Projects

### Overview: Flux AI Image Generator is an AI-powered image generation platform built using Next.js and the Flux.1 AI model. It allows users to create high-quality images from text prompts quickly and easily, with a prompt generator for optimization and international support. ### Key features: - **AI Image Generation**: Generate high-quality images based on text prompts using the Flux.1 AI model. - **Prompt Generator**: Improve and refine your prompts to get the best results from the AI. - **i18n Support**: Full internationalization support for multilingual audiences. - **Responsive Design**: Built with TailwindCSS to ensure a responsive and clean user interface on all devices. - **High Performance**: Hosted on Vercel with Cloudflare CDN for fast, secure, and reliable performance. ### Tech stack: Framework: Next.js Hosting: Vercel Domain: Dynadot CDN: Cloudflare Styling: TailwindCSS Template: Radix UI & Preline UI ### Why it matters: Flux AI Image Generator is a powerful tool for anyone looking to create high-quality images from text prompts quickly and efficiently. Its AI-powered capabilities, prompt generator, and international support make it a valuable asset for developers and creatives alike.

### Overview: MVD-Fusion is a novel approach for generating multi-view RGB-D images from a single input RGB image, leveraging depth-consistent multi-view generation. This project was presented at CVPR 2024 and aims to enhance the realism and consistency of 3D models in single-view scenes. ### Key Features: - **Depth-consistent Multi-view Generation**: Generates multi-view RGB-D images by enforcing depth consistency across views. - **Single-view 3D Reconstruction**: Produces 3D models from a single input RGB image, enhancing the realism of single-view scenes. - **Incorporation of Open-Source Libraries**: Built on top of Stable Diffusion, kiuikit, Zero-1-to-3, and SyncDreamer. ### Tech Stack: Python ### Why it Matters: MVD-Fusion offers a significant improvement in the quality and realism of single-view 3D reconstructions, making it particularly useful in applications such as augmented reality, virtual reality, and 3D visualization. ### Repository: The project is hosted on GitHub at [zhizdev/mvdfusion](https://github.com/zhizdev/mvdfusion). ### GitHub Stars: 133 ### Code Release: - Contains code for inference, training (coming soon), and pretrained weights on Objaverse. ### Environment Setup: Follow the guide in `ENVIRONMENT.md` for setting up the development environment. ### Dataset: - Provides evaluation datasets: Google Scanned Objects (GSO) and SyncDreamer in-the-wild dataset.

Overview: VTS is a tool designed for the transformation and transportation of vectors and unstructured data, developed by Zilliz based on Apache Seatunnel. Key features: - Real-time data streaming - Offline batch import - Vector embedding and transformation - High performance and scalability Tech stack: Java Why it matters: VTS simplifies the process of moving vectors and unstructured data, supporting real-time data streaming and offline batch imports, and enabling complex transformations with AI and model capabilities, ensuring data quality and operational efficiency. Repository: zilliztech/vts GitHub stars: 134

### Overview: AcademiaSD LoRAlab Trainer Studio is a comprehensive tool designed to train LoRA models efficiently on NVIDIA GPUs. It supports a wide range of models including Qwen-Image 2.1, FLUX.2 Klein 9B, FLUX.3 Image Klein, Krea 2, Z-Image, MOSS-Image, ERNIE-Image, Ideogram 4, Anima, SDXL (Pony, Illustrious, NoobAI), LTX 2.3, MiniMax-H3, and more. The tool is optimized for Windows, Linux, and RunPod, making it accessible across multiple operating systems. ### Key Features: - **Job Queue**: Manages multiple training jobs, allowing you to reorder, pause, resume, or remove them. - **Training Chart**: Visualizes the loss trend over time, with markers for previews and checkpoints. - **Saved LoRAs**: Downloads or sends LoRAs to your models folder for testing. - **3D Training Chart**: Displays the loss chart in 3D, allowing you to rotate it for better visualization. - **VRAM Read Through ADLX**: Supports VRAM read through ADLX on Windows and AMD GPUs. - **Hugging Face Upload**: Allows you to upload trained LoRAs to Hugging Face for further use. ### Tech Stack: - Python ### Why It Matters: AcademiaSD LoRAlab Trainer Studio simplifies the training process of LoRA models, making it easier to experiment with various models and

