Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 51 stars

marqo-ai/marqo-ecommerce-embeddings

State-of-the-art embedding models fine-tuned for the ecommerce domain. +67% increase in evaluation metrics vs ViT-B-16-SigLIP.

### Overview: Marqo Ecommerce Embedding Models fine-tuned for the ecommerce domain, achieving a significant improvement in evaluation metrics compared to ViT-B-16-SigLIP. ### Key Features: - **Marqo-Ecommerce-B** with 203 million parameters and 768 dimensions - **Marqo-Ecommerce-L** with 652 million parameters and 1024 dimensions - **Model Release**: Available on Hugging Face ### Tech Stack: Python ### Why it Matters: The Marqo Ecommerce models consistently outperformed all other models across various metrics, demonstrating superior performance in ecommerce image search applications.

Python 51 stars

msalhab96/SpeeQ

A framework for automatic speech recognition

Overview: SpeeQ is a Python-based speech recognition framework designed for developers and researchers to experiment with and train various speech recognition models quickly. Key features: - Pre-implemented model architectures - Supports CTC, Transducer, and Seq2Seq types - Easy installation using pip - Documentation available at https://speeq.readthedocs.io/en/latest/ Tech stack: Python Why it matters: SpeeQ simplifies the process of experimenting with speech recognition models, making it suitable for quick prototyping and testing.

Python 51 stars

mubingshen/MLC-SLM-Baseline

The project is associated with the recently-launched INTERSPEECH 2025 Workshop on Multilingual Conversational Speech Language Model (MLC-SLM) to provide participants with baseline systems for speech recognition and speaker diarization in multilingual conversational scenario.

### Overview: This project aims to provide participants with baseline systems for speech recognition and speaker diarization in multilingual conversational scenarios, focusing on the INTERSPEECH 2025 Workshop on Multilingual Conversational Speech Language Model (MLC-SLM). ### Key Features: - Baseline systems for multilingual conversational speech recognition and diarization. - Real-world multilingual conversational speech dataset for development. - Supports multiple languages including English, Korean, Japanese, and Thai. - Provides evaluation metrics for both tasks: WER (Word Error Rate) and CER (Character Error Rate). - Baseline models include vanilla Whisper-large-v3 encoder and Qwen2.5-7B, as well as Llama3.1-8B. ### Tech Stack: Python ### Why It Matters: This project is crucial for advancing the field of multilingual conversational speech language modeling, enabling the development of AI systems that can understand and respond naturally in multilingual, dynamic, and context-rich environments.

Unknown 51 stars

navalnica/be_nlp_speech_resources

Links to Belarusian NLP and Speech resources

### Overview: This repository contains a curated list of Belarusian Natural Language and Speech Processing resources, datasets, and models. It serves as a valuable resource for researchers, developers, and enthusiasts interested in Belarusian language and speech technology. ### Key Features: - **Datasets**: A collection of datasets including Common Voice, Google FLORES, KNihi Corpus, and others. - **Speech-to-Text Models**: Models like Wav2Vec2, Whisper Small, and Whisper Base fine-tuned on Belarusian datasets. - **Text-to-Speech Models**: Models like Wav2Vec2, Whisper Small, and Whisper Base fine-tuned on Belarusian datasets. - **NLP Models and Tools**: Resources for natural language processing, including models from the Universal Dependencies project and the Tatoeba corpus. ### Tech Stack: Unknown The tech stack for this repository is unknown, but it appears to be related to Belarusian language and speech processing. ### Why It Matters: This repository is crucial for anyone looking to work with Belarusian language and speech technology. It provides a comprehensive list of resources and models that can be used for research, development, and education in the field of Belarusian language processing. ### GitHub Stars: 51 The repository has received 51 GitHub stars, indicating a moderate level of interest and engagement from the community. ### Primary Language: Unknown The primary language of the repository is unknown, but it appears to be related to Belarusian language processing. ### Communities and Platforms: - **say.by**: Telegram community - **corpus.by**:

Java 51 stars

pinecone-io/pinecone-java-client

The official Java client for the Pinecone vector database

Overview: The Pinecone Java SDK provides a comprehensive interface for interacting with Pinecone's vector database, enabling developers to efficiently manage and query large-scale vector datasets. Key features: - Seamless integration with Java applications - Support for both control plane and data plane operations - Customizable OkHttpClient for enhanced performance - Support for HTTP proxy configuration for network environments Tech stack: Java Why it matters: The Pinecone Java SDK simplifies the process of working with Pinecone's vector database, making it easier for developers to manage and query large datasets efficiently.

TypeScript 51 stars

tychenjiajun/art

AI-PP3 is a command-line tool that uses artificial intelligence to analyze RAW photos and generate optimized processing profiles (PP3 files) for RawTherapee.

Overview: AI-PP3 is an experimental command-line tool designed to generate optimized RawTherapee profiles (.pp3 files) using artificial intelligence, focusing on RAW photo analysis and processing. Key features: - AI-driven analysis of RAW files - Batch creation of consistent processing parameters - Customizable development settings via natural language prompts - Multi-model support (OpenAI, Anthropic, Google, Local) - Fine-grained control over PP3 sections (Exposure, Color, Detail) - Interactive preview generation with quality controls Tech stack: TypeScript Why it matters: AI-PP3 aims to streamline the development process for RAW photographers, offering a more efficient and creative approach to processing their images. Repository: tychenjiajun/art Primary language: TypeScript GitHub stars: 51 GitHub description: AI-PP3 is an advanced command-line tool that combines multimodal AI analysis with RawTherapee's processing engine to create optimized development profiles (.pp3 files) for RAW photography. Designed for both automated workflows and creative exploration.

Python 51 stars

yulewang97/ERDiff

[NeurIPS 2023 Spotlight] Official Repo for "Extraction and Recovery of Dpatio-temporal Structure in Latent Dynamics Alignment with Diffusion Models"

### Overview: ERDiff, a state-of-the-art model for extracting and recovering spatio-temporal structure in latent dynamics alignment with diffusion models, has been introduced at NeurIPS 2023. This project showcases its effectiveness on two public datasets: CO-M and RT-M, achieving high R² scores for velocity decoding. ### Key Features: - **Model Architecture**: ERDiff is designed to enhance the extraction and recovery of spatio-temporal structure in latent dynamics alignment using diffusion models. - **Datasets**: The project is validated on the CO-M (center-out reaching, Monkey M) and RT-M (random-target reaching, Monkey M) datasets, providing real-world applications. - **Performance Metrics**: Results are presented in terms of velocity decoding R² scores, with values ranging from 34.32% to 8.38% across different sessions. ### Tech Stack: - **Python**: The primary programming language used for implementing the model and performing experiments. ### Why It Matters: ERDiff offers a robust and stable approach to diffusion model alignment, enhancing the accuracy and reliability of latent dynamics extraction and recovery in various applications, particularly in neuroscience and robotics. ### Repository: The project is hosted on GitHub, with a total of 51 GitHub stars, showcasing its popularity and utility in the research community. ### Environment Setup: - **Conda Environment**: Users can create a conda environment named `erdiff` using the `requirements.txt` file. - **Python Virtual Environment**: Alternatively, a Python virtual environment can be set up using `pip` and `

Python 50 stars

1ytic/open_stt_e2e

PyTorch end-to-end speech recognition

Overview: This repository contains a minimal set of scripts for training language and acoustic models for speech recognition tasks, specifically designed for the open_stt dataset. The training pipeline includes stages such as character-based RNN language model, CNN-RNN acoustic model with CTC loss, and fine-tuning with reinforcement learning and RNN-T loss. Key features: - Character-based RNN language model - CNN-RNN acoustic model with CTC loss - Character-based RNN language model and CNN-RNN acoustic model with RNN-T loss - Fine-tuning with Reinforcement Learning and RNN-T loss Tech stack: PyTorch Why it matters: This project provides a practical solution for training language and acoustic models for speech recognition tasks, specifically tailored for the open_stt dataset. The inclusion of various stages in the training pipeline and the use of reinforcement learning and RNN-T loss make it a valuable resource for developers looking to improve their speech recognition models.

Python 50 stars

ArmanShirzad/SocialMediaContentCreationAndPostingAutomationPlatform

This is an automated content pipeline that scrapes, curates, and summarizes data from a source, then generates and publishes ready-to-post content to Telegram channels, X (Twitter), and Instagram accounts. It also creates and uploads videos to YouTube.

### Overview: SocialMediaContentCreationAndPostingAutomationPlatform is a comprehensive automated content pipeline designed to automate the creation and distribution of content across multiple social media platforms. This tool scrapes the latest articles from a specified source, translates them into another language, summarizes the content, generates videos based on the article content, and posts the translated and summarized content to Telegram, Twitter, and Instagram accounts. Additionally, it supports video uploads to YouTube. ### Key Features: - **Automated Scraping:** Fetches the latest articles from a source, typically TechCrunch. - **Translation:** Translates articles to Persian using LLM-powered translation. - **Summarization:** Summarizes content for social media platforms, adhering to specific limits and constraints. - **Video Generation:** Creates videos from article content using stock images or footage and pexels API. - **Audio Generation:** Creates audio voice over from article content using the TTS feature. - **Multi-Platform Posting:** Posts to YouTube, Telegram, Twitter, and optionally Instagram. - **Queue & Rate Limiting:** Handles Instagram posting asynchronously with rate limits. - **Database Tracking:** Tracks which articles have been posted to which platforms. - **Scheduler:** Supports scheduled runs every 6 hours. ### Tech Stack: - Python ### Why It Matters: This platform is crucial for anyone looking to automate the creation and distribution of content across multiple social media platforms. It streamlines the process of creating engaging and relevant content, ensuring that users receive timely and accurate information in their preferred language.

Jupyter Notebook 50 stars

Burf/ModelSoups

ModelSoups for Tensorflow2 and Torch

Overview: This project implements "Model Soups" using Keras, Tensorflow, and Torch, based on a paper that improves model accuracy without increasing inference time. Key features: - Uniform Soup - Greedy Soup Tech stack: Python, Keras, Tensorflow, Torch Why it matters: This implementation aims to enhance model performance while maintaining efficiency, addressing the need for improved accuracy without sacrificing speed.

Unknown 50 stars

Cre4T3Tiv3/Cre4T3Tiv3

Overview: Cre4T3Tiv3 is a developer-focused project profile for a developer audience. Key features: - Neural network architecture design - Code evolution analysis - Hardware performance optimization - Signal processing and music analysis - Cross-domain engineering with Python, C++, CUDA, Go, Rust, and TypeScript Tech stack: Python, C++, CUDA, Go, Rust, TypeScript Why it matters: Cre4T3Tiv3 leverages domain-invariant structural patterns to diagnose system failures and improve performance across various domains, including neural networks, software engineering, and AI/ML.

Python 50 stars

DualParal-Project/DualParal

[AAAI 2026] Minute-Long Videos with Dual Parallelisms

### Overview: DualParal is a distributed inference strategy designed for Diffusion Transformers (DiT)-based video diffusion models. It optimizes efficiency by parallelizing both temporal frames and model layers, leveraging a block-wise denoising scheme. ### Key Features: - **Parallelization**: DualParal parallelizes both temporal frames and model layers to enhance performance. - **Efficiency**: By reducing the computational load, it achieves faster inference times. - **Model Compatibility**: Compatible with various DiT models, including Wan2.1 and Wan2.2. ### Tech Stack: - Python ### Why It Matters: DualParal significantly reduces the computational requirements of video diffusion models, making them more efficient and faster to run, thus improving the overall performance and usability of these models in applications like video generation and analysis.

Python 50 stars

Harahan/RTDMD

[arXiv 2026] This is the official PyTorch implementation of "RTDMD: Reinforcing Few-step Generators via Reward-Tilted Distribution Matching".

Overview: RTDMD is a reinforcement learning framework designed to improve the performance of few-step generator models, specifically for diffusion models. It combines distribution matching distillation with reward-guided reinforcement learning to achieve state-of-the-art results. Key features: - Cold-start distillation (AC-DMD) - RL fine-tune (RTDMD = GRPO + AC-DMD / BP aux) - Inference - Reward evaluation Tech stack: Python Why it matters: RTDMD significantly enhances the efficiency and accuracy of few-step generator models, particularly in diffusion tasks, by leveraging distribution matching and reward-guided reinforcement learning. Repository: Harahan/RTDMD Primary language: Python GitHub stars: 50 GitHub description: This is the official PyTorch implementation of "RTDMD: Reinforcing Few-step Generators via Reward-Tilted Distribution Matching".

Python 50 stars

HavenFeng/time_reversal

Official repo of "Time Reversal Fusion" (ECCV2024)

### Overview: Time Reversal Fusion (TRF) is an innovative method for generating sequences from given end frames, specifically designed for image-to-video tasks. This project presents a new sampling strategy called Time-Reversal Fusion, which enables the model to generate sequences without the need for tuning or backpropagation. The method is particularly useful for generating sequences that mimic subject motion, camera motion, and video looping. ### Key Features: - **Time-Reversal Fusion (TRF)**: A novel sampling strategy for generating sequences from given end frames. - **Bounded Generation**: Generalizes three scenarios in computer vision: subject motion, camera motion, and video looping. - **Domain-Specific LoRA Patches**: Fine-tuned LoRA "patches" for improved robustness and performance across various domain-specific datasets. ### Tech Stack: - Python 3.10 - PyTorch >= 2.0.1 (Diffusers compatible) - Diffusers ### Why It Matters: Time Reversal Fusion (TRF) offers a new perspective on image-to-video generation, providing a solution to generate sequences that mimic real-world scenarios without the need for fine-tuning or backpropagation. This method is particularly useful for applications requiring high-quality and robust video synthesis, such as generating sequences that capture subject motion, camera motion, and video looping. ### Repository: The project is hosted on GitHub at [HavenFeng/time_reversal](https://github.com/HavenFeng/time_reversal). It has been starred by 50 developers, indicating its

VHDL 50 stars

MohammedRashad/FPGA-Speech-Recognition

Expiremental Speech Recognition System using VHDL & MATLAB.

Overview: This project is an experimental speech recognition system designed for low-end FPGAs using VHDL and MATLAB. Key features: - Speech recognition of digit 1 or 0 - Utilizes VHDL and MATLAB for implementation - Demonstrated using Altera DE0 FPGA Tech stack: VHDL and MATLAB Why it matters: This project aims to explore the limits of low-end FPGAs for speech recognition, showcasing the feasibility of such systems on educational hardware.

Python 50 stars

NVlabs/EoRA

[ICLRW'26] EoRA: Fine-tuning-free Compensation for Compressed LLM with Eigenspace Low-Rank Approximation

Overview: EoRA is a novel fine-tuning-free method designed to enhance the accuracy of compressed Large Language Models (LLMs) without the need for retraining. It achieves this by compensating for compression errors using low-rank matrices, significantly improving performance across various tasks. Key features: - Fine-tuning-free compensation for compressed LLMs - Utilizes eigenspace low-rank approximation to enhance model accuracy - Achieves notable improvements in accuracy metrics such as 10.84% on ARC-Challenge, 6.74% on MathQA, and 11.45% on GSM8K for LLaMA3-8B compressed to 3-bit Tech stack: Python Why it matters: EoRA offers a practical solution for improving the accuracy of compressed models, enabling more efficient and flexible deployment of LLMs without the need for retraining, making it particularly useful for applications requiring high accuracy but with limited computational resources. Repository: NVlabs/EoRA Primary language: Python GitHub stars: 50 GitHub description: [ICLRW'26] EoRA: Fine-tuning-free Compensation for Compressed LLM with Eigenspace Low-Rank Approximation GitHub stars: 50

Python 50 stars

OpenSenseNova/Demystifying_Video_Reasoning

[ECCV 2026] Demystifying Video Reasoning

### Overview This repository contains the official code and resources for **[Demystifying Video Reasoning](https://huggingface.co/papers/2603.16870)**, a paper that explores the reasoning mechanisms in diffusion-based video generation models. ### Key Features - **Model Training**: The repository includes the code for training the model on the provided dataset. - **Visualization Tools**: Tools for decoding intermediate steps and visualizing token-level representations. - **Code Snippets**: Examples of how to use the tools and interpret the results. ### Tech Stack The repository uses Python for the implementation and visualization tools. ### Why It Matters This work provides a systematic understanding of how reasoning emerges in video generation models, offering insights into how to improve the reasoning dynamics of these models.

Python 50 stars

SobeyMIL/TVG

code for "TVG: A Training-free Transition Video Generation Method with Diffusion Models"

### Overview: TVG is a novel training-free transition video generation method that utilizes video-level diffusion models to enhance the quality and smoothness of generated transition videos. This approach addresses the limitations of traditional morphing methods by leveraging Gaussian Process Regression (GPR) to model latent representations and introducing interpolation-based conditional controls and a Frequency-aware Bidirectional Fusion (FBiF) architecture to improve temporal control and transition reliability. ### Key features: - **Training-free Transition Generation**: TVG eliminates the need for additional training, making it more accessible and efficient. - **Gaussian Process Regression (GPR)**: Models latent representations to ensure smooth transitions. - **Interpolation-based Conditional Controls**: Enhances temporal control and transition reliability. - **Frequency-aware Bidirectional Fusion (FBiF)**: Improves temporal consistency and dynamic range. ### Tech stack: - Python ### Why it matters: TVG offers a significant improvement in video generation performance by eliminating the need for training, making it more practical and effective for various media production scenarios. Its ability to generate smooth and dynamic transition videos without the limitations of traditional methods like morphing is particularly valuable in enhancing the flow and coherence of visual narratives. ### Repository: - **Repository**: SobeyMIL/TVG - **GitHub stars**: 50 - **GitHub description**: Code for "TVG: A Training-free Transition Video Generation Method with Diffusion Models" ### Setup: - **Install Environment via Anaconda (Recommended)**: ```bash conda create -n TVG python=3.8.5 con

Python 50 stars

TIGER-AI-Lab/VideoGenHub

A one-stop library to standardize the inference and evaluation of all the conditional video generation models.

### Overview: VideoGenHub is a comprehensive library designed to standardize the inference and evaluation of conditional video generation models across various tasks. ### Key Features: - **Text-to-Video Generation**: Supports multiple models for generating videos from textual prompts. - **Image-to-Video Generation**: Enables the creation of videos from images, enhancing visual storytelling. - **Unified Inference Pipeline**: Ensures fair and consistent evaluation across different models and datasets. - **Model Support**: Currently supports around 10 models, including LaVie, VideoCrafter2, ModelScope, StreamingT2V, Show 1, OpenSora, OpenSora-Plan, T2V-Turbo, and DynamiCrafter2. ### Tech Stack: Python ### Why it Matters: By providing a standardized and fair evaluation framework, VideoGenHub accelerates research and collaboration in the field of video generation, ensuring that advancements are made in a consistent and transparent manner. ### Repository: TIGER-AI-Lab/VideoGenHub - GitHub Stars: 50 - GitHub Description: A one-stop library for standardizing the inference and evaluation of conditional video generation models.

Python 50 stars

VAST-AI-Research/SeqTex

[SIGGRAPH Asia 2025] Official github repo of SeqTex, an end-to-end 3D texture generation method using video diffusion priors.

### Overview: SeqTex is an innovative method for generating 3D textures from video sequences, showcasing its potential in the field of computer graphics and AI research. This project leverages video diffusion priors to create detailed and realistic textures, offering a significant advancement in the field of texture generation. ### Key Features: - **Video Diffusion Priors**: Utilizes advanced video diffusion techniques to generate high-quality textures. - **End-to-End Architecture**: Seamless integration of video and texture generation, providing a complete solution. - **Versatile Applications**: Suitable for various applications in computer graphics, virtual reality, and more. ### Tech Stack: - Python - PyTorch - NVIDIA A100 GPU - Hugging Face Transformers ### Why It Matters: SeqTex represents a significant breakthrough in the field of texture generation, offering a more efficient and effective way to create detailed textures from video sequences. Its applications in computer graphics and beyond highlight its potential impact on the industry and research. ### Repository: - **Repository**: VAST-AI-Research/SeqTex - **GitHub Stars**: 50 - **Primary Language**: Python - **GitHub Description**: Official github repo of SeqTex, an end-to-end 3D texture generation method using video diffusion priors. ### Requirements: - **Environment**: Tested with 1 × NVIDIA A100 80GB GPU. Recommended to use at least 8 GPUs with 80GB memory each for optimal performance. - **Environment Setup**: `conda create -n seqtex python=3.10