Projects

Overview: ConceptLab is an official implementation of a paper that focuses on creative text-to-image generation using diffusion prior models and question-answering vision-language models. The project aims to generate new, unique concepts that have never been seen before, challenging the limitations of existing generative models. Key features: - Creative text-to-image generation - Diffusion prior models - Question-answering vision-language models - Adaptive negatives scheme - Evolutionary generation process Tech stack: Python Why it matters: ConceptLab introduces a novel approach to creative text-to-image generation by leveraging diffusion prior models and question-answering vision-language models, offering a fresh perspective on generating unique concepts that are not similar to existing examples. The project's ability to evolve concepts through an evolutionary generation process adds significant value by providing a method to create new and diverse creations iteratively. Repository: kfirgoldberg/ConceptLab GitHub stars: 255 GitHub description: Official Implementation for "ConceptLab: Creative Generation using Diffusion Prior Constraints"

Overview: MLTU is a comprehensive set of machine learning training utilities designed for TensorFlow 2.* and PyTorch, offering a suite of tutorials and examples to help developers get started with machine learning. Key features: - Text Recognition with TensorFlow and CTC network - TensorFlow OCR model for reading Captchas - Handwriting words recognition with TensorFlow - Handwritten sentence recognition with TensorFlow - Introduction to speech recognition with TensorFlow - PyTorch introduction - Using custom wrapper to simplify PyTorch models training pipeline - Handwriting words recognition with PyTorch - Transformer training with TensorFlow for Translation task - Speech Recognition in Python - YOLOv8: Real-Time Object Detection Simplified - YOLOv8: Customizing Object Detector training Tech stack: Python Why it matters: MLTU provides a valuable resource for developers looking to quickly get started with machine learning using TensorFlow and PyTorch, offering a series of practical tutorials and examples that cover a wide range of applications, from text recognition to object detection, making it an essential tool for developers in the field of machine learning.

### Overview: twelvet 是一个基于 Spring Boot 3.x 的微服务框架，采用 Spring Cloud Alibaba 和 Spring Cloud Tencent 的结合，旨在简化开发过程并提高业务代码的通用性。 ### Key Features: - 独立的前端框架 `twelvet-ui` - 公共依赖模块 `twelvet` - `twelvet-api` 包含系统、数据处理、定时任务和人工智能接口 - `twelvet-framework-core` 是核心模块，提供微服务的基础框架 - 适用于多种应用场景，如网站后台管理、会员中心、内容管理系统、客户关系管理系统和办公自动化系统 ### Tech Stack: Java 17 + Spring Boot 3.x + Spring Cloud Alibaba + Spring Cloud Tencent ### Why it Matters: 通过学习和理解微服务概念，开发者可以快速掌握微服务的核心基础知识。此外，该项目还旨在减少业务代码的重复，并提供通用的业务代码，适用于微服务和单体应用。同时，它也为学习微服务和开发提供了实践机会，可以用于网站管理后台、会员中心、内容管理系统、客户关系管理系统和办公自动化系统等开发项目。此外，它还提供轻量级、高度可移植的插件，以实现快速原型开发。后端源码位于 `https://github.com/twelvet-projects/twelvet`，前端源码位于 `https://github.com/twelvet-s/twelvet-ui`，技术文档位于 `https://doc.twelvet.cn/`，官方

### Overview: ZombieBuddy is a powerful Java agent framework designed for Project Zomboid modders. It enables runtime bytecode manipulation and patching of game classes using ByteBuddy, providing a clean, annotation-based API for modding without requiring source code access. ### Key Features: - **Annotation-based patching**: Use `@Patch` annotations to declare method patches. - **Runtime class transformation**: Patch classes that are already loaded using retransformation. - **Automatic patch discovery**: Scans for patch classes automatically. - **Lua integration**: Expose Java classes and global functions to Lua. - **Advice and Method Delegation**: Support for both advice-based and delegation-based patching. - **Verbose logging**: Configurable verbosity levels for debugging. ### Tech Stack: ZombieBuddy is built on top of ByteBuddy, a Java agent framework for bytecode manipulation. ### Why It Matters: ZombieBuddy simplifies the process of creating mods for Project Zomboid by eliminating the need for manual file replacement and provides a clean, annotation-based API for modding without requiring access to the game's source code. This makes it easier to create mods and ensures that game updates are less likely to break your mod. ### Repository: The repository is hosted on GitHub at zed-0xff/ZombieBuddy, with 249 GitHub stars. The primary language used is Java. ### GitHub Description: ZombieBuddy is a Java agent framework that enables runtime bytecode manipulation and patching of game classes using ByteBuddy. It provides a clean, annotation-based API for intercepting and modifying