### Overview: Qwen3.8-Flash-Next is a high-performance language model optimized for 2× RTX 3090 GPUs, capable of processing up to 4,191 tok/s prefill and 111.5 tok/s decode. This model features a 128 GB system memory, making it ideal for applications requiring large-scale context and high throughput. ### Key Features: - **High Performance**: Up to 4,191 tok/s prefill and 111.5 tok/s decode. - **Large Context Window**: Supports up to 262,144-token context. - **Optimized for GPU**: Designed to maximize performance on 2× RTX 3090 GPUs. - **Large System Memory**: Utilizes 128 GB of system memory for optimal performance. ### Tech Stack: - Python ### Why It Matters: Qwen3.8-Flash-Next is particularly useful for applications that require high throughput and large context, such as language modeling, text generation, and dialogue systems. Its ability to handle large inputs and outputs efficiently makes it a valuable tool in various AI and machine learning applications. ### Repository: Repository: DominikBucko/qwen38-flash-next-2x3090 GitHub Stars: 130 ### Primary Language: Python ### GitHub Description: Qwen3.8-Flash-Next on 2× RTX 3090: with 128 GB RAM up to

### Overview: VLM-Grounder introduces a novel framework using vision-language models (VLMs) for zero-shot 3D visual grounding based solely on 2D images. This approach dynamically stitches image sequences, employs a grounding and feedback scheme to find the target object, and utilizes a multi-view ensemble projection to accurately estimate 3D bounding boxes. The method significantly outperforms previous zero-shot methods on the ScanRefer and Nr3D datasets, achieving 51.6% Acc@0.25 on ScanRefer and 48.0% Acc on Nr3D. ### Key Features: - **Dynamic Image Stitching:** VLM-Grounder dynamically stitches image sequences to improve the accuracy of 3D visual grounding. - **Grounding and Feedback Scheme:** Utilizes a grounding and feedback mechanism to find the target object in 3D space. - **Multi-View Ensemble Projection:** Employing a multi-view ensemble projection to accurately estimate 3D bounding boxes. ### Tech Stack: Python ### Why it Matters: VLM-Grounder offers a novel and effective method for zero-shot 3D visual grounding, significantly improving upon existing techniques without relying on 3D geometry or object priors.

Overview: EcoAssistant is a framework designed to make LLM-assisted code-driven question answering more affordable and accurate by leveraging assistant hierarchy and solution demonstration techniques. It is built on the AutoGen framework and aims to improve the success rate and accuracy of queries without increasing the cost. Key features: - Assistant hierarchy to choose the cheapest assistant for the query. - Solution demonstration to leverage past successful query-code pairs for future queries. - Code-driven question answering that requires iterative coding and feedback. Tech stack: Python Why it matters: EcoAssistant significantly improves the success rate and accuracy of queries by 10 points while maintaining or even reducing the cost compared to using individual GPT-4 assistants. Repository: JieyuZ2/EcoAssistant GitHub stars: 132 GitHub description: EcoAssistant: using LLM assistant more affordably and accurately

Overview: ComfyUI custom node for NVIDIA PiD pixel diffusion decoding and upscale workflows. Key features: - PiD Decode - PiD Text Prompt - PiD Caption Creator - PiD Empty Latent Image - PiD KSampler Capture - PiD Prepare - PiD Sample - PiD Finalize - PiD Upscale - PixelDiT Generate Tech stack: Python Why it matters: This node streamlines the process of using ComfyUI with NVIDIA PiD, enhancing image generation and decoding capabilities for the ComfyUI ecosystem.

### Overview: Walrus Memory is a portable memory layer designed to ensure reliable and verifiable AI agent operations across applications and sessions. It provides persistent and fully controllable memory, making it essential for developing robust and trustable AI solutions. ### Key Features: - **Reliability Across Sessions**: Ensures AI agents maintain context and data integrity across different sessions and applications. - **Verifiable Memory**: Data is stored and retrieved using a verifiable and controllable memory layer. - **Developer-Friendly**: Provides a single-file guide for integration, structured documentation, and a comprehensive API reference. ### Tech Stack: TypeScript ### Why It Matters: Walrus Memory is crucial for developers who need to build AI solutions that operate reliably and securely across multiple environments and applications. Its ability to handle complex workflows and coordinate data with trust and control makes it a valuable tool in the development of AI-driven applications. ### Repository: MystenLabs/walrus-memory ### GitHub Stars: 131 ### GitHub Description: A portable memory layer that makes AI agents reliable across apps and sessions, persistent, verifiable, and fully under your control.

