AI Trending

Project directory

Projects

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

Jupyter Notebook 24 stars

sourangshupal/simple-rag-langchain

Exploring the Basics of Langchain

Overview: This comprehensive LangChain course covers the fundamentals, document loading, text splitting, embeddings, vector stores, retrieval strategies, and the complete RAG pipeline. Perfect for beginners and intermediate students. Key features: - LangChain Fundamentals - Document Loading - Text Splitting Strategies - Embeddings and Vector Representations - Vector Stores - Retrieval Strategies - Complete RAG Pipeline Tech stack: Jupyter Notebook Why it matters: This course provides a practical guide for building production-ready RAG applications using LangChain, essential for developers and researchers in the field of natural language processing and AI. Repository: sourangshupal/simple-rag-langchain GitHub stars: 24 GitHub description: Exploring the Basics of LangChain
Python 24 stars

structuredllm/itergen

Iterate on LLM-based structured generation forward and backward

### Overview: IterGen is a Python-based framework designed to enhance the control and accuracy of text generation by integrating grammar rules and offering three core functions: `forward`, `backward`, and `view`. These features enable precise navigation and refinement of the generation process, making it ideal for applications requiring structured and grammatically correct output. ### Key features: 1. **Grammar-Aware Decoding** Define custom grammars to structure generation using terminal and non-terminal symbols. IterGen leverages SynCode, a grammar-constrained generation framework, to build on top of. 2. **Forward Generation** The `forward` function generates content up to a specified count of grammar symbols (e.g., sentences, words) or until meeting stopping conditions such as EOS tokens or a maximum token limit. Each call supports customized decoding parameters like temperature and sampling methods. 3. **Backtracking Support** The `backward` function removes generated content by a specified count of grammar symbols, allowing corrections and refinements in real time. 4. **Output Inspection** The `view` function retrieves parts of the generated text corresponding to specific grammar symbols, enabling validation or further analysis. ### Tech stack: Python ### Why it matters: IterGen provides precise control and accuracy in text generation by integrating grammar rules, making it particularly useful in applications requiring structured and grammatically correct output, such as in structured language modeling and natural language processing tasks.
Jupyter Notebook 24 stars

ttop32/ArcaneAnimeGAN

AnimeGAN2 trained on Arcane

### Overview: ArcaneAnimeGAN is an anime-style generator trained on the Arcane dataset, aiming to follow the methodology of Bryandlee's AnimeGAN2. The project uses PyTorch and FastAI for training and includes features like face alignment, color correction, and style blending. ### Key Features: - **Face Alignment**: Aligns frames using the AnimeFaceDetector. - **Color Correction**: Adds color correction to the data preprocessing pipeline. - **Style Blending**: Combines the results of a fine-tuned FFHQ pretrained styleGAN with a pretrained FFHQ model. - **Training Workflow**: Includes steps for getting video data, splitting into frames, aligning images, filtering blurry frames, and creating paired data for training. ### Tech Stack: - **Python**: Jupyter Notebook - **Libraries**: FastAI, Kornia, OpenCV, PyTorch, AnimeFaceDetector, FFHQ, StyleGAN3, AnimeGAN2-PyTorch, AnimeGAN2-PyTorch-Face-Portrait-v1 ### Why It Matters: This project aims to create high-quality anime-style images by blending a fine-tuned FFHQ model with a pretrained FFHQ model, leveraging the Arcane dataset for training. The inclusion of face alignment and color correction ensures the generated images are visually appealing and consistent with the anime genre. ### GitHub Stars: 24 ### Acknowledgement and References: - **AnimeGAN**: Original AnimeGAN project. - **AnimeGANv2**: Follows the methodology of Bryandlee's AnimeGAN2. - **anime
Python 24 stars

uni-medical/retina-text2cfp

[MICCAI 2025] Official implementation of RetinaLogos, a caption-guided framework for high-resolution retinal image synthesis. Enables fine-grained control over retinal anatomy, disease stages, and pathological features using a 1.4M-scale captioned fundus dataset.

