Projects

### Overview: This project presents "gen2seg: Generative Models Enable Generalizable Instance Segmentation," a code implementation for ICLR 2026. The repository includes pre-trained models, inference scripts, and training instructions for generative models that enable generalizable instance segmentation. The project is designed to be accessible and useful for researchers and developers in the field of computer vision. ### Key Features: - **Pretrained Models**: Available models include Stable Diffusion 2 (SD) and ImageNet-1K-pretrained Masked Autoencoder-Huge (MAE-H). - **Inference Scripts**: Provides scripts for running inference on SD and MAE models. - **Prompting**: Offers a script for generating prompts to enhance the quality of generated images. - **Training**: Includes scripts for training the generative models, with specific requirements for GPU memory. - **Data**: Utilizes datasets Hypersim and Virtual Kitti 2 for training and evaluation. ### Tech Stack: - Python ### Why It Matters: This project is significant for its innovative approach to generative models that enable generalizable instance segmentation, making it a valuable tool for researchers and developers in computer vision. The ability to generate high-quality images with improved segmentation accuracy can significantly enhance various applications in fields such as autonomous vehicles, medical imaging, and more. ### Repository: - **Repository**: UCDvision/gen2seg - **GitHub Stars**: 78 - **Primary Language**: Python - **GitHub Description**: "ICLR 2026" - Code for "gen2seg:

Overview: Comp4D is a compositional 4D scene generation project that extends previous work by focusing on larger-scale movements and realistic interactions between objects. It integrates GPT-4 to generate proper trajectories and designs scenes with more complex compositions. Key features: - Compositional 4D scene generation - Integration of GPT-4 for trajectory design - Support for more objects and larger-scale movements Tech stack: Python Why it matters: Comp4D addresses the challenge of generating complex 4D scenes with realistic interactions, making it valuable for applications requiring detailed and dynamic visual content. Repository: VITA-Group/Comp4D GitHub stars: 78 GitHub description: "Comp4D: Compositional 4D Scene Generation"

### Overview: SGDiff is an official implementation of the paper "Diffusion-Based Scene Graph to Image Generation with Masked Contrastive Pre-Training" by Yang et al. It provides a comprehensive framework for scene graph to image generation using masked contrastive pre-training, leveraging high-quality structural annotations from the LAION-SG dataset. ### Key Features: - **Scene Graph to Image Generation**: Utilizes a masked contrastive pre-training approach to generate images from scene graphs. - **Structured Annotations**: Incorporates high-quality structural annotations from the LAION-SG dataset to represent complex scene structures accurately. - **Pre-trained Models**: Provides new foundations models, including SDXL-SG, for incorporating structural annotations into the generation process. - **Large-Scale Dataset**: Offers a large-scale dataset (LAION-SG) with precise structural annotations for scene graphs, enhancing the model's ability to represent semantic structures in complex scenes. ### Tech Stack: - **Python**: The primary programming language used for implementing the framework. ### Why It Matters: SGDiff offers a robust solution for generating images from scene graphs, leveraging advanced pre-training techniques to enhance the accuracy and quality of the generated images. The inclusion of structured annotations from the LAION-SG dataset ensures that the generated images are highly representative of the scene's semantic structure, making it particularly useful for applications requiring precise representation of complex scenes. ### Repository: - **Repository**: `YangLing0818/SGDiff` - **GitHub Stars**: 78 - **GitHub Description**: Official implementation for "Diffusion-Based Scene Graph to Image

### Overview: This repository contains the official implementation of the paper "Rgb-event fusion for moving object detection in autonomous driving" by Zhuyun Zhou, Zongwei Wu, Rémi Boutteau, Fan Yang, Cédric Demonceaux, and Dominique Ginhac. The dataset used in the paper is called DSEC-MOD, which consists of 16 sequences with a total of 13,314 frames, including 10,495 frames for training and 2,819 frames for testing. ### Key features: - **Dataset**: DSEC-MOD, a moving object detection dataset for autonomous driving. - **Code Implementation**: The repository includes the official code for the paper. - **Publication**: The paper was presented at ICRA'23. ### Tech stack: Python ### Why it matters: This dataset and implementation are crucial for researchers and developers working on autonomous driving and moving object detection, providing a benchmark for evaluating and improving algorithms in this field.

