Projects

### Overview: OpenReshot 是一个原生 iOS 应用程序，旨在为 iPhone 和 iPad 用户提供一种独特的照片动效体验。通过拖动照片并调整视角，用户可以重新拍摄照片，从而获得全新的视角和构图。这款应用特别适用于那些希望体验到空间照片和可变视角照片新体验的用户，但系统版本或设备限制可能会暂时阻止他们使用。 ### Key Features: - **重新拍摄照片**：用户可以拖动照片，找到比原图更合适的视角和构图。 - **生成最终照片**：当前角度满意后，点击“重构”生成一张更完整、更清晰的照片。 - **空间画廊**：重建结果会缓存在本机，再次打开同一张照片可直接加载空间，无需重复推理。 ### Tech Stack: Swift ### Why it Matters: OpenReshot 将普通照片的拍摄体验提升到了一个全新的维度，为用户提供了更多的创意空间。通过简单的拖动操作，用户可以轻松地改变照片的视角和构图，创造出截然不同的视觉效果。这款应用不仅提升了照片的观赏性，还为摄影爱好者和创意工作者提供了更多的创作可能性。

### Overview: Exploring Diffusion Transformer Designs via Grafting aims to evaluate the impact of architectural decisions on model quality through the use of pretrained diffusion transformers. The project presents a novel approach called grafting, which allows researchers to study new architecture designs without the need for costly pretraining. ### Key Features: - **Grafting**: A method for editing pretrained diffusion transformers to create new architectures. - **Testbed Design**: Based on DiT-XL/2 design, a testbed is constructed to study the impact of grafting on model quality. - **Pretrained Models**: Utilizes pretrained diffusion transformers to evaluate architectural decisions. ### Tech Stack: Jupyter Notebook ### Why it Matters: This project provides a practical tool for architectural investigation by leveraging pretrained models, thus reducing the computational cost associated with pretraining. The grafting approach offers a simple yet effective way to explore and evaluate new architecture designs without the need for extensive pretraining, making it accessible and valuable for researchers in the field of deep learning.

Overview: This project is a 4K image preview API designed for Gemini 3 Pro, offering a cost-effective alternative to OpenAI's API. It supports both default and 4K pricing models. Key features: - 4K image generation and editing - Pricing: $0.03 per image, $0.04 per 4K image - Supports multiple image generation tasks - Real-time billing and refund upon failure - Asynchronous processing Tech stack: Python Why it matters: This API provides a cost-effective solution for generating high-resolution images, making it ideal for applications requiring detailed visual content without the need for a subscription-based service. Repository: nmkshs742wz1yrh/nanobananapro-nano-banana-pro-4k Primary language: Python GitHub stars: 77 GitHub description: Gemini-3-pro-image-preview API (nanobananapro / gemini 3 pro image preview): 4K $0.04 per image; default $0.03. Model id, 4k settings, curl and Python examples over an OpenAI-compatible endpoint.

Overview: This repository contains official code for the Manning book "Rearchitecting LLMs: Structural Techniques for Efficient Models," focusing on optimization techniques for transforming large pre-trained models into specialized Small Language Models (SLMs). The project aims to address the growing need for AI professionals with the skills to create highly efficient models tailored to specific data and tasks. Key features: - Depth pruning - Width pruning in GLU architectures - Knowledge distillation Tech stack: Jupyter Notebook Why it matters: The book teaches optimization techniques for transforming large pre-trained models into efficient Small Language Models, addressing the need for specialized models over generic LLMs. It provides hands-on notebooks that run on Google Colab's free tier, offering a practical approach to understanding and applying these techniques to open-source models like Llama, Gemma, and Qwen. Repository: peremartra/Rearchitecting-LLMs GitHub stars: 77

### Overview: kANNolo is a Rust library designed for approximate nearest neighbors (ANN) search, focusing on usability and performance. It supports both dense and sparse embeddings and implements the HNSW graph index and Product Quantization. ### Key features: - **Modular Design**: Easy to use and modify. - **Performance Optimization**: Built for maximum performance, including both dense and sparse indexes. - **Compatibility**: Seamless integration with both dense and sparse embeddings. - **Algorithm Support**: Implements the HNSW graph index and Product Quantization. ### Tech stack: Rust ### Why it matters: Designed for researchers and developers, kANNolo simplifies the prototyping of new ANN search algorithms and data structures, making it ideal for AI and machine learning applications requiring efficient nearest neighbor search.