### Overview: RetinaLogos is a novel method designed to generate high-resolution Color Fundus Photographs (CFPs) from textual descriptions of retinal conditions. Leveraging a large-scale synthetic caption dataset (1.4 million entries), RetinaLogos provides fine-grained control over anatomical structures and disease progression, offering valuable insights for early detection and clinical training purposes. ### Key Features: - **Large-Scale Synthetic Dataset**: RetinaLogos-1400k, a synthetic dataset containing 1.4 million retinal caption-image pairs, enabling the generation of high-resolution, clinically relevant Color Fundus Photographs (CFPs) across a wide range of retinal conditions. - **Fine-Grained Control Over Retinal Images**: The method allows for precise control over retinal anatomical structures and disease stages using free-form textual prompts, providing unprecedented customization for lesion types, disease progression, and anatomical variations. - **State-of-the-Art Performance**: Demonstrates superior performance in generating clinically accurate retinal images, achieving a 10%-25% improvement in classification tasks such as diabetic retinopathy grading and glaucoma detection. ### Tech Stack: Python ### Why It Matters: RetinaLogos addresses the limitations of existing retinal image synthesis models by providing more detailed and fine-grained control over retinal anatomy and disease progression. This method can synthesize retinal images that capture subtle anatomical variations and disease progression, offering valuable insights for early detection and clinical training purposes. The state-of-the-art performance across multiple datasets demonstrates its effectiveness in
Rust 24 stars

writerslogic/holographic-memory

Local-first associative memory and vector search for AI agents, with Rust, Python, Node.js, and MCP integrations.

### Overview: WARM is a Rust library designed for local-first associative memory and vector search, integrating with Python, Node.js, and MCP. It provides a memory API, vector index, and storage capabilities, ensuring no network calls and no telemetry. ### Key features: - **Document Search**: Supports chunking, stemmed BM25, Boolean and phrase queries, and metadata filters. - **Vector Index**: Graph index with a memory-mapped vertex file, updatable delta for inserts and deletes, and optional query-aware routing. - **Memory API**: Provides `remember`, `recall`, `forget`, and `explain` with time validity and supersession; conversation memory with source provenance. - **Storage**: Atomic replace and delete, crash-safe journal and checkpoints, verified compaction, and one writer per store. - **Security**: Optional AES-256-GCM storage encryption with documented coverage and limits, signed audit log, and COSE/SCITT provenance. - **Vector-Symbolic Layer**: Sparse hypervectors with bind, bundle, and permute, triples with multi-hop queries, and Hopfield cleanup. ### Tech stack: Rust ### Why it matters: WARM is a comprehensive solution for local-first associative memory and vector search, making it ideal for agents, apps, and services that need efficient and secure memory storage and retrieval.
Jupyter Notebook 24 stars

xndien2004/Multimodal-Sarcasm-Detection-for-UITC2024

The Multimodal Sarcasm Detection System detects sarcasm in multimedia content using image-caption generation and NLP. It achieved 1st place in UITC2024 by classifying sarcasm into four categories: image, text, multi sarcasm, and not sarcasm.

### Overview The **Multimodal Sarcasm Detection System** is designed to detect sarcasm in multimedia content using image-caption generation and natural language processing. It achieved 1st place in the UITC2024 competition by classifying sarcasm into four categories: image, text, multi sarcasm, and not sarcasm. ### Key features - **Multimodal Input Handling**: The system can process both text and image-based inputs. - **Sarcasm Classification**: It classifies content into four categories: image sarcasm, text sarcasm, multi sarcasm, and not sarcasm. - **Model Architecture**: Utilizes state-of-the-art models like VIntern-1B-v2 for image captioning and transformers for text analysis. - **Voting Model Integration**: Combines predictions from four different models trained for 2-class, 3-class, and 4-class tasks to ensure accurate final predictions. ### Tech stack - Python 3.10 - Transformers 4.x - SentencePiece 0.2.0 - PyVi 0.1.1 - Einops 0.8.0 ### Why it matters This system is developed for the UITC2024 competition, where it achieved the highest F1 score of 0.4475, placing it in the top 1st place. The system's performance across various metrics in the competition demonstrates its effectiveness in detecting sarcasm in multimodal contexts.
Python 24 stars

yantrikos/yantrikdb-mcp

Persistent cognitive memory for AI agents — drop-in MCP server for Claude Code, Cursor, and Windsurf with temporal decay, contradiction detection, a knowledge graph, and autonomous consolidation, backed by the YantrikDB Rust engine.