### Overview: SoundPy is an experimental Python package designed for exploring sound and machine learning in the context of sound. It includes features such as signal processing, audio manipulation, feature extraction, and deep learning applications. ### Key features: - **Signal Processing**: Filters, noise reduction, feature extraction, and visualization. - **Audio Manipulation**: Conversion, augmentation, and manipulation of audio files. - **Deep Learning**: CNN, LSTM, and ResNet50 models for speech and sound classification. ### Tech stack: Jupyter Notebook ### Why it matters: SoundPy offers a comprehensive toolkit for researchers and developers interested in exploring sound and applying machine learning techniques to audio data. Its features cater to various applications, including deep learning models, signal processing, and audio manipulation, making it useful for both research and practical use cases.

Overview: The Cursor10x Memory System, now known as DevContext, enhances AI assistants with persistent project context, conversation history, and code relationships across sessions. Key features: - Continuous project-centric context - Intelligent code connections - Minimal resource needs - Self-managing context system Tech stack: JavaScript Why it matters: DevContext takes AI development to the next level by providing continuous, project-centric context awareness, structured metadata, adaptive learning, and seamless workflow integration. GitHub stars: 78 Repository: aiurda/cursor10x-mcp

### Overview: AI Blog Generator is a multi-agent system designed for generating professional blog posts using the Agno framework. It supports leading LLM providers such as OpenAI, Gemini, Claude, and Grok, ensuring flexibility and research-driven content. ### Key features: - **Multi-Agent Workflow:** Orchestrates research, content extraction, and writing using specialized agents. - **Provider Flexibility:** Easily switch between OpenAI, Gemini, Claude, and Grok models. - **Research Driven Content:** Finds, evaluates, and cites authoritative sources. - **Content Scraping:** Extracts and summarizes article content for reference. - **Gradio UI:** Simple web interface for generating and viewing blog posts. ### Tech stack: Python ### Why it matters: AI Blog Generator streamlines the process of generating professional blog posts, ensuring that the content is well-researched and supported by authoritative sources. This tool is particularly useful for developers and content creators looking to leverage the latest advancements in AI and LLMs.

Overview: FALCO is an official PyTorch implementation of the "Attribute-preserving Face Dataset Anonymization via Latent Code Optimization" paper, presented at CVPR 2023. It addresses the challenge of anonymizing face datasets while preserving facial attributes for downstream tasks. Key features: - Anonymizes face datasets while preserving facial attributes. - Uses a pre-trained GAN to optimize the latent representation of images. - Implements a novel feature-matching loss to ensure facial attribute preservation. Tech stack: Python Why it matters: FALCO provides a practical solution for anonymizing face datasets while maintaining important facial attributes, which is crucial for applications like face recognition and training machine learning models without compromising privacy.

### Overview: This project uses Generative Adversarial Networks (GANs) to synthesize images from textual descriptions, showcasing the potential of text-to-image synthesis. ### Key features: - Conditional Wasserstein GAN for generating high-quality images. - Preprocessed dataset of flowers and birds for demonstration. - Detailed instructions on how to download and preprocess the required dataset. ### Tech stack: Python ### Why it matters: This project demonstrates the capabilities of using GANs for text-to-image synthesis, offering a practical application in the field of generative models.

Overview: GreenBoost for Windows is a GPU VRAM extension system that allows larger language models to run efficiently on Windows, using both system RAM and NVMe for larger LLMs. Key features: - Experimental MoE backend for larger models - Portable allocation shim for seamless model loading - Supports quantized MoE models requiring more VRAM than available - GPU VRAM extension for larger LLMs using system RAM and NVMe Tech stack: C/C++ binaries built with MSVC, Windows CUDA driver, NVIDIA Windows driver Why it matters: GreenBoost provides a solution for running larger language models on Windows without the need for additional hardware, making it ideal for developers and researchers looking to maximize model performance on their existing hardware. Repository: denoflore/greenboost-windows GitHub stars: 78 GitHub description: Windows port of GreenBoost - GPU VRAM extension with system RAM & NVMe for larger LLMs.

