AI Trending

Project directory

Projects

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

Python 45 stars

a0okiboi4c5/qwen-image-2-api-es

Qwen Image 2.0 API (`qwen-image-2.0`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

### Overview: Qwen Image 2.0 API provides powerful image generation capabilities, making it ideal for various creative and design projects. ### Key Features: - **Model ID and Routes**: Access to Qwen Image 2.0 via the `qwen-image-2.0` model id, which supports multiple image, video, and language models. - **Pricing**: $0.02 per image, with a minimum cost of $1 per task, and billing is applied directly to the task. - **Asynchronous Nature**: Tasks are processed asynchronously, with results available through the `GET /v1/tasks/{task_id}` endpoint. ### Tech Stack: Python ### Why It Matters: - **Cost Efficiency**: Ideal for projects that require high-quality images without the need for a monthly subscription. - **Versatility**: Supports a wide range of image, video, and language models, making it a versatile tool for creative projects. - **Billing Directly**: No need for upfront payments or monthly plans, making it ideal for projects that require flexible pricing. ### Repository: a0okiboi4c5/qwen-image-2-api-es ### Primary Language: Python ### GitHub Stars: 36 ### GitHub Description: Qwen Image 2.0 API (`qwen-image-2.0`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.
Python 45 stars

azle5biyd9td956/z-image-turbo-api-fr

Z-Image Turbo API (`z-image-turbo`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Z-Image Turbo API is a fast, OpenAI-compatible image generation service offered through APIMart, providing one key access to multiple models for a cost-effective $0.01 per image. Key features: - Supports the `z-image-turbo` model for fast text-to-image generation - Pricing is $0.01 per image, with an optional $0.02 per additional image - No subscription required, pay-as-you-go - Charges are reported in the response, and tasks are refunded for failed submissions - Batches and retries are handled by the API, not a custom integration Tech stack: Python Why it matters: Z-Image Turbo API simplifies access to a wide range of image generation models through a single, cost-effective endpoint, making it ideal for teams looking to leverage multiple models without the complexity of custom integrations or monthly subscriptions.
Python 45 stars

bentoml/transformers-nlp-service

Online Inference API for NLP Transformer models - summarization, text classification, sentiment analysis and more

### Overview: This project is an online inference API designed for NLP Transformer models, including summarization, text classification, and sentiment analysis. It utilizes BentoML and HuggingFace's transformers to provide a modular, composable, and scalable solution for building NLP services. ### Key features: - **Model Serving**: Deploy and serve HuggingFace's transformers models. - **Swagger UI**: Interactive Swagger UI for exploring example endpoints. - **BentoML Integration**: Best practices for setting up model services, pre/post-processing, and deployment. - **Container Support**: Pre-built containers for CPU and GPU, requiring a container engine like Docker or Podman. - **Python API**: BentoML Python API for serving models directly. - **gRPC Support**: Included for gRPC compatibility, with additional documentation available. ### Tech stack: BentoML + HuggingFace ### Why it matters: This project demonstrates how to easily deploy and serve HuggingFace's transformers models for various NLP tasks, making it accessible and scalable for developers. The use of BentoML and HuggingFace ensures best practices in model serving and deployment, making it a valuable resource for anyone looking to build NLP services.
Jupyter Notebook 45 stars

davide-coccomini/Detecting-Images-Generated-by-Diffusers

Overview: This project explores the task of detecting images generated by text-to-image diffusion models, specifically focusing on Stable Diffusion and GLIDE. The research evaluates the feasibility of using simple Multi-Layer Perceptrons (MLPs) and traditional Convolutional Neural Networks (CNNs) to detect generated images. The findings highlight the potential for detecting generated images using associated textual information and suggest that incorporating such information can significantly impact performance. Key features: - Detecting images generated by text-to-image diffusion models - Evaluating the feasibility of using MLPs and CNNs for detection - Observing that Stable Diffusion models can detect GLIDE-generated images relatively well - Finding that incorporating textual information rarely leads to significant improvement in detection results - Implications for security and privacy concerns in real-world applications Tech stack: Jupyter Notebook Why it matters: This work provides insights into the feasibility of detecting generated images and has implications for security and privacy concerns in real-world applications. Repository: davide-coccomini/Detecting-Images-Generated-by-Diffusers Primary language: Jupyter Notebook GitHub stars: 45 GitHub description: No description provided.
Python 45 stars