### Overview: This repository is a comprehensive survey of decoding methods for large language models and large vision-language models, focusing on three primary paradigms: contrastive decoding, guided decoding, and parallel decoding. ### Key features: - 📜 Published in ACM SIGKDD Explorations Newsletter - 📧 Contact: haoran.wang@emory.edu - 📑 Compiles research on decoding methods for LLMs and LVLMs - 🌐 Links to related papers and tutorials ### Tech stack: Unknown ### Why it matters: This survey provides a valuable resource for researchers and practitioners looking to enhance the decoding methods of large language and vision-language models, ensuring more effective and efficient output generation.

Overview: MegaRAG is a cutting-edge multimodal graph-based retrieval augmented generation system that enables global visual question answering on documents. It combines graph-based reasoning with page retrieval techniques to provide precise and rich responses. Key features: - Multimodal Knowledge Graph (MMKG) - Graph-based reasoning for precise answers - Page retrieval for contextual understanding - Integration of visual information for comprehensive queries Tech stack: Python Why it matters: MegaRAG is a significant advancement in the field of document retrieval and generation, offering a robust solution for complex visual questions on documents. Its ability to integrate visual information and provide rich, contextually rich responses makes it highly valuable for applications requiring detailed and accurate information extraction. Repository: AI-Application-and-Integration-Lab/MegaRAG Primary language: Python GitHub stars: 74 GitHub description: [ACL 2026] MegaRAG: Multimodal Graph-based Retrieval Augmented Generation

### Overview: cc-VisionRouter 是透明代理，用于在 Claude Code 中安全地处理长程任务。它自动识别请求中的图片，并将图片请求分流到多模态模型，确保主力模型不会因图片而崩溃。此外，它还支持背景任务的多模态处理，确保 Claude Code 的后台请求也能安全通过。 ### Key features: 1. **图片请求分流**：自动识别请求中的图片，并将图片请求分流到多模态模型，确保主力模型不会因图片而崩溃。 2. **背景任务处理**：支持 Claude Code 的后台请求（如压缩、总结等），确保这些请求也能安全通过。 3. **模型自动切换**：代理会自动将 Claude Code 的模型名替换成实际的模型 ID，无需手动配置。 4. **日志和状态查看**：提供交互式菜单、配置选项、启动和停止代理功能，以及查看状态和日志的命令。 5. **语言自动跟随**：界面会根据系统 locale 自动切换语言，支持环境变量和交互菜单切换。 ### Tech stack: JavaScript ### Why it matters: 让不支持图片的主力模型也能在 Claude Code 中安全跑长程任务，避免因图片请求导致的崩溃，从而保证长任务的顺利进行。

### Overview: FSS is a pure-Rust, evidence-native, local-first sensor fabric designed for owner-authorized cameras and capture drones. It offers a unified system for recording, processing, and analyzing video streams, providing a calibrated world model and event transaction capabilities. ### Key Features: - **Packet Truth**: Tracks packet integrity and source continuity. - **Authority**: Ensures source custody and operator control. - **Cognition**: Models and graphs for event detection and tracking. - **Effect**: Provides idempotent effects and observed receipts. ### Tech Stack: Pure-Rust implementation with evidence-native processing. ### Why It Matters: FSS transforms heterogeneous sensors into a cohesive system, providing a calibrated world model and event transaction capabilities. It addresses the challenges of consumer security hardware by integrating a unified system for recording, processing, and analyzing video streams, ensuring a robust and reliable security infrastructure.

Overview: This is a Stable Diffusion plugin for Krita that integrates with Automatic1111, SD.Next, and Forge WebUIs, allowing users to generate and manipulate images using Stable Diffusion models. The plugin is designed to be lightweight and easy to use, with a focus on compatibility and performance. Key features: - Automatic1111 integration - SD.Next compatibility - Forge WebUI support - Support for Soft Inpainting - Max size option for large canvas - ControlNet compatibility improvements - UI bug fixes - Batch controls and Simplify UI - RemBG, ControlNet Preprocessor Previews, Upscaler, and Inpainting support Tech stack: Python Why it matters: This plugin provides a seamless interface for creating and modifying images using Stable Diffusion models, making it useful for artists and designers who prefer to work within Krita's environment. The ability to integrate with multiple web UIs and support for various features ensures versatility and compatibility with different tools and environments.