### Overview: This project provides a Speech Recognition module for Asterisk, utilizing the Vosk API server. It is designed to enhance Asterisk's capabilities by integrating speech recognition technology, making it easier to manage and automate voice-based interactions. ### Key Features: - **Speech Recognition Module**: Integrates speech recognition capabilities directly into Asterisk. - **Vosk API Server Integration**: Utilizes the Vosk API server for speech recognition, ensuring seamless integration with Asterisk. - **Docker Support**: The project includes a Docker container for easy deployment and management. ### Tech Stack: - **Language**: C - **GitHub Stars**: 132 ### Why It Matters: This module enhances the functionality of Asterisk by providing advanced speech recognition capabilities, making it more versatile and efficient for voice-based applications and automated systems.

Overview: This project aims to detect human emotions using Mel frequency cepstral coefficients (MFCC) features extracted from speech data, employing a recurrent neural network (LSTM) for classification into six predefined categories: happy, sad, angry, frustrated, neutral, and fear. Key features: - Utilizes Mel frequency cepstral coefficients (MFCC) for feature extraction - Implements a recurrent neural network (LSTM) for emotion classification - Dataset sourced from the IEMOCAP database, a collection of 10 emotions with a focus on six major emotions: angry, neutral, frustrated, sad, excited, and happy - Applies bucket padding and batch shuffling for data consistency and batch generation - Uses two layers of Bidirectional LSTM followed by attention in the last layer - Dropout regularization to mitigate overfitting - Tensorflow implementation of the model available for both CPU and GPU Tech stack: Python, TensorFlow, Jupyter Notebook Why it matters: This project demonstrates the application of deep learning techniques to emotion recognition from speech, providing a valuable tool for understanding and analyzing human emotions through audio data.

Overview: at16k is a Python library designed for automatic speech recognition (ASR). It provides a simple and efficient way to perform speech-to-text conversion, making it accessible for developers to quickly build applications that require this functionality. Key features: - Supports multiple models for ASR, including English models trained at 8 KHz and 16 KHz. - Includes a command-line utility for easy experimentation with different models and use cases. - Provides a library API for developers to integrate at16k into their projects. Tech stack: Python Why it matters: at16k simplifies the process of integrating speech-to-text capabilities into applications, making it a valuable tool for developers working on projects that require speech recognition. Repository: at16k GitHub stars: 132 GitHub description: Trained models for automatic speech recognition (ASR). A library to quickly build applications that require speech to text conversion.

Overview: This project is a demo ChatBot application built using Amazon Bedrock's KnowledgeBase, Agent, and other AWS's serverless GenAI solutions. Key features: - Amazon Bedrock's KnowledgeBase integration - Amazon Bedrock's Agent for multi-step task execution - Amazon OpenSearch Serverless for vector database - AWS Lambda for serverless compute - Amazon S3 for storage - Amazon Athena for efficient query execution - Elastic Load Balancer Tech stack: Python Why it matters: This application showcases Amazon Bedrock's capabilities to convert natural language into Amazon Athena queries and to process complex data sets, demonstrating its utility in integrating various AWS services for serverless GenAI solutions.

Overview: HarmonyView is an official PyTorch implementation of the paper "HarmonyView: Harmonizing Consistency and Diversity in One-Image-to-3D" presented at CVPR 2024. This repository provides codes for improving the Syncdreamer model with a novel diffusion sampling technique, enabling the generation of realistic 3D content from a single image. Key features: - Improved Syncdreamer model with novel diffusion sampling technique - Generates realistic 3D content from a single image - Enhances visual and geometric consistency across generated views - Increases diversity of novel views, even in complex scenes Tech stack: Python Why it matters: HarmonyView offers a significant advancement in the field of image-to-3D rendering by harmonizing consistency and diversity, making it a valuable tool for researchers and developers looking to create more realistic and varied 3D models from a single image. Repository: Byeongjun-park/HarmonyView GitHub stars: 132

Overview: RagBase is a completely local RAG system that integrates with an open LLM, providing a chat interface for documents using LangChain, Streamlit, Ollama (Llama 3.1), Qdrant, and advanced techniques like reranking and semantic chunking. Key features: - Local RAG with a chat interface - Integration with an open LLM - Uses LangChain, Streamlit, Ollama (Llama 3.1), Qdrant, and advanced methods like reranking and semantic chunking Tech stack: Python Why it matters: RagBase offers a seamless integration of local RAG with an open LLM, allowing for efficient and personalized document-based conversations, making it ideal for developers and researchers looking to enhance their productivity and knowledge retrieval capabilities.