ddlBoJack/MT4SSL

[INTERSPEECH 2023 Best Paper Shortlist] Official implementation for MT4SSL: Boosting Self-Supervised Speech Representation Learning by Integrating Multiple Targets

Overview: MT4SSL is an official implementation of a multi-task learning framework for speech-based self-supervised learning, specifically designed for the Boosting Self-Supervised Speech Representation Learning by Integrating Multiple Targets paper presented at INTERSPEECH 2023. This project aims to enhance speech recognition performance through the integration of multiple targets. Key features: - Integrates multiple targets for speech representation learning - Optimizes model with offline and online targets simultaneously - Achieves good performance and low WER on speech recognition tasks with few pre-training steps Tech stack: Python, PyTorch, fairseq Why it matters: This implementation provides a practical solution for improving speech recognition accuracy by leveraging multiple targets in a self-supervised learning framework, offering a significant boost in performance and efficiency compared to traditional methods.
Python 45 stars

eltonpan/zeosyn_gen

DiffSyn: A Generative Diffusion Approach to Materials Synthesis Planning (Nature Computational Science, 2026)

### Overview: DiffSyn: A Generative Diffusion Approach to Materials Synthesis Planning This project introduces **DiffSyn**, a generative diffusion model designed to assist materials scientists in planning and optimizing synthesis routes. The model leverages advanced neural techniques to predict and generate synthesis recipes for zeolite materials, significantly accelerating the discovery and development of new materials. ### Key Features: 1. **Generative Diffusion Model**: DiffSyn uses a generative diffusion model to generate synthesis recipes, optimizing the process by predicting the most efficient and effective synthesis routes. 2. **Materials Synthesis Planning**: The model is specifically tailored for zeolite materials, which are crucial in various industrial applications. 3. **Neural Network Architecture**: The model incorporates a neural network architecture that can efficiently handle the complexity of zeolite synthesis, providing accurate predictions. 4. **Visualization Tools**: The project includes tools to visualize the generated synthesis recipes, aiding in the interpretation and validation of the model's predictions. ### Tech Stack: - **Python**: The core of the project is implemented in Python, leveraging libraries such as PyTorch for deep learning and NumPy for numerical computations. - **CUDA**: The model is optimized for NVIDIA GPUs, utilizing CUDA for parallel processing to enhance performance. - **Zeosyn Dataset**: The project utilizes the Zeosyn dataset, a comprehensive collection of zeolite synthesis routes, to train and evaluate the model. ### Why It Matters: - **Efficiency**: DiffSyn significantly reduces the time required for materials synthesis planning, making it an invaluable tool for researchers and engineers. - **
Python 45 stars

fegounna/LLM-Fine-Tuning-for-Music-Generation

Overview: This repository focuses on fine-tuning Llama2 for music generation tasks. The process involves converting MIDI files to a text-based representation using `midi_textefinal.py` script. The text representation encodes musical notes as quadruplets, where each variable is separated by a colon (`:`). After converting MIDI files to text, the Llama2 model is fine-tuned for improved music generation performance. Key features: - Converting MIDI files to text representation - Converting text to MIDI file - Generating music Tech stack: Python Why it matters: This project demonstrates the application of Llama2 for music generation, showcasing how to fine-tune a language model for creative tasks, which can be particularly useful in the field of music composition and performance.
C 45 stars

fotonik-sk/stm32n6-speech-recognition

Offline speech-to-text on the STM32N6 microcontroller: free English speech recognition on the NPU, no cloud

