Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 93 stars

duongngocbinh56599-ui/local-llm-lab

Honest local LLM deployment planning and benchmarking for high unified-memory Macs and future Linux/NVIDIA rigs.

Overview: local-llm-lab is a lightweight CLI tool designed for developers with unified-memory machines, offering a quick and accurate assessment of local hardware suitability for large language models (LLMs). Key features: - Detects local hardware specifications - Saves and reuses measured profiles - Plans whether a model fits within local resources - Generates lightweight CLI scripts for local deployment - Benchmarks with deterministic mock data - Performs stress simulations to test resource contention - Reports in Markdown, JSON, SVG charts, and static HTML - Serves reports locally with a localhost-only Web viewer Tech stack: Python Why it matters: It simplifies the process of deploying large LLMs on local hardware, ensuring optimal performance and resource management. Repository: duongngocbinh56599-ui/local-llm-lab Primary language: Python GitHub stars: 94 GitHub description: Honest local LLM deployment planning and benchmarking for high unified-memory Macs and future Linux/NVIDIA rigs.

TypeScript 93 stars

erhanmeydan/kilnframe

Free, open source plugin for Premiere Pro and DaVinci Resolve: turns what you say into animated inserts and music, rendered on your own computer

### Overview: Kilnframe is a free, open-source plugin that transforms spoken content into animated inserts and music for Adobe Premiere Pro, DaVinci Resolve Studio, and FilmCraft. It automatically generates HTML, CSS, and JavaScript animations based on the speech, rendering them locally on your computer. ### Key Features: - **Eleven Styles and Your Own**: Each style is designed to match the content, such as a bar chart in Navy Clay or a Whiteboard drawing. - **Fullscreen or Over**: Scenes can be rendered in full screen or placed over your footage. - **You Approve Before Rendering**: Each scene is reviewed before rendering, allowing for easy modifications. - **Sound Effects on Their Own Track**: Sound effects are synthesized and placed on their own track. - **Music for the Whole Video**: The Music tab scores the timeline or a range with Google Lyria, placing music on its own track. - **Your Plan or Key**: Use Claude Code, Codex, or a subscription with an API key. - **Renders Locally**: Your audio and video remain on your computer, never leaving it. ### Tech Stack: TypeScript ### Why It Matters: Kilnframe simplifies the process of adding dynamic, engaging content to your projects, making them more interactive and visually appealing. It's particularly useful for creating educational videos, presentations, and other media where visual storytelling is crucial.

Unknown 93 stars

fabiotosi92/Diffusion4RobustDepth

[ECCV 2024] Diffusion Models for Monocular Depth Estimation: Overcoming Challenging Conditions

### Overview: This repository contains information about our work "Diffusion Models for Monocular Depth Estimation: Overcoming Challenging Conditions," presented at ECCV 2024 by Fabio Tosi, Pierluigi Zama Ramirez, and Matteo Poggi from the University of Bologna. ### Key Features: - Novel use of text-to-image diffusion models with depth-aware control for generating challenging scenes with known depth information. - A unified framework that addresses multiple challenges in single-image depth estimation, including adverse weather conditions and non-Lambertian surfaces. - Pioneering approach to enhancing the robustness of monocular depth estimation models in out-of-distribution scenarios. ### Tech Stack: Unknown ### Why It Matters: This work introduces innovative techniques for improving the robustness of monocular depth estimation models in challenging conditions, making it a significant contribution to the field of computer vision.

Python 93 stars

giannisdaras/ambient-diffusion

[NeurIPS 2023] Official Implementation: "Ambient Diffusion: Learning Clean Distributions from Corrupted Data"

Overview: This repository contains the official PyTorch implementation of the paper "Ambient Diffusion: Learning Clean Distributions from Corrupted Data," presented at NeurIPS 2023. Authored by Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alexandros G. Dimakis, and Adam Klivans, it introduces a novel diffusion-based framework capable of learning an unknown distribution using highly-corrupted samples. Key features: - The implementation of the Ambient Diffusion framework. - Training of generative models that do not memorize individual training samples. - Ability to learn the conditional expectation of the full uncorrupted image given the additional measurement corruption. - Training on standard benchmarks such as CelebA, CIFAR-10, and AFHQ. - Demonstrated capability to learn the distribution even when all training samples have 90% of their pixels missing. - Finetuning of foundation models using small corrupted datasets (e.g., MRI scans with block corruptions). Tech stack: The repository uses Python for implementation and relies on the Diffusers library for diffusion models. Why it matters: This approach offers a unique solution to the problem of learning from highly-corrupted data, which is crucial in scientific applications where access to uncorrupted samples is limited or expensive. By training generative models that do not memorize individual training samples, it ensures that the learned distribution is more generalizable and robust. GitHub stars: 93 GitHub description: [NeurIPS 2023] Official Implementation:

Go 93 stars

greynewell/infermux

Route inference across providers.

