Projects

Overview: This project aims to train a 65M-parameter Visual Language Model (VLM) from scratch using only 3 dollars and 2 hours of GPU time. The MiniMind-V model is designed to be minimal in size and efficient in training and inference, making it accessible even on personal GPUs. Key features: - MiniMind-V: A 65M-parameter VLM trained from scratch. - MiniMind-V-MOE: A variant with multi-branch optimization. - MiniMind-V-Small: A smaller version of MiniMind-V with 65M parameters. - MiniMind-V-3V: An optimized version with 200M parameters. - MiniMind-V-3V-MOE: An optimized variant with multi-branch optimization. Tech stack: Python Why it matters: This project serves as a minimal implementation of a VLM model, providing a straightforward guide for beginners to understand and experiment with visual language models. It also showcases the feasibility of training large models without significant computational resources, making it an excellent resource for educational and research purposes. Repository: jingyaogong/minimind-v Primary language: Python GitHub stars: 8755 GitHub description: 👀 Train a 65M-parameter VLM from scratch in just 2h!

Overview: VAR is an ultra-simple, user-friendly yet state-of-the-art codebase for autoregressive image generation, achieving NeurIPS 2024 Best Paper Award and scaling laws in visual generation. Key features: - Scalable image generation - Next-scale prediction - User-friendly interface - NeurIPS 2024 Best Paper Award Tech stack: Jupyter Notebook Why it matters: VAR offers a simple yet powerful solution for image generation, demonstrating significant advancements in visual autoregressive modeling and scaling laws in the field of image generation. Repository: FoundationVision/VAR Primary language: Jupyter Notebook GitHub stars: 8758 GitHub description: [NeurIPS 2024 Best Paper Award][GPT beats diffusion🔥] [scaling laws in visual generation📈] Official impl. of "Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction". An *ultra-simple, user-friendly yet state-of-the-art* codebase for autoregressive image generation!

Overview: Reor is a private and local AI personal knowledge management app designed for high-entropy individuals, leveraging models like Ollama, Transformers.js, and LanceDB to enhance semantic search capabilities. Key features: - Local AI personal knowledge management - Automatic note linking and semantic search - Supports local LLMs and embedding models - Features an Obsidian-like markdown editor Tech stack: JavaScript Why it matters: Reor simplifies the process of managing personal knowledge by enabling local AI-powered insights, making it an invaluable tool for those who prefer to keep their thoughts and data private. Repository: reorproject/reor GitHub stars: 8543