### Overview: This project implements speech recognition that runs entirely on a microcontroller, the STM32N6. It recognizes any words, not a fixed list of commands, at a power of 0.2 W. The firmware is designed to run on the STM32N6570-DK board and is available on GitHub. ### Key features: - Speech recognition on the STM32N6 microcontroller - Recognition of any words, not a fixed list of commands - Recognition at a power of 0.2 W - Recognition of English text - Recognition works best within about one meter of the microphone - Recognition is plain text in capital letters: no punctuation, and numbers come out as words - Recognition works offline, no connection, no server and no subscription ### Tech stack: C ### Why it matters: This project aims to provide free speech recognition on a microcontroller, addressing the gap between cheap chips that know a dozen fixed commands and processors or cloud services that understand free speech but need far more power, a network connection, or both.
JavaScript 45 stars

havingautism/Qurio

Qurio brings multi-provider models, custom agents, reusable skills, MCP servers, HTTP tools, retrieval, long-term memory, Deep Research, and Scrapbook into one operating surface.

### Overview: Qurio is a comprehensive AI knowledge workspace that integrates multiple providers, custom agents, reusable skills, and advanced features like Deep Research and structured reasoning. It offers a unified interface for teams demanding more than basic chat, making it easier to switch between tools and access the latest research. ### Key Features: - **Multi-Provider Support**: Connects to various AI providers like Gemini, OpenAI, and more. - **Custom Agents**: Enables specialized workflows with customizable prompts and tool groups. - **Deep Research**: Dedicated workflow for complex tasks, including structured steps and exportable outputs. - **MCP Tool Groups**: Orchestration of tools and enable/disable controls for agents. - **Structured Memory**: Keeps domain-indexed summaries for efficient retrieval and personalization. - **Production-Grade UX**: Features like streaming rendering, Markdown/Code rendering, and thoughtful UI for research/think/tool panels. ### Tech Stack: JavaScript ### Why It Matters: Qurio provides a unified workspace for teams that require more than basic chat, making it easier to switch between tools and access the latest research. It supports multiple providers, custom agents, and advanced features like Deep Research and structured reasoning, making it a valuable tool for teams looking to enhance their AI capabilities. ### GitHub Stars: 45 ### Repository: havingautism/Qurio
Python 45 stars

hopit-ai/Moda

MODA: open fashion retrieval benchmark and models by Hopit AI. MODA (203M, open source), MODA Pro Lite (213M, open weights), MODA Pro (hosted). Full-corpus benchmarks vs FashionSigLIP, SigLIP-SO400M and ZooClaw — one harness, losses shown. #1 open model on LookBench.

### Overview: MODA is an open-source fashion retrieval benchmark and models by Hopit AI. It measures full-corpus performance against FashionSigLIP, SigLIP-SO400M, and ZooClaw, providing losses for comprehensive evaluation. MODA stands out as the top open model on LookBench, with a fine R@1 of 67.63 on LookBench, surpassing a 1.24B-parameter model. ### Key Features: - **Image to Image Retrieval:** Fine R@1 67.63 on LookBench. - **Text to Image Retrieval:** MODA leads the ≤250M class on catalogue search, improving KAGL by 10.9% and Polyvore by 8.7%. - **Text to Image Retrieval:** MODA routes queries by shape, outperforming its constituent models by +5.0% and +5.4%. - **Text to Image Retrieval:** MODA achieves 5 of 6 full-corpus wins over its own base model. ### Tech Stack: - Python ### Why It Matters: MODA provides a comprehensive benchmark and models for fashion retrieval, filling the gap in open-source benchmarks for purchase-grounded queries. It includes a complete retrieval pipeline, isolated component contributions, and publishes all code, results, and methodology.
Jupyter Notebook 45 stars

hyeon-cho/Tangential-Amplifying-Guidance

[ICML2026] Official Implementation of "TAG: Tangential Amplifying Guidance for Hallucination-Resistant Sampling"