Overview: Infermux is an inference router designed for seamless communication across multiple providers. Key features: - Supports multiple providers for different models - Tracks tokens and cost per request - Reports spans to TokenTrace - Provides an HTTP API and gRPC API - Includes a standard gRPC health service and server reflection - Offers graceful drain on SIGINT/SIGTERM - Implements error contract for retryable errors Tech stack: Go Why it matters: Infermux simplifies the process of routing inference requests across different providers, ensuring efficient and reliable communication.

Python 93 stars

icon-lab/SelfRDB

Official PyTorch implementation of SelfRDB, a diffusion bridge model for multi-modal medical image synthesis

Overview: SelfRDB is an official PyTorch implementation of a novel diffusion bridge model designed for multi-modal medical image synthesis. It introduces a soft-prior mechanism and self-consistent recursion in the reverse process to enhance generalization and information transfer between modalities. Key features: - Soft-prior mechanism for boosting generalization performance - Self-consistent recursion for improved sampling accuracy in reverse steps - Novel noise scheduling with monotonically increasing variance towards the end-point Tech stack: Python Why it matters: SelfRDB offers a significant improvement in the synthesis of medical images by enhancing the transfer of information between modalities and improving sampling accuracy, making it a valuable tool in medical image processing and analysis.

Python 93 stars

jjwang/laibot-client

开源人工智能，基于开源软硬件构建语音对话机器人、智能音箱……人机对话、自然交互，来宝拥有无限可能。特别说明，来宝运行于Python 3！

Overview: 来宝是一个基于开源软硬件构建的语音对话机器人项目，旨在通过Python 3运行，提供人机对话和自然交互体验。 Key features: - 基于树莓派的语音对话机器人 - 自动化语音服务集成 - 优化用户体验的智能音箱解决方案 Tech stack: Python 3 Why it matters: 来宝项目致力于探索基于开源软硬件和免费语音服务的语音助手解决方案，旨在提供更好的用户体验和智能化服务。

HTML 93 stars

nolangz/data2motion

Overview: Data2Motion is an AI skill designed to convert JSON data, KPIs, rankings, trends, and small datasets into polished animated charts. It simplifies the process of creating data stories into a standalone HTML file with built-in controls and deterministic QA hooks. Key features: - Converts JSON data into animated charts - Supports various chart types including KPIs, bar, column, diverging, line, multiline, area, stacked, scatter, lollipop, and dumbbell - Provides a fixed chart vocabulary and two launch-ready styles - Includes built-in replay, speed, freeze, and deterministic QA hooks - Outputs a standalone HTML file from a single JSON spec Tech stack: The repository uses Python scripts and Codex and Claude skills for the Lite engine. Why it matters: This tool streamlines the process of creating data stories, making it easier to visualize trends, comparisons, and other data points in a dynamic and engaging way. Repository: nolangz/data2motion GitHub stars: 92 GitHub description: No description provided. README excerpt: # Data2Motion - AI Data Animation Skill for Animated Charts [![GitHub stars](https://img.shields.io/github/stars/nolangz/data2motion?style=social)](https://github.com/nolangz/data2motion/stargazers) [![GitHub forks](https://img.shields.io/github/forks/nolangz/data2motion?style=social)](https://github.com/nolangz/data2motion/forks) [![Codex Skill](https://img.sh