Overview: This repository contains sample codes for a Generative AI Language Workshop, focusing on integrating Langchain with Google Cloud's Generative AI Language offerings. Key features: - Langchain integration - Google Cloud resources provisioning - Vertex AI Workbench setup Tech stack: Python Why it matters: This workshop provides practical, hands-on experience in integrating Generative AI with Langchain and leveraging Google Cloud's Vertex AI Workbench for a seamless workshop environment.

### Overview: AI Megalist is a curated index of 200+ AI tools, organized by the job you want to do, not the lab that built the model. It offers a quick find feature, eight ready-made stacks tailored to engineers, researchers, marketers, creators, designers, founders, students, and knowledge workers, and a comprehensive browse of 225+ tools across five clusters and 22 categories. ### Key features: - **Quick find**: One-line "I want to..." → category map. - **Pick your role**: Eight ready-made stacks. - **Browse the index**: 225+ tools across five clusters and 22 categories. ### Tech stack: - JavaScript, TypeScript, React, and Markdown. ### Why it matters: AI Megalist provides a practical and user-friendly way to find the right AI tool for your specific needs, ensuring you can focus on your job without worrying about the technical details.

### Overview: DDPS is an innovative approach to semantic segmentation that leverages a recently developed denoising diffusion generative model to enhance the quality of existing discriminative methods. This project aims to improve the semantic segmentation performance on datasets like ADE20K and Cityscapes. ### Key Features: - **DDPS-DeeplabV3+**: Implements a MobilenetV2 backbone with 27.9M parameters, achieving a mean IoU of 38.07%. - **DDPS-Segformer**: Utilizes a MiT-B0 backbone with 15.1M parameters, offering a mean IoU of 41.67%. - **DDPS-DeeplabV3+**: Features a ResNet50 backbone with 54.2M parameters, delivering a mean IoU of 46.32%. - **DDPS-Segformer**: Implements a MiT-B2 backbone with 65.5M parameters, achieving a mean IoU of 49.73%. ### Tech Stack: The project is implemented in Python, utilizing the Hugging Face framework. ### Why It Matters: DDPS introduces a novel method to improve semantic segmentation quality by incorporating a mask prior modeled through a denoising diffusion generative model. This approach aims to enhance the accuracy and robustness of existing segmentation techniques, making it a significant contribution to the field of computer vision. ### Repository: The project is hosted on GitHub at [OpenGVLab/DDPS](https://

### Overview: You can reproduce the experiments of our paper "Don’t Stop Pretraining? Make Prompt-based Fine-tuning Powerful Learner" [arXiv:2305.01711](https://arxiv.org/pdf/2305.01711.pdf). ### Key features: - Implementation of the Prompt-based Continued Pre-training (PCP) method. - Implementation of state-of-the-art prompt-based fine-tuning approaches, including hard and soft variants. - Implementation of self-training methods based on hard prompt-based fine-tuning approaches. ### Tech stack: Python ### Why it matters: Our approach aims to improve the performance of prompt-based fine-tuning by presenting both task-related texts and prompt templates to language models through unsupervised pre-training objectives before fine-tuning for the target task. This method consistently improves the performance of state-of-the-art prompt-based fine-tuning approaches across various benchmarks, even with only hundreds of unlabelled examples.

Overview: This repository contains various installation guides for Large Language Models, providing essential resources for fine-tuning and inference using Jupyter Notebooks and Google Colab. Key features: - Basic installation guide using Jupyter or Google Colab - Intermediate scripts for model parallel, distributed data parallel, and Fully Sharded Data Parallel training - Guide on pushing models to HuggingFace Hub - 8-bit quantization notebook - LLM comparison notebook - Running LLM on laptop using llama.cpp - Setting up jupyter lab on laptop Tech stack: Jupyter Notebook Why it matters: This repository serves as a comprehensive guide for developers looking to install and fine-tune Large Language Models, making it a valuable resource for anyone interested in advancing their LLM skills.

Overview: AIHTTPAnalyzer is an AI-powered security analysis assistant that integrates with Burp Suite, enhancing web application security testing capabilities. It automates vulnerability detection and provides intelligent analysis, including real-time vulnerability assessments and PoC examples. Key features: - 🔍 Analyze HTTP requests and responses for security vulnerabilities - 🛠️ Provide technical analysis and exploitation steps - 📄 Include PoC examples and payloads - 🖥️ Integrate with Burp Suite's UI and context menu - 🚀 Real-time vulnerability assessments - 🤖 AI-powered context-aware analysis - 🎯 Generate Proof-of-Concept exploits - 📝 Custom PoC script generation - 🔧 Payload customization for specific scenarios Tech stack: Java Why it matters: AIHTTPAnalyzer revolutionizes web application security testing by leveraging artificial intelligence to automate vulnerability detection and provide intelligent analysis, making it easier for security professionals to identify and address complex security issues efficiently.