Overview: This project compares two local Persian speech-to-text models: a fine-tuned Whisper model from nezamisafa and a general large-v3 model from mlx-community, side by side on Apple Silicon. Key features: - Side-by-side comparison of two models - Transcription of audio files in Markdown format - Support for various audio file formats (.ogg, .wav, .mp3, .m4a, .flac, .aac, .opus) - Output includes metadata and raw model transcriptions - Transparent metadata for easy comparison of differences Tech stack: Python Why it matters: This experiment provides a clear comparison of a Persian-specialized model and a fast general model, allowing developers to understand the strengths and weaknesses of each in Persian speech recognition. Repository: erfnzdeh/persian-asr-compare GitHub stars: 79

### Overview: This project implements Particle Guidance, an extension of diffusion-based generative sampling, to enhance diversity and efficiency in generative models. The method is particularly useful in improving the quality and efficiency of sampling in conditional image generation and molecular conformer generation. ### Key Features: - Particle Guidance for Conditional Image Generation - Particle Guidance for Molecular Conformer Generation - Theoretical Analysis and Empirical Experiments ### Tech Stack: Python ### Why it Matters: The project provides a practical implementation of Particle Guidance, which offers a significant improvement in diversity and sample efficiency over traditional methods. It is particularly useful in applications like conditional image generation and molecular conformer generation, where it can lead to better quality and faster convergence. ### Repository: gcorso/particle-guidance ### GitHub Stars: 78 ### Primary Language: Python ### Stable Diffusion Assets: -  -  -  ### Running Visualization: - Code depends on PyTorch >= 1.2, numpy, and matplotlib. - Stable Diffusion-1.5 model is required. - Code is available in the `stable_diffusion` folder. ### Quantitative Comparison: - Data processing: - Step 1: Follow the instruction on top of `data_process.py` - Step 2: Run `python3 data_process.py` to randomly sample 5K image-text pair from COCO

Overview: The Glami-1M dataset is the largest multilingual image-text classification dataset, comprising over 1 million images and corresponding captions in 100 languages. Key features: - Multilingual support across 100 languages - Comprehensive image-text classification dataset - Over 1 million images and captions Tech stack: Jupyter Notebook Why it matters: This dataset is crucial for researchers and developers working on multilingual image-text classification tasks, providing a rich resource for training and evaluating models in various language environments.

### Overview: SNKV is a lightweight key-value store designed for simplicity, security, and performance. It leverages SQLite's B-tree storage engine for efficient storage and retrieval, eliminating the need for a SQL layer. The store supports ACID transactions, native TTL functionality, and encryption, making it suitable for applications requiring high performance and security. ### Key Features: - **ACID Transactions and WAL Mode**: Ensures data consistency and concurrent readers, single writer. - **Native TTL**: Per-key expiry, lazy eviction, and bulk `kvstore_purge_expired()`. - **Encryption**: Per-value XChaCha20-Poly1305 with Argon2id key derivation, transparent to all APIs. - **Column Families**: Logical namespaces within one database. - **Iterators**: Forward, reverse, prefix, and seek. - **Vector Search**: HNSW index via `usearch` stored in the same `.db` file. - **Thread-Safe**: Shares one handle or each thread's handle for parallel reads in WAL. - **Single-Header**: Clean Valgrind build. - **SQLite Compatibility**: Compatible with SQLite page/WAL-level tools. ### Tech Stack: - **C** - The primary language used for implementation. ### Why It Matters: SNKV is designed for applications that require high performance, security, and simplicity. Its lightweight nature and direct SQLite integration make it an excellent choice for embedded systems and applications where storage and performance are critical. ### Build: - **Linux / macOS**: ```bash make make snkv

### Overview: This repository contains a comprehensive collection of the most important papers related to contrastive pretraining for vision, language, and audio models, focusing on the most impactful and innovative research in the field. ### Key Features: - **CLIP (Contrastive Language-Image Pretraining)**: Contains papers directly related to CLIP, including its primary pretraining objective and its extensions. - **Align (Scaling Up Visual and Vision-Language Representation Learning)**: Focuses on extending CLIP to include noisy alt-text for improved efficiency. - **CLOOB (CLOOB: Modern Hopfield Networks with InfoLOOB)**: Addresses the issue of saturation in the InfoNCE objective by using modern Hopfield networks. - **DeCLIP (Supervision Exists Everywhere)**: Improves data efficiency by incorporating supervision into contrastive learning. - **FILIP (Fine-grained Interactive Language-Image Pre-Training)**: Enhances semantic alignment by adding token-wise maximum similarity between visual and textual features. - **DeFILIP (Combining DeCLIP and FILIP)**: Provides a benchmark for data, model, and supervision, combining the strengths of DeCLIP and FILIP. - **PyramidCLIP (Hierarchical Feature Alignment)**: Allows for relaxed assumptions about the correspondence between images and metadata, improving feature alignment. ### Tech Stack: Python ### Why It Matters: The papers in this repository are crucial for researchers and practitioners looking to advance the state-of-the-art in multimodal pretraining, providing a comprehensive overview of the most impactful and

Overview: This project is a Python-based voice control utility designed to allow users to control Spotify through the terminal. Key features: - Play a track by name - Play a random track - Play an album - Play songs by an artist - Play a playlist from Your Library - Pause, resume, skip, and set volume - Toggle repeat and shuffle - Show the current song - Go back to the last played song - Quit the program Tech stack: Python, spotipy (Spotify API wrapper), speech_recognition Why it matters: This project provides a convenient way to control Spotify using voice commands, making it easier for users to manage their music without needing to navigate through the Spotify app.

Overview: Tokovo is an AI-native studio designed for creating multi-device shows that happen inside phones. It simplifies the production process by providing a phone-native production stage with features like one or many simulated devices, app worlds, OS surfaces, camera direction, sound, voice, and deterministic rendering. Key features: - Generate or write a script - Place the action inside one or many simulated devices - Direct pacing, camera moves, captions, keyboard input, sound, voice, and cliffhangers - Preview the episode in browser - Render a vertical MP4 for Shorts, Reels, TikTok, and serialized story formats Tech stack: TypeScript Why it matters: Tokovo streamlines the production process for phone-native stories, allowing creators to focus on the story rather than the technical details, and provides a controlled environment for AI-assisted production, making the process more efficient and consistent.

Overview: This is the official n8n node for interfacing with Qdrant, a vector similarity search engine that provides a production-ready service with a convenient API for storing, searching, and managing vectors. Key features: - List Collections - Create Collection - Update Collection - Get Collection - Collection Exists - Delete Collection - Upsert Points - Retrieve Point - Retrieve Points - Delete Points - Count Points - Scroll Points - Batch Update Points - Update Vectors - Delete Vectors - Query Points - Query Points In Batch - Query Points Groups - Matrix Pairs - Matrix Offsets - Set Payload - Overwrite Payload - Delete Payload - Clear Payload - Payload Facets - Create Payload Index Tech stack: TypeScript Why it matters: This n8n node simplifies the integration of Qdrant into n8n workflows, enabling efficient vector similarity search and retrieval for various applications.

Overview: This project introduces a diffusion-based framework for spatio-temporal point processes, presented in the research track paper at KDD 2023. The framework is implemented in Python and tested on Linux with PyTorch 1.7 and Python 3.7.10. Key features: - Spatio-temporal diffusion point processes - Implementation based on DDPM - Training on various datasets including Earthquake, COVID19, Citibike, and Independent Tech stack: Python Why it matters: This framework offers a novel approach to spatio-temporal point processes, potentially improving the accuracy and efficiency of modeling temporal dynamics in various applications.

Overview: IntellyWeave is an AI-powered OSINT platform that automates the extraction of entities from documents, visualizes relationships on maps and networks, and supports hypothesis-driven investigation and document processing. Key features: - Entity extraction using GLiNER - Geospatial visualization with Mapbox - Network analysis and graph visualization - Archive discovery and hypothesis-driven investigation - Multi-agent reasoning for complex questions Tech stack: FastAPI, Next.js, Weaviate, DSPy Why it matters: IntellyWeave simplifies OSINT analysis by automating the tedious tasks of entity extraction, relationship mapping, and geographic visualization, enabling teams to focus on the analysis of intelligence. Repository: vericle/intellyweave GitHub stars: 78