Overview: YantrikDB MCP Server is a persistent cognitive memory system designed for AI agents, offering features like semantic recall, knowledge graph, contradiction detection, and procedural learning. It ships as an embeddable engine, network database, or MCP server, backed by YantrikDB Rust. Key features: - Persistent semantic recall - Knowledge graph - Contradiction detection - Autonomous consolidation - Temporal decay - A portable Agent Skills skill for Claude Code, Cursor, Windsurf, and more Tech stack: Python Why it matters: YantrikDB MCP Server provides a reliable and efficient solution for AI agents, enhancing their ability to store and recall information across sessions, making them more effective in complex tasks. Repository: yantrikos/yantrikdb-mcp GitHub stars: 25
Python 24 stars

yueyu1030/Patron

[ACL 2023] The code for our ACL'23 paper Cold-Start Data Selection for Few-shot Language Model Fine-tuning: A Prompt-Based Uncertainty Propagation Approach

Overview: This project presents a novel approach to cold-start data selection for few-shot language model fine-tuning, leveraging prompt-based uncertainty propagation. The method is applied to improve fine-tuning performance on various datasets, including IMDB, Yelp-full, AGNews, Yahoo!, DBPedia, and TREC. Key features: - Novel cold-start data selection method - Prompts-based uncertainty propagation for fine-tuning - Performance on multiple datasets (IMDB, Yelp-full, AGNews, Yahoo!, DBPedia, TREC) - Detailed instructions for running fine-tuning and prompt-based learning experiments Tech stack: Python, Transformers, PyTorch, Scikit-learn, Faiss, SentencePiece, tqdm, tensorboardX, nltk, OpenPrompt Why it matters: This approach significantly enhances the performance of few-shot language model fine-tuning by effectively selecting relevant cold-start data, thereby improving the model's ability to generalize to unseen tasks.
Python 24 stars

zeyofu/Commonsense-T2I

Code for Commonsense-T2I Challenge: Can Text-to-Image Generation Models Understand Commonsense? [COLM 2024]

### Overview: This repository contains code for the paper "Commonsense-T2I Challenge: Can Text-to-Image Generation Models Understand Commonsense" presented at COLM 2024. The challenge involves evaluating text-to-image generation models' ability to produce images that align with common sense, using adversarial prompts to test visual commonsense reasoning. ### Key features: - **Adversarial Challenge**: Evaluates models' ability to generate images that fit commonsense in real-life scenarios. - **Dataset**: Hand-curated and annotated with fine-grained labels, including commonsense type and likelihood of expected outputs. - **Benchmarking**: Compares state-of-the-art (sota) T2I models, achieving a 48.92% accuracy on the Commonsense-T2I dataset. ### Tech stack: Python ### Why it matters: This challenge highlights the limitations of current text-to-image models in understanding and generating images that align with common sense, providing valuable insights into the capabilities and deficiencies of these models.
Python 24 stars

zhaosuifeng/FinRAGBench-V

FinRAGBench-V: A Benchmark for Multimodal RAG with Visual Citation in the Financial Domain (EMNLP 2025)

Overview: FinRAGBench-V is a comprehensive benchmark designed specifically for the financial domain, focusing on visual retrieval-augmented generation (RAG) with visual citation. This benchmark aims to evaluate the capabilities of Multimodal Large Language Models (MLLMs) by integrating rich visual content alongside textual data. Key features: - A multimodal financial corpus containing 60,780 Chinese and 51,219 English pages from 1,104 Chinese and 1,105 English documents. - A high-quality, human-annotated question-answering dataset with seven question categories. - An RGenCite baseline that seamlessly integrates retrieval, generation, and fine-grained visual citation. Tech stack: Python Why it matters: FinRAGBench-V addresses the underrepresentation of visual content in financial RAG research, providing a robust benchmark for evaluating the capabilities of MLLMs in the financial domain.
Unknown 24 stars

zxcvfd13502/PanoFree

[ECCV 2024] PanoFree: Tuning-Free Holistic Multi-view Image Generation with Cross-view Self-Guidance

Overview: PanoFree is a novel method for generating holistic multi-view image sequences, achieving state-of-the-art results without the need for tuning parameters. Key features: - Holistic multi-view image generation - Cross-view self-guidance for improved quality - Tuning-free approach Tech stack: Python, PyTorch Why it matters: PanoFree offers a significant improvement in image generation quality by leveraging cross-view self-guidance, making it a valuable tool for applications requiring high-resolution multi-view imagery without the need for extensive parameter tuning.
Python 24 stars

zyhector/token-rush

Single-stream LLM inference engine for Qwen3.8-27B on one RTX 5090 — int4 quantization + speculative decoding + custom Triton/CUDA kernels; 229–379 tok/s vs llama.cpp/SGLang/vLLM at batch size 1.

