Projects

Overview: Coauthor is an English-to-LaTeX translator designed for use in Overleaf, converting natural language descriptions into LaTeX code. It serves as an indispensable tool for non-math majors, simplifying the process of writing mathematical content directly in Overleaf. Key features: - Converts English descriptions of formulae to LaTeX - Describes matrices in Matlab style - Translates theorems and equations directly - Automatically creates typeset tables - Serves as a lightweight Chrome extension Tech stack: JavaScript Why it matters: Coauthor streamlines the process of writing mathematical content in Overleaf, making it accessible and efficient for users who are not math majors. Repository: varunshenoy/coauthor Primary language: JavaScript GitHub stars: 118 GitHub description: Convert natural language to LaTeX within Overleaf using LLMs

### Overview: AlchemyLake for Databricks is a governed, provenance-sealed creative rendering solution that turns any data source into finished deliverables such as PowerPoint decks, PDFs, infographics, video briefings, podcasts, and scores. It integrates seamlessly with Databricks, ensuring data integrity and compliance. ### Key Features: - **Governed Creative Rendering**: Seamlessly transforms any governed source into finished deliverables. - **PowerPoint Decks**: Board-ready decks with read-aloud scripts and Q&A prep. - **PDF Dossiers**: Enterprise-level PDFs with statistical appendices and Excel evidence. - **Deep-Research Dossiers**: Multi-step investigations with governed SQL follow-ups. - **Infographics**: Visual representations of data trends and insights. - **Video Briefings**: Animated video briefings with narration in various formats. - **Data Podcasts**: Two-host, multi-format data podcasts. - **Sonified Scores**: Sonified scores in six genres, following growth patterns. ### Tech Stack: Python ### Why It Matters: AlchemyLake for Databricks ensures data integrity and compliance by rendering data into finished deliverables, making it a valuable tool for data-driven decision-making and presentation.

Overview: HALC is an official PyTorch implementation of the paper "HALC: Object Hallucination Reduction via Adaptive Focal-Contrast Decoding". It supports online object hallucination reduction using various decoders and post-hoc methods. Key features: - Supports online object hallucination reduction with decoders like Minigpt4-v2, Instructblip, LLaVA-1.5, mPLUG-OWL2, and post-hoc methods like Woodpecker and LURE. - Currently, it supports beam search for the HALC decoder. - Grounding DINO is used as the external detector to ground hallucinatory objects. Tech stack: Python Why it matters: HALC provides a comprehensive solution for reducing object hallucination in generative models, making it valuable for researchers and practitioners in the field of computer vision and artificial intelligence.

### Overview: This project introduces a novel framework for generating high-fidelity 4D content from single videos, specifically focusing on the generation of dynamic 3D animations. The approach combines the use of Gaussian Variation Fields (GVFs) and Direct 4D Mesh-to-GS Variation Field (DSVGF) Variational Autoencoders (VAEs) to efficiently represent and synthesize complex 3D animations. ### Key Features: - **Gaussian Variation Fields (GVFs)**: The project utilizes GVF-based VAEs to directly encode and compress high-dimensional animations into a compact latent space. - **Direct 4D Mesh-to-GS Variation Field VAE (DSVGF VAE)**: This VAE directly encodes canonical Gaussian Splats (GS) and their temporal variations from 3D animation data without per-instance fitting. - **Gaussian Variation Field Diffusion Model (GVFDiffusion)**: The model is trained with temporal-aware Diffusion Transformer conditioned on input videos and canonical GS, demonstrating superior generation quality compared to existing methods. - **Efficient Representation**: The use of GVF and DSVGF VAEs allows for efficient representation and synthesis of complex 3D animations, which is crucial for generating high-fidelity 4D content from single videos. ### Tech Stack: The project uses Jupyter Notebook as the primary language for development and experimentation. ### Why It Matters: This project represents a significant advancement in the field of video-to-4D synthesis by providing a more efficient and effective method for generating high-quality 4