### Overview: CamLiFlow and CamLiRAFT are PyTorch implementations of two recent papers on optical flow and scene flow estimation. These models utilize bidirectional camera-LiDAR fusion to achieve superior performance on various datasets, including FlyingThings3D, KITTI, and MPI Sintel. ### Key Features: - **CamLiFlow**: A bidirectional camera-LiDAR fusion method for joint optical flow and scene flow estimation. - **CamLiRAFT**: An extended version of CamLiFlow that incorporates a new type of bidirectional fusion pipeline called CamLiRAFT, achieving significant improvements over the original method. ### Tech Stack: - **Python**: The codebase is written in Python, leveraging PyTorch for deep learning. ### Why It Matters: - **Performance Improvement**: CamLiRAFT significantly outperforms existing stereo and LiDAR-only scene flow methods across various datasets, demonstrating its effectiveness in handling complex and non-rigid motion scenarios. - **Generalization**: The method shows good generalization performance, reducing errors on non-rigid motion tasks. - **Training Efficiency**: The simplified training schedule of CamLiRAFT allows for faster training times compared to the original CamLiFlow, making it more practical for real-world applications. ### Repository: - **Repository**: MCG-NJU/CamLiFlow - **GitHub Stars**: 254 - **Primary Language**: Python - **GitHub Description**: Official PyTorch implementation for two papers on optical flow and scene flow estimation. ### Changes to the Conference Paper: - **Comparison with Stereo Scene Flow

**Overview:** SEPIA is a self-hosted, extendable, personal, intelligent assistant framework built with a modular, open-source architecture. It includes speech recognition (STT), wake-word detection, text-to-speech (TTS), natural-language understanding, dialog-management, a cross-platform client app, and more. The framework is lightweight and optimized for Linux, Windows, and Mac, and runs smoothly on a Raspberry Pi. **Key features:** - Speech Recognition (STT) - Wake-Word Detection - Text-to-Speech (TTS) - Natural-Language Understanding - Dialog-Management - Cross-Platform Client App - Lightweight Java Server and Elasticsearch DB - SEPIA Open Assistant - SEPIA Client - SEPIA STT Server - SEPIA SDK - SEPIA Wiki - SEPIA Twitter Feed - SEPIA Mastodon Feed - SEPIA Blog **Tech stack:** Java, Elasticsearch, HTML, JavaScript, Docker, Docker Hub, Linux, Windows, Mac, Raspberry Pi. **Why it matters:** SEPIA is a powerful, open-source framework that allows developers to create their own, fully customizable digital voice assistants. It includes a wide range of features such as news, music, timers, alarms, reminders, to-do and shopping lists, smart home, navigation, places, weather, Wikipedia, web-search, soccer results, and more. The framework is designed to be highly customizable and can be used to build a variety of smart devices, from smart speakers to mobile assistants.

### Overview: EPNet: Enhancing Point Features with Image Semantics for 3D Object Detection (ECCV 2020) is a state-of-the-art method for 3D object detection that leverages point cloud and image semantic information to improve detection accuracy. This project introduces a novel fusion module for point-wise feature enhancement without requiring additional image annotations, resulting in significant improvements over existing methods. ### Key Features: - **Multi-Scale Point-wise Fusion**: Enhances point features with semantic image features in a point-wise manner. - **Consistency Loss**: Encourages the consistency between localization and classification confidence. - **End-to-End Learning**: Integrates localization and classification confidence using an end-to-end trainable framework. - **No Extra Image Annotations**: Achieves accurate multi-scale point-wise fusion without requiring additional image annotations. ### Tech Stack: - **Python**: The primary programming language used for implementing the project. ### Why It Matters: EPNet significantly improves the accuracy of 3D object detection tasks by leveraging point cloud and image semantic information, making it a valuable contribution to the field of computer vision and robotics. ### Repository: - **Repository**: `happinesslz/EPNet` - **GitHub Stars**: 254 - **GitHub Description**: EPNet: Enhancing Point Features with Image Semantics for 3D Object Detection (ECCV 2020) - **Primary Language**: Python ### Installation: - **Environment Requirements**: - **Linux (tested on Ubuntu 16.04)** -