Overview: Token Rush is a single-stream LLM inference engine optimized for Qwen3.8-27B on a single RTX 5090 GPU, achieving up to 379 tokens per second with speculative decoding and int4 quantization. Key features: - Single-stream inference - Built for one user, one card, one stream - Int4 quantization for 379 tokens per second - Speculative decoding for improved throughput - Custom Triton and CUDA kernels for performance Tech stack: Python Why it matters: Token Rush offers the fastest inference speed for Qwen3.8-27B on a single RTX 5090, surpassing llama.cpp and vLLM at batch size 1, while maintaining low memory usage and text-only functionality. Repository: zyhector/token-rush Primary language: Python GitHub stars: 24 GitHub description: Single-stream LLM inference engine for Qwen3.8-27B on one RTX 5090 — int4 quantization + speculative decoding + custom Triton/CUDA kernels; 229–379 tok/s vs llama.cpp/SGLang/vLLM at batch size 1.
Python 24 stars

zzz47zzz/CET

[ACL2023] Preserving Commonsense Knowledge from Pre-trained Language Models via Causal Inference

Overview: This repository contains the source code for a method called CET, which aims to preserve commonsense knowledge from pre-trained language models through causal inference. The project provides a fair comparison with various fine-tuning methods and includes implementations of different fine-tuning techniques for a comprehensive evaluation. Key features: - CET fine-tuning method for preserving commonsense knowledge from pre-trained language models - Vanilla fine-tuning method - BSS (Batch-Spectral-Shrinkage) method - ChildTune-F and ChildTune-D methods - Mixout method - NoisyTune method - R3F method - RecAdam method - ReInit method Tech stack: Python Why it matters: This project provides a valuable resource for researchers and developers interested in fine-tuning pre-trained language models to retain their commonsense knowledge, making it easier to adapt these models for specific tasks while maintaining their original capabilities.
Python 23 stars

2U1/Pixtral-Finetune

An open-source implementaion for fine-tuning Pixtral by MistralAI.

Overview: This repository provides a script for fine-tuning Pixtral, a model compatible with MistralAI, focusing on training Transformers compatible with Pixtral-12b. The script supports batch size of 1, making it suitable for long training times. Key features: - Deepspeed - LoRA/QLoRA - Full-finetuning - Enable finetuning `vision_model` while using LoRA - Disable/enable Flash Attention 2 - Multi-image and video training - Training optimized with liger kernel Tech stack: Python Why it matters: This script is useful for fine-tuning Pixtral, which is essential for improving language models and applications in natural language processing and vision tasks. Repository: 2U1/Pixtral-Finetune Primary language: Python GitHub stars: 23 GitHub description: An open-source implementation for fine-tuning Pixtral by MistralAI. README excerpt: An open-source implementation for fine-tuning Pixtral by MistralAI. This repository contains a script for training Transformers compatible [Pixtral-12b](https://huggingface.co/mistral-community/pixtral-12b). However, the model only supports batch size=1, making it suitable for long training times.
Kotlin 23 stars

Adnan9011/github-search

If you're interested in learning about KMP and CMP with Ktor, Koin, Clean architecture, and multi-module setups, this GitHub project provides minimal sample code to help you dive into these technologies!