Overview: This repository offers a suite of tools and examples for building AI applications using the Watsonx.ai Flows Engine, a powerful framework for creating AI agents that can be deployed in the cloud. Key features: - Build, run, and deploy AI tools from any data source - Integrations with LangGraph, LangChain, Watsonx.ai, and OpenAI - Examples of various AI applications, including chat apps, text-to-SQL agents, and YouTube transcription agents Tech stack: Python & JavaScript Why it matters: The Watsonx.ai Flows Engine simplifies the process of building AI applications, making it accessible for developers and enabling the creation of AI agents that can be easily integrated into various frameworks. Repository: IBM/wxflows GitHub stars: 117 GitHub description: Examples and tutorials for building AI applications with watsonx.ai Flows Engine README excerpt: 📝 [BLOGS](https://developer.ibm.com/search/?q=%22flows%20engine%22) | 📝 [DOCUMENTATION](https://wxflows.ibm.stepzen.com/docs) | 📝 [DISCORD](https://ibm.biz/wxflows-discord) | 📄 [FREE SIGNUP](https://ibm.biz/wxflows)

Overview: motion_generate_tools is a Blender addon designed to generate motion using the MDM: Human Motion Diffusion Model, offering a versatile tool for animators and developers. Key features: - Generate motion using MDM: Human Motion Diffusion Model - Compatible with SMPL armature - Supports text conditions for precise motion descriptions Tech stack: Python Why it matters: This addon provides a powerful and flexible solution for creating realistic motion sequences in Blender, making it an essential tool for both animators and developers working with motion capture data. Repository: UuuNyaa/blender_motion_generate_tools Primary language: Python GitHub stars: 117

### Overview: ForgeWM is an open-source project that trains a real-time, playable video world model (VWM) using a progressive causal pipeline on 8 GPUs. It is designed for keyboard, mouse, or gamepad control and offers 72 FPS performance on a single H20. ### Key Features: - **Progressive Causal Pipeline**: Four stages (bidirectional SFT → teacher-forced causal AR → consistency distillation → on-policy DMD) that transform the Matrix-Game 2 base into a real-time interactive student. - **Few-step Students**: ForgeWM-1, ForgeWM-2, and ForgeWM-4 models from one Stage-3 run per step budget, achieving up to 72 FPS on a single H20. - **Cross-Domain Transfer**: The same recipe ported to a gamepad-driven FPS by widening only the action interface. ### Tech Stack: Python ### Why It Matters: ForgeWM provides a reproducible and open-source solution for training video world models that can be controlled with keyboard, mouse, or gamepad inputs, achieving high performance on real-time applications.

Overview: This project introduces the Chrome Web Speech API, a feature that allows users to convert spoken words into text using their voice. It's a powerful tool for accessibility and voice recognition. Key features: - Voice to Text functionality - Live demo for testing - Support for Google Chrome version 25 or later Tech stack: JavaScript Why it matters: This API is crucial for developers looking to create applications that can understand and respond to spoken commands, enhancing user interaction and accessibility.

Overview: uttertype is a Python project designed for dictation using OpenAI's Whisper model, offering a simple and efficient solution for transcription. Key features: - Transcription using OpenAI's Whisper model - Support for both OpenAI's official API and local Whisper server - Hotkey functionality for easy integration into existing workflows Tech stack: Python Why it matters: This project simplifies dictation and transcription, making it accessible for developers and users alike.

### Overview: DocWire SDK is a C++20 data-processing infrastructure designed for deterministic, auditable, on-premise workflows. It supports 100+ formats, OCR, local AI, and opt-in cloud AI pipelines. ### Key features: - Ingests files, email boxes, archives, HTTP request payloads, JSON, XML, CSV, and more. - Normalizes and structures raw input using format detection, parsers, OCR, document element extraction, and structured-data handling. - Transforms data with filters, translations, summaries, classification, entity extraction, sentiment analysis, embeddings, and optional local or cloud AI steps. - Export or act through plain text, HTML, CSV, metadata, stdout, files, HTTP POST, or the built-in HTTP server. - Composes pipelines with `operator|` and embeds them directly into your own C++20 application. - Runs on-premise by default; cloud processing is used only when you explicitly add a cloud pipeline step. ### Tech stack: C++ ### Why it matters: DocWire SDK is built for production environments, ensuring predictable execution, memory ownership, and auditability. It supports complex data processing tasks without compromising on computational speed, making it ideal for applications in healthcare, compliance, trading, and autonomous systems.