Jupyter Notebook 93 stars

poloclub/LLM-Attributor

LLM Attributor: Attribute LLM's Generated Text to Training Data

### Overview: LLM Attributor is a tool designed to help visualize and understand how large language models (LLMs) generate specific text phrases. By selecting text phrases and observing the corresponding training data points, users can gain insights into the training data that influenced the model's output. ### Key Features: - Interactive visualization of training data attribution for text generation. - Ability to modify model-generated text and see the impact on attribution. - Compatible with Jupyter Notebooks, making it easy to integrate into existing computational workflows. ### Tech Stack: - Python - Jupyter Notebook ### Why It Matters: LLM Attributor provides a powerful way to understand the origins of text generated by LLMs, making it easier to debug and improve model performance by identifying and modifying the training data that significantly influences the model's output. ### Repository: poloclub/LLM-Attributor - GitHub stars: 93 - GitHub description: LLM Attributor: Attribute LLM's Generated Text to Training Data - Primary language: Jupyter Notebook - GitHub description: LLM Attributor helps you visualize training data attribution of text generation of your large language models (LLMs). Interactively select text phrases and visualize the training data points responsible for generating the selected phrases. Easily modify model-generated text and observe how your changes affect the attribution with a visualized side-by-side comparison. ### Getting Started: - Installation: `pip install llm-attributor` - Initialization: `from LLMAttributor import LLMAttributor; attributor = LLMAttributor

TypeScript 93 stars

pouyahasanamreji/continuum

Shared memory + orchestration for your coding agents — one MCP server, persistent vector memory, agent registry

### Overview: Continuum is a shared memory and orchestration system designed for coding agents, providing a single MCP server, persistent vector memory, and an agent registry. It allows clients to read the plot, dispatch fresh agents in isolated worktrees, and share lessons across all Claude Code, Codex, and Cline chats on the project. ### Key Features: - **Shared Memory and Orchestration**: One MCP server, persistent vector memory, and an agent registry. - **Orchestrated Agents**: Clients can read the plot and become orchestrators, dispatching fresh agents in isolated worktrees. - **Persistent Vector Memory**: Every lesson learned by agents is indexed in `sqlite-vec`, providing semantic and fuzzy recall. - **Canonical Workflow**: A 4-phase dispatch protocol (Intake → Research → Verify → Handoff) to stop re-explaining the process. - **Multi-Agent Registry**: State machine and reserved-path tracking to prevent parallel agents from clobbering each other. - **Two-Tier Knowledge**: Fundamental lessons are loaded on every dispatch, while situational lessons surface via semantic search when relevant. - **Human Web Panel**: Astro + React UI for browsing projects, agents, plot, and knowledge, with AI clients driving everything via MCP. - **Client-Agnostic**: Standard MCP / Streamable HTTP, works with Claude Code, Codex, Cline, Cursor, or anything that speaks the protocol. - **Local First**: One SQLite file, WAL mode, no cloud, optional embedder is your call (Ollama, T ### Tech

Unknown 93 stars

tingruew/DiffusionModels-In-RecSys

Overview: This repository provides a comprehensive survey of diffusion models applied to recommendation systems. It includes detailed information on various core tasks such as collaborative filtering, implicit feedback, explicit ratings, item graph, and user graph, as well as multimodal recommendation and cross-domain recommendation. The survey is based on recent research papers and is accepted by ACM TORS. Key features: - Comprehensive survey of diffusion models in recommendation systems - Detailed information on core recommendation tasks - Papers on core recommendation tasks, including implicit feedback, collaborative filtering, explicit ratings, item graph, user graph, and multimodal recommendation - Papers on data modality and domain, including image generation, text-to-recommendation, multimodal recommendation, and cross-domain recommendation - Papers on trustworthy objectives and constraints, including fairness, accountability, transparency, and out-of-distribution Tech stack: Unknown Why it matters: This survey is significant as it provides a clear overview of the current state of research in diffusion models applied to recommendation systems, making it a valuable resource for developers and researchers in the field.

Unknown 93 stars

toki-plus/ai-mixed-cut

颠覆性AI内容再创作引擎，通过“解构-重构”爆款视频模式，全自动生成高度原创短视频。AIGC, Content Creation, Video Generation, Automation, LLM, Python, FFmpeg.