### Overview: This project provides a simple, self-contained codebase for training Text to Image Diffusion models using Keras, specifically focusing on the training of CLIP conditioned models. The repository includes a notebook for training a 128x128 Latent Diffusion model on the Kaggle kernel hardware, demonstrating improved performance and resolution compared to traditional methods. ### Key Features: - **Training a CLIP Conditioned Text to Image Diffusion Model**: The codebase allows for training models conditioned on CLIP embeddings, achieving better results with lower computational resources. - **Latent Diffusion Models**: Utilizes a latent diffusion model architecture to generate high-resolution images from textual prompts. - **Cross-Attention and Self-Attention**: Incorporates additional text/CLIP/masking embeddings and cross/self attention mechanisms for improved performance. - **Training on Kaggle Hardware**: The notebook demonstrates training a 128x128 Latent Diffusion model on the Kaggle kernel hardware, showcasing faster training times and improved resolution. ### Tech Stack: - **Python**: The primary language used for implementing the codebase. - **Keras**: Framework used for building and training the neural network models. ### Why it Matters: - **Simplicity and Efficiency**: The project provides a straightforward and efficient way to train Text to Image Diffusion models, making it accessible for developers and researchers without extensive computational resources. - **Improved Results**: The code achieves better performance and resolution compared to traditional methods, making it a valuable resource for anyone interested in the field of image generation and diffusion models

Overview: This repository serves as a meta repository for the official ChatBotKit Node SDK, providing comprehensive tools and packages for developers working with ChatBotKit. It includes SDKs for Node.js, React, Next.js, and more, along with a CLI and various widget types and utilities. Key features: - ChatBotKit CLI - ChatBotKit API SDK - ChatBotKit Agent SDK - ChatBotKit React SDK - ChatBotKit Next.js SDK - ChatBotKit NextAuth.js SDK - ChatBotKit Widget types and utilities Tech stack: JavaScript, TypeScript, React, Next.js, and more. Why it matters: This repository streamlines the development process for ChatBotKit applications, providing developers with the necessary tools to create and deploy conversational AI solutions with custom data and integrations.

### Overview: NinjaChefs + AI is a web application that leverages generative AI to simplify online recipe sharing, making it accessible and user-friendly. ### Key Features: - **AI Recipe Assistance**: Automatically generates introductions, descriptions, ingredient lists, and cover images for recipes. - **Public Recipe Viewing**: Allows users to view and search shared recipes. - **Recipe Sharing**: Enables users to add recipes with AI assistance, including writing introductions, ingredient lists, and cover images. - **Community Safety**: Features an AI-powered spam filter and insights on allergens and health implications. ### Tech Stack: React, Tailwind, Node.js with Express, MongoDB, CloudFlare Images, Clerk, OpenAI Platform SDK, Cloudflare Workers. ### Why it Matters: NinjaChefs + AI aims to democratize online recipe sharing by removing the tedious tasks associated with writing detailed recipes, making it easier for everyone to share their culinary creations.

### Overview: RoomFuser is an acoustic impulse response generation project that uses diffusion models to efficiently simulate room acoustics. The project explores the use of Denoising Diffusion Probabilistic Models (DDPMs) to generate RIRs, which are crucial for audio applications such as room acoustics and audio processing. ### Key Features: - **Model Details:** - **DiffWave Model:** Uses a vector of size 10 for conditioning, including room dimensions, source position, receiver position, and RT60. - **Unet1D Model:** Introduces a mean prior on the initial noise signal to generate RIRs, using a low-order RIR computed via the Image Source Method (ISM) as a mean. - **Feature Extractors:** - Time Domain - Time Domain (Scaled using Scikit-learn) - Frequency Domain (Transfer Function) - **Datasets:** - Original dataset from FAST-RIR - Custom dataset generated using the GPU-RIR library ### Tech Stack: - **Programming Languages:** Jupyter Notebook ### Why It Matters: The project showcases the potential of diffusion models in generating high-quality RIRs, offering a more efficient alternative to computationally expensive methods like the Image Source Method (ISM). The ability to visualize the diffusion process adds an educational value, making it a valuable resource for researchers and developers in the field of audio processing and acoustics.