Overview: DiffBEV is an official PyTorch implementation for a conditional diffusion probability model specifically designed for Bird's Eye View (BEV) perception in autonomous driving. It aims to improve the quality of BEV representations by utilizing diffusion models to denoise noisy samples and refine semantic features. Key features: - Conditional diffusion model for BEV perception - Training with three types of conditions - Cross-attention module for context fusion - 25.9% mIoU improvement on nuScenes dataset compared to the best-performing existing approach Tech stack: Python Why it matters: DiffBEV leverages the denoising capabilities of diffusion models to enhance the quality of BEV representations in autonomous driving, leading to improved performance in semantic segmentation and 3D object detection tasks. Repository: JiayuZou2020/DiffBEV GitHub stars: 253 GitHub description: Official PyTorch implementation for a conditional diffusion probability model in BEV perception GitHub URL: https://github.com/JiayuZou2020/DiffBEV

### Overview: MixEval is an official evaluation suite and dynamic data release for MixEval, designed to evaluate Large Language Models (LLMs) with a high correlation and low cost. It supports both free-form and multiple-choice splits and is updated every month to ensure stability and efficiency. ### Key Features: - **Dynamic Benchmarks**: MixEval consists of two benchmarks, `MixEval` and `MixEval-Hard`, with two splits each. - **High Correlation**: Achieves a 0.96 correlation with Chatbot Arena, making it highly effective for benchmarking. - **Cost Efficiency**: Runs 6% faster and cheaper than MMLU, with queries updated monthly to avoid contamination. - **Local Evaluation**: Supports local model parsers and evaluating local checkpoints, making it practical for various environments. ### Tech Stack: - Python ### Why It Matters: MixEval provides a reliable and efficient benchmark for evaluating LLMs, ensuring accurate performance metrics without the need for expensive crowdsourced evaluations.

Overview: ScholarCopilot is an intelligent academic writing assistant designed to enhance research writing by integrating AI-powered text completion and citation suggestions. Built by TIGER-Lab, it aims to streamline the writing process while maintaining high scholarly standards. Key features: - Next-3-Sentence Suggestions - Full Section Auto-Completion - Context-Aware Writing - Real-time Citation Suggestions - One-Click Citation Insertion - Citation Bibtex Generation Tech stack: Python Why it matters: ScholarCopilot is a valuable tool for researchers looking to improve their writing efficiency and ensure academic integrity. By integrating AI with academic writing, it helps in generating coherent and contextually relevant text, as well as managing citations accurately, thereby streamlining the research process. Repository: TIGER-AI-Lab/ScholarCopilot GitHub stars: 253 GitHub description: ScholarCopilot: Training Large Language Models for Academic Writing with Accurate Citations [COLM 2025]

### Overview: KADATH is an evolutionary multi-agent runtime designed to optimize agents for a given goal through repeated competition and selection across multiple epochs. It uses a kernel that manages the run, benchmark, deadlines, and database, while organisms perform the goal with an editable agent framework. ### Key Features: - **Goal Optimization**: Evolves agents to achieve a specific goal through multiple epochs. - **Reproducible Epochs**: Agents are evaluated and improved across multiple epochs to converge on the optimal solution. - **Editable Agent Framework**: Each genome contains a complete editable agent framework, allowing agents to evolve independently. - **Evolutionary Control Plane**: The evolutionary control plane is kept outside the genome, preventing agents from rewriting their scores or lineage. ### Tech Stack: Python ### Why it Matters: KADATH is crucial for developers looking to automate the optimization of agents for specific goals. By continuously improving agents through evolutionary algorithms, KADATH ensures that the best possible solution is consistently found, making it an invaluable tool for AI development and optimization.

### Overview: Inferflow is an efficient and highly configurable inference engine for large language models (LLMs). It allows users to serve most of the common transformer models by modifying configuration files, without the need to write source code. ### Key Features: - **Extensible and Highly Configurable**: A typical way of using Inferflow to serve a new model is editing a model specification file, but not adding/editing source codes. - **3.5-bit Quantization**: Implementing 2-bit, 3-bit, 3.5-bit, 4-bit, 5-bit, 6-bit, and 8-bit quantization. - **Hybrid Model Partition for Multi-GPU Inference**: Supports three model partitioning strategies: partition-by-layer, partition-by-tensor, and hybrid partitioning. - **Wide File Format Support**: Supports loading models of multiple file formats directly, including pickle, safetensors, llama.cpp, gguf, etc. - **Wide Network Type Support**: Supports decoder-only, encoder-only, and encoder-decoder models. - **GPU/CPU Hybrid Inference**: Supports GPU-only, CPU-only, and GPU/CPU hybrid inference. ### Tech Stack: C++ ### Why it Matters: Inferflow simplifies the process of deploying large language models by providing a modular framework that allows users to serve new models without modifying source code. It supports efficient inference with 3.5-bit quantization and flexible multi-GPU partitioning, making it suitable for deployment on various hardware configurations.