### Overview: AI Mixed Cut 是一个颠覆性的AI内容再创作引擎，通过“解构-重构”爆款视频模式，实现全自动生成高度原创短视频。它结合了AIGC、Content Creation、Video Generation、Automation、LLM、Python、FFmpeg等技术，旨在提高内容团队的工作效率，同时保留人工审核的必要性。 ### Key features: - **解构与重构爆款视频模式**：通过“解构”提取视频内容，再通过“重构”生成高度原创的短视频。 - **素材库结构化管理**：将视频、文案和音频素材分解为结构化的素材库，便于管理和再生产。 - **脚本重组与语音生成**：从素材库中重组脚本，生成高质量的音频，支持多语言语音合成。 - **视频合成与二次创作**：支持视频的顺序、随机拼接，提供多种画面处理选项，增强内容差异化。 ### Tech stack: Python, FFmpeg, LLM, Edge TTS, NVIDIA (NVENC) ### Why it matters: AI Mixed Cut 通过自动化流程显著提高了内容团队的工作效率，同时保留了人工审核的必要性，从而提升了内容的原创性和质量。它为内容团队提供了一个高效、灵活的工作流，适用于大量历史素材的整理与再生产，以及需要将长内容拆分为多个主题内容的编辑人员。

Python 92 stars

AstraBert/ingest-anything

From data to vector database effortlessly

Overview: `ingest-anything` is a Python package designed to automate the process of ingesting non-PDF files into vector databases, making it easier to convert text files into vector embeddings. Key features: - Ingests text files into vector databases - Ingests code files into vector databases - Ingests HTML content from URLs into vector databases - Provides a fully-automated pipeline for document ingestion Tech stack: Python Why it matters: Simplifies the process of converting text files into vector embeddings, making it easier to integrate with vector databases for text and code analysis.

Jupyter Notebook 92 stars

FareedKhan-dev/create-stable-diffusion-from-scratch

Implemented a stable diffusion architecture using PyTorch.

Overview: This project details the implementation of a stable diffusion architecture using PyTorch, focusing on creating a small-scale model from scratch using the MNIST dataset. Key features: - Implementing a stable diffusion architecture using PyTorch - Using the MNIST dataset for training and inference - Available models checkpoints in the `models_checkpoints` folder for inference Tech stack: PyTorch Why it matters: This project showcases the feasibility of creating a stable diffusion model from scratch, providing insights into the computational and data requirements for such a task.

Unknown 92 stars

JingyeChen/awesome-game-generation

🕹️ Explore cutting-edge techniques in game generation

### Overview: Explore cutting-edge techniques in game generation, including interactive video generation, world models, and more. ### Key features: - 📄 Academic research papers and publications - 📝 Technical writeups and blog posts - 🎮 Game demos and playable implementations - 🎲 Board game and tabletop game implementations - 📊 Datasets and training data - 📹 Video demonstrations and tutorials - 📰 Latest updates and announcements - 📚 Source code and implementations - 🔗 Useful links and resources ### Tech stack: Python, PyTorch, TensorFlow, Blender, Unity, and other game development tools. ### Why it matters: This repository serves as a curated list of resources for using artificial intelligence models in game development, focusing on game generation. It provides a comprehensive collection of research papers, blogs, games, demos, and other materials, encouraging contributions and welcoming contributions to maintain and expand the list.

TypeScript 92 stars

SheikhAminul/ChatGPT-voice-control

Voice control for ChatGPT. Talk to ChatGPT and hear ChatGPT's responses in a natural voice.

Overview: This project allows users to interact with ChatGPT through voice input, providing a natural and conversational experience. Key features: - Voice control for ChatGPT - Read aloud feature for ChatGPT's responses - Support for multiple languages Tech stack: TypeScript Why it matters: This extension offers a unique and user-friendly way to communicate with ChatGPT, enhancing the overall conversational experience for users. Repository: SheikhAminul/ChatGPT-voice-control Primary language: TypeScript GitHub stars: 92

Python 92 stars

Umesh-01/Python-Assistant

Python Assistant (PA) is a voice command based assistant service written in Python 3.9+. It can recognize human speech or voice, talk to user and execute basic commands.

Overview: Python Assistant (PA) is a voice command-based assistant service written in Python 3.9+. It can recognize human speech or voice, talk to users, and execute basic commands. Key features: - Tells the current time or date or month or day - Opens a web page - Tells about general questions - Play music on YouTube - Increase/decrease the speakers master volume - Tells the internet speed - Tells the weather for a particular city - Opens MS Office suite applications - Set an alarm - Tells the latest daily news - Write notes for reminder - Read notes - Search on Wikipedia - Guide us about something - Send messages on WhatsApp - Send emails - Tells the distance between two cities - Clean the Recycle Bin - Tells Quotes - Tells the COVID-19 cases worldwide - Tells the COVID-19 cases for specific country - Tells Jokes - Do Arithmetic operations - Convert currency - Take screenshots - Tells Poems - Close applications - Open applications - Shutdown or Restart computer - Searching anything on the internet Tech stack: Python 3.9+ Why it matters: Python Assistant (PA) is a powerful tool for automating tasks and providing quick responses to users, making it highly useful for developers and anyone looking to streamline their workflow.