Overview: LLM-SRBench is a comprehensive benchmark designed to evaluate Large Language Models (LLMs) for scientific equation discovery and symbolic regression across four scientific domains. It includes two categories: LSR-Transform for reasoning beyond memorized forms and LSR-Synth for synthetic, discovery-driven problems. Key features: - 239 challenging problems across four scientific domains - LSR-Transform for common models to less common mathematical representations - LSR-Synth for synthetic, discovery-driven problems requiring data-driven reasoning - Comprehensive benchmark for evaluating LLMs in scientific equation discovery and symbolic regression Tech stack: Python Why it matters: LLM-SRBench provides a robust and diverse benchmark for assessing the capabilities of LLMs in scientific equation discovery and symbolic regression, ensuring that models can reason beyond memorized forms and tackle discovery-driven problems effectively.

Overview: Chronos Engine is a groundbreaking temporal simulation platform that empowers researchers, storytellers, game developers, and futurists to construct, explore, and manipulate dynamic timelines. Unlike conventional world simulation tools that focus on static spatial environments, Chronos Engine treats time as the primary interactive dimension. Key features: - Alternate History Modeling: Fork a timeline at significant historical turning points. - Forecasting & Scenario Planning: Input macro-economic indicators, climate variables, or population dynamics. - Generative Narrative Design: Automate plot weaving for writers and game masters. - Educational Sandboxing: Interactive observation of how small changes in initial conditions lead to vastly different historical outcomes. Tech stack: Temporal Graph Network (TGN) Why it matters: Chronos Engine provides a computational backbone to bring temporal tapestries to life, unlocking capabilities that static world simulators cannot approximate, making it perfect for researchers, storytellers, game developers, and futurists alike. Repository: grandgaming9321-prog/reality-engine Primary language: HTML GitHub stars: 132 GitHub description: Top Dynamic AI World Simulation & Storytelling Tools 2026

Overview: Piragi is a powerful and intuitive interface for querying and retrieving information from various sources, including PDFs, Word documents, URLs, and more. It supports local and remote storage, auto-updates, and advanced features like smart citations and semantic chunking. Key features: - Zero Config - Works with free local models out of the box. - All Formats - Supports PDF, Word, Excel, Markdown, Code, URLs, Images, Audio. - Remote Storage - Reads from S3, GCS, Azure, HDFS, and SFTP with glob patterns. - Web Crawling - Recursively crawls websites with `/**` syntax. - Auto-Updates - Background refresh, queries never blocked. - Smart Citations - Every answer includes sources. - Pluggable Stores - Supports LanceDB, PostgreSQL, Pinecone, Supabase, and custom. - Advanced Retrieval - HyDE, hybrid search, cross-encoder reranking. - Semantic Chunking - Context-aware and hierarchical chunking. - Knowledge Graph - Entity/relationship extraction for better answers. - Async Support - Non-blocking API for web frameworks. Tech stack: Python. Why it matters: Piragi simplifies the process of integrating various sources of information into a single, accessible interface, making it an ideal tool for developers looking to enhance their applications with dynamic and up-to-date content. Repository: hemanth/piragi Primary language: Python GitHub stars: 132 GitHub description: The best RAG interface yet.

### Overview: This project integrates Meta Rayban Glasses with a WhatsApp bot, utilizing Google Gemini, Redis for data management, Notion for note-taking, and Google Calendar for event and reminder management. It leverages Python 3.x and the `pip` package manager for development. ### Key Features: - **Meta Rayban Glasses Integration**: Enables the use of Meta Rayban Glasses with a WhatsApp bot. - **Google Gemini Integration**: Utilizes Google Gemini for image analysis and AI capabilities. - **Redis for Data Management**: Manages data using Redis for storing images and managing other data. - **Notion for Note-Taking**: Integrates with Notion for organizing and managing notes. - **Google Calendar for Event and Reminder Management**: Manages events and reminders using Google Calendar. ### Tech Stack: Python ### Why it Matters: This project provides a comprehensive solution for integrating Meta Rayban Glasses with a WhatsApp bot, enhancing user experience and functionality.

Overview: TinyGPT is a C++ implementation of a language model inference system, designed for fast and efficient processing of large language models like Llama 3.2, Qwen 2.5, Qwen 3, and Mistral. It features a fast BPE tokenizer, CUDA inference, and support for multiple data types and parallel processing. Key features: - Fast BPE tokenizer inspired by tiktoken - CUDA inference with FP32, FP16, and BF16 data types - Paged KV Cache with auto-sizing - Continuous Batching with chunked prefill - Flash Attention via TinyFA - Multi-GPU Tensor Parallel inference via NCCL Tech stack: C++ Why it matters: TinyGPT is a lightweight, efficient implementation of a language model inference system, making it suitable for real-time applications and systems with limited resources.