Overview: Tangential Amplifying Guidance (TAG) is an innovative technique that addresses the hallucination problem in diffusion models by directly guiding the sampling trajectory. It decomposes the update around the radial and tangential directions, selectively amplifying the tangential component to correct the drift towards off-manifold states. Key features: - Directly counteracting hallucination by steering samples back to high-probability regions. - Compatible with standard `diffusers` pipelines for SD 1.5, SD 2.1, SDXL, and SD 3, with minimal overhead. - Suppresses off-manifold drift while maintaining prompts and seeds unchanged. - Conditional TAG variant that amplifies the aligned tangential component of the guidance signal. Tech stack: The implementation is based on Jupyter Notebook. Why it matters: TAG provides a practical solution to the hallucination problem, ensuring that generated images remain faithful to the input prompts, thus improving the overall quality and reliability of the generated content. Repository: The project is hosted on GitHub at [hyeon-cho/Tangential-Amplifying-Guidance](https://github.com/hyeon-cho/Tangential-Amplifying-Guidance). GitHub stars: The repository has received 45 stars, indicating a growing interest in this technique and its potential applications. Primary language: The implementation is written in Jupyter Notebook, making it accessible and easy to understand for developers and researchers alike. Installation: The project requires Python 3.10 and the `torch` library, which is automatically installed from the Hugging Face Hub on first use.
C# 45 stars

j3soon/speech-to-windows-input

Perform speech-to-text (STT/ASR) with Azure speech service and simulate keyboard to input the recognized text; Supports English, Chinese, Japanese, and more.

### Overview: This project allows users to convert speech to text using Azure Speech Service, simulating a keyboard input for Windows applications. It supports multiple languages including English, Chinese, Japanese, and more, offering improved accuracy compared to built-in Windows features. ### Key Features: - **Speech-to-Text (STT/ASR) with Azure Speech Service** - **Simulated Keyboard Input for Windows Applications** - **Supports Multiple Languages (English, Chinese, Japanese, and more...)** - **Free & Open-Source Implementation** ### Tech Stack: .NET Framework >= 4.6.1 ### Why It Matters: - **Windows 10 Speech Recognition** lacks support for Chinese and more languages. - **Old Speech Recognition Tool** in Windows 10 has poor recognition accuracy. - **Office 365 Speech Recognition** offers better accuracy but requires Office 365 subscription. - **Microsoft Office 365 Speech Recognition** supports English, Chinese, Japanese, and more languages but is limited to Office applications. - **Free & Open-Source Implementation** of a speech recognition tool for all Windows applications. ### Why Use Azure Speech Service as Backend? - **High Recognition Accuracy** similar to Office 365 Speech Recognition. - **Offline Model** not practical due to hardware requirements. - **Free Tier Quota** (5 hours/month) is sufficient and free. - **No Additional Costs** beyond the free tier. - **Custom Model Training** for higher accuracy in domain-specific terminologies. ### How to Use: - **Prerequisites:
Python 45 stars

kayua/MalDataGen

MalDataGen is an advanced Python framework for generating and evaluating synthetic tabular datasets using modern generative models, including diffusion and adversarial architectures.

Overview: MalDataGen is an advanced Python framework designed for generating and evaluating synthetic tabular datasets using modern generative models, including diffusion and adversarial architectures. It aims to support state-of-the-art generative models, enhance model generalization through data augmentation, enable fair benchmarking with reproducible evaluations, and provide publication-ready metrics and visualizations. Key features: - **Generative Models Support**: Supports various generative models like GANs, VAEs, and diffusion architectures. - **Data Augmentation**: Enhances model generalization by augmenting the training data. - **Reproducible Pipelines**: Provides reproducible pipelines for generating synthetic datasets. - **Fair Benchmarking**: Enables fair benchmarking via TS-TR and TR-TS evaluation strategies. - **Publication-Ready Metrics**: Offers publication-ready metrics and visualizations. - **GPU Acceleration**: Supports GPU acceleration for efficient processing. - **CSV/XLS Ingestion**: Supports ingestion of CSV and XLS files for data preparation. - **Custom CLI Scripts**: Allows for custom command-line scripts for data manipulation. - **Integration with Academic Pipelines**: Integrates with academic pipelines for reproducible research. Tech stack: Python Why it matters: MalDataGen is essential for researchers and practitioners in the field of malware detection, providing a robust and reproducible framework for generating synthetic tabular datasets. This framework supports modern generative models, enhances model generalization, and enables fair benchmarking, making it a valuable tool for improving the accuracy and reliability of malware detection systems.
Python 45 stars