JavaScript 92 stars

UncensoredLabs/uncensored-ai-api

Uncensored AI API and browser chat: DeepSeek and GLM models, streaming, Python and TypeScript examples, pricing and setup guides.

### Overview: Access uncensored and abliterated **DeepSeek and GLM text models** through one OpenAI-compatible API, or use them directly in browser Chat. Keep your client, change the base URL, and choose a model. ### Key features: - **DeepSeek and GLM Models**: Access uncensored and abliterated versions of DeepSeek and GLM text models. - **Streaming Support**: Streamlined responses for real-time interactions. - **Python and TypeScript Examples**: Examples in Python and TypeScript for seamless integration. - **Pricing and Setup Guides**: Detailed guides on how to set up and use the API. ### Tech stack: JavaScript ### Why it matters: UncensoredLabs provides a powerful and flexible way to access uncensored and abliterated text models, making it ideal for security researchers, reverse engineers, and developers looking to explore difficult themes and unconventional characters. The API is designed to be user-friendly and easy to integrate, ensuring that users can quickly start using the models in their projects.

Go 92 stars

VQLite/VQLite

VQLite - Simple and Lightweight Vector Search Engine based on Google ScaNN

Overview: VQLite is a lightweight and simple vector similarity search engine based on Google ScaNN, designed to meet the needs of projects that require a lightweight solution without the overhead of Kubernetes clusters. Key features: - RESTful API - Simple and Lightweight - Fast Tech stack: Go Why it matters: VQLite provides a simple and efficient way to build vector similarity search services without the need for heavy infrastructure, making it suitable for projects that require a lightweight solution without the burden of Kubernetes clusters. GitHub stars: 92 Repository: VQLite/VQLite Primary language: Go Design: VQLite's design allows for the storage of multiple vectors per document while still allowing the storage of metadata without the need for additional storage solutions such as MySQL or Redis. This makes it suitable for projects that require a one-to-many relationship between documents and vectors. Benchmark: VQLite was tested on 13 million vectors and achieved high QPS (queries per second) on AWS machines, with the best performance on c5.9xlarge and c7g.8xlarge machines. Requirements: Linux or MacOS, Golang. Get Started: VQLite can be compiled from source code or run in a Docker container. The Docker container is recommended for easy deployment. Tips: To improve search speed, try to make each segment as large as possible, set it close to the memory limit, and minimize the number of segments. The reorder value has a significant impact on both speed and recall rate, with larger values resulting in slower retrieval speed but higher

Python 92 stars

athrael-soju/Snappy

🐊 Snappy's unique approach unifies vision-language late interaction with structured OCR for region-level knowledge retrieval. Like the project? Drop a star! ⭐

### Overview: Snappy is a cutting-edge project that integrates vision-language models with OCR to achieve region-level document retrieval. Unlike traditional systems, Snappy uses patch-level similarity scores to filter OCR-extracted regions, providing efficient and interpretable results. ### Key Features: - **Region-Level Document Retrieval**: Snappy retrieves specific paragraphs, tables, or figures instead of returning entire pages. - **Patch-to-Region Relevance Propagation**: This approach maps vision transformer patch grids to OCR bounding boxes, enabling efficient and interpretable results. - **Efficient Candidate Retrieval**: Uses mean-pooled embeddings for candidate retrieval. - **Full-Resolution Region Reranking**: Follows up with full-resolution region reranking for improved accuracy. ### Tech Stack: - **Python**: The primary programming language used for the project. - **FastAPI**: For building the API. - **Next.js**: For server-side rendering and efficient rendering. - **TypeScript**: For type safety and better code readability. - **Qdrant**: For efficient vector similarity search. - **Docker**: For containerization and easy deployment. ### Why It Matters: - **Efficiency**: Snappy operates purely at inference time, eliminating the need for additional training. - **Interpretability**: Provides visual heatmaps showing which document regions match query tokens, enhancing understanding. - **Integration**: Combines the strengths of VLMs and OCR systems, offering a production-ready solution for document retrieval. ### Repository: athrael-soju/Snappy - **GitHub Stars**: 92 - **GitHub Description**: