matching, and a reward-guided ancestral sampler to enhance training stability, model coverage, and inference quality. ### Key features: - **Training-based distillation method**: Initializes from a pre-trained dLLM and distills a few-step student. - **64× acceleration**: Matches or surpasses the performance of the dLLM teacher and the GPT-2 baseline. - **Theoretical foundation**: Novel framework based on integral KL-divergence minimization. - **Improved training algorithms**: Grouped reward normalization, intermediate-state matching, and reward-guided ancestral sampler. - **Benchmark performance**: Achieves perplexity ranging from 62.2 (8 NFEs) to 18.4 (128 NFEs) on OpenWebText, outperforming prior accelerated dLLMs and the GPT-2 baseline. - **Robustness and effectiveness**: Extensive ablation studies, model scaling, downstream task evaluations, and unconditional protein sequence generation validate the robustness and effectiveness of DiDi-Instruct. ### Tech

### Overview: SmarterRouter is an intelligent, multi-backend AI router designed for local AI labs. It automatically selects the best model for each prompt, profiles models for performance, and routes queries to the most efficient provider—whether it's Ollama, llama.cpp, or OpenAI. ### Key Features: - **Zero Manual Model Selection** - AI automatically picks the right model for each prompt. - **All Local, Zero Cost** - No cloud API fees, works with your existing models. - **Production-Ready** - Built-in monitoring, metrics, and error handling. - **Drop-in Replacement** - Works with any OpenAI-compatible client. ### Tech Stack: Python ### Why It Matters: SmarterRouter simplifies the process of routing queries to the optimal model for tasks, making it ideal for local AI labs where cost and performance are critical concerns. Its production-ready features ensure seamless integration with existing infrastructure, and its zero-code changes and zero-cost setup make it an attractive alternative to more complex solutions. ### GitHub Stars: 154 ### Repository: peva3/SmarterRouter ### Quick Start (5 minutes): ```bash # 1. Clone the repository git clone https://github.com/peva3/SmarterRouter.git cd SmarterRouter # 2. Start with Docker Compose docker-compose up -d # 3. Verify it's running curl http://localhost:11436/health ``` ### Interactive Setup Wizard (New in v2.1.5): ```bash

### Overview: This project is a vector search demo using RedisVL, HuggingFace, OpenAI, and Cohere embeddings, implemented with FastAPI, React, and Redis. The app is built as a Single Page Application (SPA) and utilizes Redis as the vector database, RedisVL for Python, and FastAPI for the backend API. ### Key Features: - **Vector Search**: Utilizes RedisVL for vector database and RedisVL for Python client. - **Embedding Providers**: Supports HuggingFace, OpenAI, and Cohere embeddings. - **FastAPI**: Backend API for handling requests. - **React**: Frontend for the SPA application. - **Redis**: Highly performant vector database for efficient vector search. ### Tech Stack: - Python - Redis - FastAPI - Pydantic - React - Docker Compose - MaterialUI - React-Bootstrap ### Why It Matters: This project showcases the integration of AI-powered semantic search with RedisVL, providing a practical application for semantic document retrieval. The use of vector embeddings and RedisVL demonstrates how AI can enhance the performance and efficiency of vector databases, making it a valuable tool for developers looking to implement semantic search in their applications.

Overview: This repository contains the official implementation of "Learning to Generate Realistic LiDAR Point Clouds" (LiDARGen), a paper presented at ECCV 2022. The repository includes the code for training, evaluation, and denoising visualization of the model. Key features: - Training and evaluation of the LiDARGen model using Python. - Pretrained model for KITTI-360 dataset. - Sampling from the pretrained KITTI-360 model. - Metrics for evaluation, including FID, MMD, and JSD. - Denoising visualization and densification of the LiDARGen model. Tech stack: Python. Why it matters: This repository provides a comprehensive implementation of the LiDARGen model, allowing researchers and developers to easily reproduce and evaluate the results of the paper "Learning to Generate Realistic LiDAR Point Clouds" (ECCV 2022).

Overview: GraphRAG-UI is a user-friendly interface designed for the GraphRAG tool, utilizing the Retrieval-Augmented Generation (RAG) approach to index and query large text data. It supports the latest version of GraphRAG-0.3.3 and provides a convenient management and interaction method for GraphRAG, including configuration of local large language models like Ollama. Key features: - Intuitive web interface for easy configuration and use of GraphRAG. - Quick creation, update, and management of text data indexes. - Ability to submit natural language queries and retrieve relevant content from indexed data, followed by responses from a large language model. - Customizable settings and parameters to fine-tune indexing and querying processes. - Monitoring of indexing and querying tasks through detailed logs and status updates. Tech stack: Python Why it matters: GraphRAG-UI simplifies the management and interaction with GraphRAG, making it easier for users to leverage powerful large language models like Ollama for indexing and querying text data.