### Overview: This repository is a curated list of 152 open-source tools designed to run, fine-tune, and build with Large Language Models (LLMs) locally, without relying on cloud services, API keys, or censorship. The ultimate goal is to provide a comprehensive and up-to-date toolkit for developers and enthusiasts to create local AI applications. ### Key features: - **One-Click Runners & Installers**: Tools like Ollama, LM Studio, and Llama.cpp allow users to run and interact with various LLMs directly from their local machines. - **Desktop & Web UIs**: Features such as Open WebUI, LobeChat, and Gradio provide user-friendly interfaces for local model interaction. - **Agent Frameworks**: Tools like CrewAI and AutoGen enable the creation of multi-agent systems, enhancing the capabilities of LLMs in complex environments. ### Tech stack: The repository is built using C++ and Python, leveraging the GGUF format for high-performance inference and the GGUF format for efficient model compression. ### Why it matters: This repository is crucial for developers and researchers who need to experiment with LLMs locally, ensuring they have the necessary tools to fine-tune models, build AI applications, and collaborate with other local AI enthusiasts.

Overview: 这个项目是一个基于魔塔社区Paraformer模型的中文语音识别工具，支持将音频和视频中的中文语音转换为SRT字幕。 Key features: - 中文语音识别 - 支持音频和视频转为SRT字幕 - 提供API接口和简单界面 Tech stack: Python Why it matters: 该项目填补了国外模型在中文支持方面的不足，为用户提供了一个便捷的中文语音识别解决方案。

### Overview: Alpaca-RLHF is a project that fine-tunes LLaMA using Reinforcement Learning with Human Feedback (RLHF). The repository leverages DeepSpeed Chat for efficient training and includes modifications to the original DeepSpeed Chat codebase to improve performance and accuracy. ### Key Features: - **Special Tokens**: Added to enhance the model's ability to handle human feedback. - **Rewarded Fitting**: Improved the training process by rewarding correct responses. - **Multi-Turn Alpaca Dataset**: Enhanced the dataset for better performance. - **Step-by-Step Training**: Includes modifications to three steps of the training process, including step 1, 2, and 3. ### Tech Stack: - Python ### Why It Matters: Alpaca-RLHF is a significant improvement over traditional RLHF methods, offering better performance and accuracy through the incorporation of special tokens, rewarded fitting, and modifications to the training process.

### Overview: Stable Diffusion Segmentation (SDSeg) is an official implementation of the paper "Stable Diffusion Segmentation for Biomedical Images with Single-step Reverse Process" presented at the 2024 MICCAI conference. This project aims to provide a comprehensive and efficient solution for biomedical image segmentation using Stable Diffusion, a powerful generative model. ### Key Features: - **Stable Diffusion Model Integration**: Utilizes Stable Diffusion to enhance image segmentation capabilities. - **Single-Step Reverse Process**: Implements a novel single-step reverse process to improve segmentation accuracy. - **Efficient Training**: Optimized for faster training speeds by reducing unnecessary logging. ### Tech Stack: - Python ### Why It Matters: SDSeg offers a significant improvement in biomedical image segmentation by leveraging the strengths of Stable Diffusion, leading to better accuracy and efficiency. This project is valuable for researchers and practitioners in the field of medical imaging, providing a practical tool for enhancing image analysis and diagnosis. ### Repository: - **Repository**: `lin-tianyu/Stable-Diffusion-Seg` - **Stars**: 117 - **Description**: Official implementation of "Stable Diffusion Segmentation for Biomedical Images with Single-step Reverse Process" at MICCAI 2024. ### Contributors: - **Tianyu Lin**: Main contributor - **Zhiguang Chen**: Additional contributor - **Zhonghao Yan**: Additional contributor - **Weijiang Yu**: Additional contributor - **Fudan Zheng**: Additional contributor ### News:

### Overview: Harfiyen, tarayıcıda çalışan Türkçe konuşma tanıma ve altyazı aracıdır. Otomatik noktalama, canlı transkript ve altyazılı video dışa aktarma özelliği sunar. Tamamen yerel, WebGPU destekli. ### Key features: - **Türkçe konuşmayı tarayıcıda yazıya döker, altyazıyı videonun içine yazar.** - **Seda-v0.2 Türkçe konuşma tanıma modelini doğrudan tarayıcıda, ekran kartı üzerinde (WebGPU) çalıştırır.** - **Canlı mikrofon: kelimeler konuştuktan yaklaşık bir saniye sonra belirir, kayıt aynı anda saklanır.** - **Döküm durdurulunca o ana kadar yazılan proje olarak kalır ve sonra kaldığı yerden sürdürülebilir; sayfa yenilense de kalınan yer unutulmaz.** - **Canlı mikrofon: kelimeler konuştuktan yaklaşık bir saniye sonra belirir, kayıt aynı anda saklanır.** - **Canlı mikrofon: kelimeler konuştuktan yaklaşık bir saniye sonra belirir, kayıt aynı anda saklanır.** - **Döküm sürerken önizleme yazılanı takip eder: görüntü ve altyazı o an yazılan yeri gösterir. Oyn

### Overview: TorchEBM is a simulation-free, GPU-first library for generative modeling in PyTorch, designed for scalable and stable training of energy-based models, diffusion, flow matching, and equilibrium and energy matching. ### Key Features: - Composable primitives for EBMs, diffusion, flow matching, and equilibrium/energy matching. - GPU-first training for efficient and stable models. - Simpler, more efficient training compared to traditional methods. - Consistent and stable training across different datasets and models. ### Tech Stack: Python ### Why It Matters: TorchEBM simplifies and accelerates the training of complex generative models, making it easier and faster to develop and deploy high-quality generative AI applications.

Overview: Cozy Auto Texture is a Blender add-on designed to simplify the process of generating AI textures using Stable Diffusion, making it accessible even for beginners. Key features: - Easy setup without requiring admin privileges or advanced coding knowledge - Open-source software - Simple download and installation process - Free weights hosted by This Cozy Studio Inc. - Compatible with multiple Blender versions - Intuitive user interface - Ethical use of Stable Diffusion weights Tech stack: Python Why it matters: Cozy Auto Texture provides a user-friendly way to generate textures using AI, making it accessible to a wide range of users without the need for complex installations or extensive knowledge of programming.

### Overview: The VideoDB Python SDK is a powerful tool designed to help developers integrate video processing capabilities into their AI applications. It provides a robust API for semantic search, scene extraction, transcript generation, and more, making it ideal for applications requiring structured video data. ### Key Features: - **Semantic Search**: Facilitates search and retrieval of videos based on semantic content. - **Scene Extraction**: Extracts scenes from videos for detailed analysis. - **Transcript Generation**: Automatically generates transcripts from video content. - **Multimodal Content Generation**: Generates comprehensive multimedia content for enhanced understanding. ### Tech Stack: The SDK is built on Python, utilizing libraries such as `requests`, `backoff`, and `tqdm` for its functionality. ### Why It Matters: This SDK simplifies the integration of advanced video processing capabilities into AI applications, enabling more intelligent and interactive applications that can understand and process video data as structured data.

### Overview: Bring [**ChatGPT Images 2.0**](https://openai.com/index/introducing-chatgpt-images-2-0/) (`gpt-image-2`) to [OpenCode](https://opencode.ai) using your ChatGPT subscription (no API costs). ### Key features: - **Subscription-friendly**: Generations ride on the same Codex backend channel OpenCode already uses for ChatGPT subscription chat — billed against your existing ChatGPT Plus, Pro, or Business plan, not your API credits. - **Reference images**: Pass up to five input images alongside the prompt for style guidance, edit targets, or compositing inputs. - **Transparent backgrounds**: Ask for a transparent background, a cutout, or background removal, and the agent passes `background: "transparent"` to get an RGBA PNG. The cutout edges can be rough, with a faint semi-transparent halo. For a clean cutout, generate the image on a plain background and remove the background yourself. ### Tech stack: TypeScript ### Why it matters: This plugin allows you to leverage the powerful image generation capabilities of ChatGPT through your existing ChatGPT subscription, without the need for API costs.

Overview: German-OCR is a high-performance German document Optical Character Recognition (OCR) tool designed to extract text from invoices, receipts, forms, and other business documents. Key features: - Vision-OCR: Extracts text from German documents, including invoices and receipts, into structured JSON format. - Text-Assistant: Translates, summarizes, rewrites, and chats with text from German documents. - Supports local and cloud backend options. Tech stack: Python Why it matters: German-OCR is optimized for real German invoices, ensuring zero hallucination and generating 100% valid JSON output, making it suitable for businesses requiring accurate document processing. Repository: Keyvanhardani/german-ocr GitHub stars: 116 GitHub description: German-OCR is specifically trained to extract text from German documents including invoices, receipts, forms, and other business documents. README excerpt: German-OCR is a high-performance German document Optical Character Recognition (OCR) tool designed to extract text from invoices, receipts, forms, and other business documents.