lperezmo/real-time-translator

A quick app to translate speech in real time using the Whisper API for transcribing audio, translating, and then using Google Text-to-Speech (gTTS) to play out the translation.

Overview: A quick app to translate speech in real time using the Whisper API for transcribing audio, translating, and then using Google Text-to-Speech (gTTS) to play out the translation. Key features: - Transcribe speech from microphone or audio file - Translate speech or text into English from any language - Display translation results on screen as text - Translation synthesized as speech by Google Text-to-Speech module (gTTS) Tech stack: Python Why it matters: This app allows users to quickly translate speech in real time, making it useful for applications requiring instant translation, such as language learning tools or remote communication platforms.
Python 45 stars

lukasuz/MotionDreamer

[3DV 2025] MotionDreamer: Exploring Semantic Video Diffusion features for Zero-Shot 3D Mesh Animation

### Overview: MotionDreamer is a novel method for zero-shot 3D mesh animation that leverages semantic video diffusion features. This approach enables the automatic re-animation of arbitrary 3D shapes without specialized training, improving efficiency and versatility. ### Key Features: - **Semantic Video Diffusion**: Extracts motion features from videos to guide the animation process. - **Explicit Mesh Representation**: Utilizes a mesh-based representation compatible with existing computer graphics pipelines. - **Zero-Shot Performance**: Achieves superior re-animation accuracy for diverse 3D shapes compared to existing techniques. - **User Study Validation**: Demonstrates superior performance in a user study. ### Tech Stack: Python ### Why It Matters: MotionDreamer offers a time-efficient and versatile solution for re-animating 3D shapes, making it particularly useful in applications where manual animation is impractical or costly. The integration of semantic video diffusion features enhances the accuracy and efficiency of the animation process, making it a significant advancement in the field of 3D animation and computer graphics.
Python 45 stars

mihirahuja1/vectorwrap

Universal vector search wrapper for Postgres, MySQL, SQLite (pgvector, Vector Store, sqlite‑vss).

### Overview: vectorwrap is a universal vector search wrapper that simplifies the process of integrating vector search capabilities into various databases such as PostgreSQL, MySQL, SQLite, DuckDB, and ClickHouse. It provides a seamless way to switch between different vector backends with just a single line of code, making it ideal for prototyping, testing, and production deployments. ### Key Features: - **Universal Vector Search**: Supports multiple databases including PostgreSQL, MySQL, SQLite, DuckDB, and ClickHouse. - **Single Line of Code**: Switch between vector backends with a single line of code, making it easy to integrate. - **Stable API**: Core methods follow semantic versioning, ensuring backward compatibility. - **Production Ready**: PostgreSQL 16+ with pgvector, MySQL HeatWave, and MariaDB 11.8+ for native vector support. - **SQLite and DuckDB**: Great for prototyping with SQLite and DuckDB, which can be used in both production and development environments. - **ClickHouse**: High-performance analytics with ClickHouse, perfect for complex queries and large datasets. ### Tech Stack: vectorwrap uses Python as the primary language. ### Why It Matters: vectorwrap simplifies the process of adding vector search capabilities to a wide range of databases, making it easier to integrate and deploy vector search solutions quickly and efficiently. The stable API ensures compatibility across different versions, making it a reliable choice for both development and production environments.
C++ 45 stars

pablocael/pynear

A python library for efficient KNN search within metric spaces using multiple distance functions.