### Overview: Dodaios is an open-source agent OS designed for the agentic era. It runs your coding agents, browser, and apps by voice or chat, on your own machine. Formerly known as Personal Jarvis, it supports Windows, macOS, and Linux, and is compatible with various coding agents, including Claude and ChatGPT. ### Key Features: - **Runs Coding Agents**: Dodaios runs your coding agents, coding CLI, browser, and apps by voice or chat. - **Voice Control**: Supports wake phrases and live conversations with speech recognition. - **Multiple Coding Agents**: Runs multiple coding agents side by side in live panes with call signs. - **Model Compatibility**: Supports local models through Ollama and hosted models through Claude and ChatGPT. ### Tech Stack: - Python ### Why It Matters: - **Local, No Account**: The app and data stay on your machine, eliminating the need for an account. - **Any Model or Subscription**: Supports local models, hosted models, and your existing Claude and ChatGPT plans. - **Multiple Coding Agents**: Runs multiple coding agents side by side in live panes, allowing you to watch and steer multiple agents. - **Voice Control**: Supports wake phrases and live conversations with speech recognition, enabling voice control. ### GitHub Stars: 252 ### Why It Matters: - **Local, No Account**: The app and data stay on your machine, eliminating the need for an account. - **Any Model or Subscription**: Supports local models, hosted models, and your existing Claude and ChatG

Overview: DEEPPOWERS is a Fully Homomorphic Encryption (FHE) framework specifically designed for the Model Context Protocol (MCP), aiming to provide end-to-end privacy protection and high-efficiency computation for the MCP protocol's upstream and downstream ecosystems. Key features: - **End-to-end Encryption**: Ensures that user-LLM interactions are fully encrypted, maintaining data privacy throughout the process. - **Secure Task Execution**: Encrypted data transmission and computation for sensitive operations without exposing data to third parties. - **Seamless MCP Integration**: Deeply integrates with the MCP protocol, enhancing its capabilities with privacy-preserving computations. - **Accelerated MCP Workflows**: Processes encrypted data locally, reducing unnecessary data transfers, and significantly improving MCP operational efficiency. - **Enhanced Tool Security**: Enables secure execution of MCP tools that handle sensitive data, expanding the scope of privacy-preserving tasks. - **Cross-Model Compatibility**: Works with various MCP servers and LLM providers, including DeepSeek, GPT, Gemini, and Claude. - **Reduced Latency**: Minimizes delays in MCP interactions by optimizing encryption, computation, and decryption processes. - **Privacy-Preserving AI**: Provides complete AI assistance workflows while maintaining strict data privacy throughout. - **Scalable Architecture**: Designed to grow with the MCP ecosystem, supporting emerging models and computational tools. Tech stack: C++ Why it matters: DEEPPOWERS leverages the power of Fully Homomorphic Encryption (FHE) to ensure data privacy and high efficiency in the MCP protocol's ecosystem, making it a crucial

### Overview: This repository contains a survey of harmful fine-tuning attacks and defenses for large language models, focusing on the ACM CSUR conference. The survey is updated weekly and includes recent papers accepted by NeurIPS, ICLR, ICML, and ICML2025. It also provides slides and a detailed survey paper accepted by ACM CSUR. ### Key features: - **Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Survey** - **Attacks**: - Shadow Alignment: The Ease of Subverting Safely-Aligned Language Models - Fine-tuning aligned language models compromises safety, even when users do not intend to! - On the Vulnerability of Safety Alignment in Open-Access LLMs - Language Model Unalignment: Parametric Red-Teaming to Expose Hidden Harms and Biases - Lora fine-tuning efficiently undoes safety training in llama 2-chat 70b - BadLlama: cheaply removing safety fine-tuning from Llama 2-Chat 13B - Removing RLHF Protections in GPT-4 via Fine-Tuning - Exploiting Novel GPT-4 APIs - **Defenses**: - Alignment Stage Defenses - Fine-tuning Stage Defenses - Post-Fine-tuning Stage Defenses - **Mechanical study**: - Benchmark - Chef Recommendation: Risk of harmful fine-tuning attack can be more pronounced with jailbreak tuning and