### Overview: This repository contains the official implementation of "Maximum Likelihood Training of Score-Based Diffusion Models" by Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon. The paper was published in NeurIPS 2021 (spotlight) and presents a new method for training score-based generative models that is more efficient than neural ODEs. ### Key features: - **Continuous Normalizing Flow Model**: The repository provides the official implementation of a continuous normalizing flow model called ScoreFlow, which is trained with a variational objective. - **Efficiency**: ScoreFlow is significantly more efficient than neural ODEs, achieving state-of-the-art likelihood on CIFAR-10 and ImageNet 32x32. - **Theoretical Contributions**: The repository proves the connection between the Kullback–Leibler divergence and the weighted combination of score matching losses, generalizing de Bruijn's identity in information theory and the evidence lower bound in variational inference. ### Tech stack: Python ### Why it matters: The repository's theoretical contributions and practical efficiency make it a significant contribution to the field of generative modeling, particularly in the area of score-based diffusion models. Its ability to achieve state-of-the-art performance on benchmark datasets and its potential for bridging the gap between continuous and discrete models makes it a valuable resource for researchers and practitioners in the field.

### Overview This repository contains a PyTorch implementation of our paper titled "A Recipe for Watermarking Diffusion Models". Our paper explores the addition of an "invisible watermark" to diffusion models (DMs) generated content, focusing on two main cases: unconditional/class-conditional DMs and multimodal text-to-image DMs. ### Key Features - **Unconditional/class-conditional DMs**: We investigate embedding a binary bit-string into generated images and pursue perfect detection/recovery from generated images using a pretrained bit-string decoder. - **Multimodal text-to-image DMs**: We investigate embedding a predefined image-text pair into the pretrained models and pursue perfect reconstruction/generation of the predefined image conditioned on the predefined trigger (text) prompt. ### Tech Stack The repository uses Python with PyTorch as the primary framework. ### Why It Matters Our work is pioneering in the field of watermarking diffusion models, providing a comprehensive approach to embedding watermarks in DMs and ensuring their detection and recovery. This method can be crucial for verifying the authenticity of generated content, enhancing security, and ensuring the integrity of the watermarking process. ### Repository Repository: [yunqing-me/WatermarkDM](https://github.com/yunqing-me/WatermarkDM) GitHub Stars: 155 Primary Language: Jupyter Notebook GitHub Description: A Recipe for Watermarking Diffusion Models README Excerpt: <!-- <p align='center' style="text-align:center;font-size:1.25em;"> <a href="https://scholar.google

### Overview: AIBot is an autonomous Minecraft AI agent that uses a DeepSeek-compatible language model to plan and execute tasks. It runs on a server-side Fabric mod for Minecraft 1.21.3, allowing users to give natural-language commands to mine, build, farm, fight, and survive. The mod uses deterministic task execution and does not allow improvisation, making it suitable for environments where precise control is necessary. ### Key Features: - **Goal Planning and Execution**: The language model plans goals and executes them through deterministic task state machines. - **63 Tools and 34 Task State Machines**: The mod uses 63 tools and 34 task state machines to execute goals. - **Strict Survival Mode**: The default profile disables certain features to ensure a strict survival environment. - **Compatibility Profile**: The `operator` profile allows for independent configuration of privileged capabilities. - **Versioned Runtime Snapshot**: The mod maintains a versioned runtime snapshot for safety and reliability. ### Tech Stack: Java ### Why It Matters: AIBot provides a server-side Minecraft AI that can be used for various tasks, ensuring precise and controlled gameplay. Its deterministic execution and strict survival mode make it suitable for environments where precise control is necessary, such as survival or crafting.

Overview: This sample demonstrates how to build an AI chat experience using Retrieval-Augmented Generation (RAG) with LangChain.js and OpenAI language models. The application is hosted on Azure Static Web Apps and Azure Container Apps, utilizing Azure AI Search as the vector database. It serves as a starting point for developing more complex AI applications. Key features: - Azure Static Web Apps - Azure Container Apps - Azure AI Search (vector database) - LangChain.js - OpenAI SDK - Node.js Tech stack: Bicep Why it matters: This sample provides a practical guide on how to build an AI chatbot using RAG, showcasing the integration of LangChain.js with OpenAI models, and the use of Azure services for hosting and deployment. It is designed to be a foundational resource for developers interested in creating AI-driven applications.

Overview: Petrichor is a self-hosted knowledge platform designed for humans and AI agents, enabling the creation and sharing of structured knowledge through Markdown, evidence, and Agent Skills. Key features: - Structured writing with PlateJS, Markdown, code blocks, formulas, tables, mind maps, and media embedding. - Knowledge library and publishing capabilities, including multi-level directories, tags, full-text search, article sharing, public Q&A, and evidence-based articles. - Agentic RAG (Relevance and Aggregation of Knowledge) with title-aware slicing, question recommendations, BM25 / Vector / Wiki fusion, directory navigation, and Evidence / Trace. - Semantic Wiki for entity and concept extraction, multi-article aggregation, relationship graph, source citations, and structural checks. - Pi Agent Runtime with unified support for internal and external MCP clients, Skill files, Jina-compatible model reordering, sandboxed isolation, and Playwright browser execution. - Agent Skill extension with external MCP clients, fileized SKILL.md, compatibility with Jina models, isolation code sandbox, and Playwright browser execution. - Open integration with API keys, MCP, REST, Skill packages, capability list, and full call audit. - Self-hosted infrastructure with Go + Gin, PostgreSQL + pgvector, Redis + Asynq, S3-compatible storage, and Caddy. Why it matters: Petrichor offers a robust, self-contained platform that separates knowledge retrieval and reading, enabling efficient and context-aware interactions with AI agents. It supports multiple integration points and customizable models, making it a versatile tool for knowledge management