### Overview: PyNear is a Python library for efficient K-NN search within metric spaces using multiple distance functions. It provides a C++ core with a NumPy-only API, enabling near-duplicate search with recall guarantees, memory-tight ANN, and exact search for specific use cases. ### Key Features: - **Near-Duplicate Retrieval with Pigeonhole Completeness Guarantee**: Finds all nearest neighbors within a specified radius with a guaranteed recall. - **Quantized ANN**: Tracks or beats Faiss's float HNSW up to 91% recall with a quarter of the memory. - **Exact Search**: Provides exact nearest neighbors for specific use cases, such as CV matching and deduplication. - **Scikit-Learn Drop-In Compatibility**: Integrates seamlessly with scikit-learn, making it easy to switch between different indexing methods. - **SIMD Support**: Utilizes SIMD instructions on x86 and ARM architectures for improved performance. - **Zero Native Dependencies**: Requires only NumPy for installation, making it lightweight and easy to use. ### Tech Stack: PyNear is built with C++ at its core, providing a fast and efficient solution for K-NN search. ### Why It Matters: PyNear is designed to handle the trade-offs between recall and performance in K-NN search, offering solutions for near-duplicate detection, image/video deduplication, and other applications where recall is crucial. ### Repository: The repository for PyNear is located at [pablocael/pynear](https://github.com/pablocael/pynear).
Python 45 stars

plainerman/DiffDock-Pocket

Implementation of DiffDock-Pocket: Diffusion for Pocket-Level Docking with Side Chain Flexibility

Overview: DiffDock-Pocket is an all-atom docking model designed for pocket-level docking with side chain flexibility, optimized for computationally generated structures. Key features: - Training and inference capabilities - Visualization tools - Support for both CPU and CUDA environments - Custom complex inference with flexible side chains - Energy minimization option - SMILES representation for ligand input Tech stack: Python Why it matters: DiffDock-Pocket provides a powerful tool for docking studies, especially in the context of flexible side chains, making it suitable for various applications in computational biology and drug discovery. Repository: plainerman/DiffDock-Pocket Primary language: Python GitHub stars: 45 GitHub description: Implementation of DiffDock-Pocket: Diffusion for Pocket-Level Docking with Side Chain Flexibility
Python 45 stars

seahore/PPG-GradVC

A diffusion-based cross-lingual voice conversion model, as my bachelor's thesis

Overview: This project is a diffusion-based cross-lingual voice conversion model developed as part of a bachelor's thesis. It leverages the architecture of Grad-TTS and PPG from an SSL-based phoneme recognizer, aiming to convert any-to-many voice conversion. Key features: - Diffusion-based cross-lingual voice conversion model - Uses Grad-TTS architecture and PPG from an SSL-based phoneme recognizer - Supports multilingual corpora and multilingual voice conversion Tech stack: Python Why it matters: This project is a significant contribution to the field of voice conversion, offering a versatile solution for multilingual voice conversion tasks. Repository: seahore/PPG-GradVC Primary language: Python GitHub stars: 45
Swift 45 stars

spokestack/spokestack-ios

Spokestack: give your iOS app a voice interface!

### Overview: Spokestack provides an extensible speech recognition pipeline for iOS apps, including voice activity detection, wakeword activation, and automated speech recognition. ### Key features: - Voice activity detection - Wakeword activation with two implementations - Simplified Automated Speech Recognition interface - Speech pipeline seamlessly integrates VAD-triggered wakeword detection using on-device machine learning models with transcribing utterances using platform Automated Speech Recognition - On-device Natural Language Understanding utterance classifier - Simple Text to Speech API ### Tech stack: Swift ### Why it matters: Spokestack simplifies the process of integrating speech recognition and voice interaction into iOS apps, making it easier for developers to create voice-activated interfaces.