Overview: This repository provides a comprehensive framework for fine-tuning image editing tasks, supporting various architectures including **FLUX Kontext**, **DreamOmni2**, **Qwen-Image-Edit**, and **Qwen-Image-Edit-2509**. The implementation features efficient training through LoRA and an optimized embedding cache system for 2-3x training acceleration. Key features: - **DreamOmni2 Trainer Support (v3.3.0)**: Full support for DreamOmni2 LoRA fine-tuning with cumulative offset positioning for multi-image conditioning. - **Multi-Logger Support (v3.2.0)**: Added unified logging interface supporting TensorBoard, Weights & Biases, and SwanLab. - **Validation Visualization Improvements (v3.1.0)**: Enhanced validation sampling visualization with step-by-step progress display. - **FSDP LoRA Checkpoint Fix (v3.0.2)**: Fixed FSDP training issues with LoRA checkpoint saving and documented performance benchmarks across different training strategies. Tech stack: PyTorch Why it matters: This framework is essential for developers and researchers working on image editing tasks, offering a robust and efficient solution for fine-tuning models with various architectures and performance optimizations.

Overview: This project is a real-time web-based Speech-to-Text app built with Streamlit, featuring the DeepSpeech engine. It leverages Streamlit's capabilities for easy deployment and real-time interaction. Key features: - Real-time speech-to-text functionality - Streamlit-based interface for easy deployment and interaction - Uses the DeepSpeech engine for accurate speech recognition Tech stack: Python Why it matters: This project showcases the integration of Streamlit with real-time media streams, demonstrating how to build interactive applications with minimal code and high performance.

### Overview: Echo-Memory is an official code repository for a controlled study of memory in action-conditioned video world models. It focuses on preserving identity, layout, and viewpoint across action-world generation and revisit trajectories. ### Key Features: - **Memory Modules:** Echo-Memory includes Context, Compression, Spatial, and State-Space memory modules to ensure that the generated scenes maintain historical observations. - **Revisit Assets:** It provides GT replay, in-domain revisit, and open-domain revisit assets for detailed analysis and evaluation. - **Training Recipes:** The code includes reproducible memory rows, ablation scripts, and training recipes for various memory modules. ### Tech Stack: Python ### Why It Matters: This project is crucial for understanding how action-world models preserve identity, layout, and viewpoint through controlled memory studies, providing valuable insights into the mechanisms of memory in artificial intelligence systems.

### Overview: Rectified Diffusion: Straightness Is Not Your Need Rectified Diffusion is a groundbreaking approach that challenges the conventional understanding of straightness in the context of image generation. This project introduces Rectified Diffusion, a novel method that identifies that straightness is not the essential training target and extends the scope of rectified flow. The method aims to improve training efficiency and generation quality, making it a significant advancement in the field of diffusion models. ### Key Features: - **Rectified Diffusion (RD):** A novel diffusion model that focuses on improving training efficiency and generation quality. - **CM (Conditional Model):** A conditional model that enhances the ability to generate high-quality images. - **Phased Diffusion:** A variant of diffusion models that improves the flow of gradients, leading to better training and generation results. ### Tech Stack: - **Python:** The primary programming language used for implementing the Rectified Diffusion model. ### Why It Matters: - **Improves Training Efficiency:** Rectified Diffusion significantly reduces the number of training steps required to achieve high-quality results, making it more efficient. - **Enhances Generation Quality:** The method produces images that are more visually appealing and closer to the intended output, leading to better user experience. - **Extends Scope of Rectified Flow:** Rectified Diffusion extends the application of rectified flow models beyond just straightness, making them more versatile and applicable to a wider range of tasks. ### Repository: G-U-N/Rectified-Diffusion - **Primary Language:** Python - **GitHub Stars:**

Overview: Noobi.ai is a local-first desktop agent that converts a natural-language game idea into a reviewed, playable browser or Godot game using Codex App Server and Godot 4. Key features: - Converts a prompt into a reviewed, playable game - Uses Codex App Server for game production - Supports Web and Godot 4 game formats - Provides a developer preview for macOS Tech stack: TypeScript Why it matters: Noobi.ai streamlines the game production process by breaking it into a bounded game-production loop, ensuring a high-quality, reviewed game that can be played locally.