Overview: This project is a minimal sample code repository focusing on Kotlin, KMP, MultiModule, Clean Architecture, Koin, and Ktor technologies. It provides a practical introduction to these advanced development practices. Key features: - KMP (Kotlin Multiplatform) - MultiModule - Clean Architecture - Implement fake favorite - Koin - Ktor - Shimmer - Error Handling - Tested on both Android and iOS - Working on testing with Kotest Tech stack: Kotlin Why it matters: This project offers a concise yet comprehensive guide to Kotlin and related technologies, making it an excellent resource for developers looking to deepen their understanding of modern software development practices.
Python 23 stars

Anil-matcha/gemini-omni-comfyui

ComfyUI custom nodes for Google Gemini Omni — text-to-video, image-to-video, and video editing via the Gemini Omni API on muapi.ai

### Overview: ComfyUI custom nodes for Google Gemini Omni — text-to-video, image-to-video, and video editing via the Gemini Omni API on muapi.ai. Generate, animate, and edit AI videos directly inside ComfyUI using the muapi.ai Gemini Omni API. ### Key Features: - **Text-to-Video**: Generate video from a text prompt. - **Image-to-Video**: Animate up to 5 reference images. - **Video Edit**: Restyle or transform an existing video clip. ### Tech Stack: Python ### Why it Matters: Use Google Gemini Omni directly within ComfyUI workflows without coding. Ideal for video generation and editing tasks.
Python 23 stars

CodNoob100/null-drift

Overview: This project is an experimental, self-hosted memory-state service designed for AI agents. It combines a Rust HTTP gateway, a Rust state daemon, and Python client adapters. Key features: - A Rust HTTP service for incoming text. - A 10,000-dimensional Hyperdimensional Representation with Salience-Weighted Updates and Anchor Matching. - A state daemon for per-thread 10,000-dimensional HRSA state. - Python client adapters for interacting with the service. Tech stack: Rust. Why it matters: This project explores fixed-width hyperdimensional representations with salience-weighted updates and anchor matching, aiming to provide a more efficient and cost-effective memory state service for AI agents compared to traditional infinite lossless memory. Repository: CodNoob100/null-drift Primary language: Python GitHub stars: 23 GitHub description: No description provided. README excerpt: The repository combines a Rust HTTP gateway, a Rust state daemon, and Python client adapters. It explores fixed-width hyperdimensional representations with salience-weighted updates and anchor matching; it does **not** provide infinite lossless memory, proven cost savings, or a production security boundary.
Kotlin 23 stars

DarrylBayliss/Simon-Says-Android

An Android App recreating the Simon Says game. Uses MediaPipe to run an LLM on device

Overview: Simon-Says-Android is an Android app that recreates the Simon Says game, utilizing MediaPipe to run a Gemma 3.0b language model on the device. Key features: - Simon Says game experience - MediaPipe-powered Gemma 3.0b language model Tech stack: Kotlin Why it matters: This app showcases the integration of artificial intelligence and game development, demonstrating how machine learning can be applied to create engaging and interactive experiences on mobile devices.
Python 23 stars

ForJadeForest/LIVE-Learnable-In-Context-Vector

【NeurIPS 2024】The implementation of LIVE: Learnable In-Context Vector for Visual Question Answering https://arxiv.org/abs/2406.13185

Overview: LIVE: Learnable In-Context Vector for Visual Question Answering is a project that introduces a novel approach to visual question answering using the LIVE model, which is specifically optimized for visual context learning. This project is based on the NeurIPS 2024 paper and aims to improve the accuracy and efficiency of visual question answering tasks. Key features: - Implementing the LIVE model for visual question answering - Optimized for in-context learning - Supports multiple models (IDEFICS-9b, IDEFICS2-8b-base) - Provides data preparation instructions for COCO, VQAv2, and OKVQA datasets Tech stack: Python Why it matters: This project is significant for researchers and developers in the field of computer vision and artificial intelligence, as it offers a more efficient and accurate solution for visual question answering compared to existing methods. The implementation of the LIVE model and the support for multiple models make it a valuable tool for improving the performance of visual question answering tasks. Repository: ForJadeForest/LIVE-Learnable-In-Context-Vector Primary language: Python GitHub stars: 23 GitHub description: 【NeurIPS 2024】The implementation of LIVE: Learnable In-Context Vector for Visual Question Answering https://arxiv.org/abs/2406.13185
Python 23 stars

FreedomIntelligence/TRIM

We introduce new approach, Token Reduction using CLIP Metric (TRIM), aimed at improving the efficiency of MLLMs without sacrificing their performance.

Overview: TRIM, a new approach for improving the efficiency of Multi-Modal Language Models (MLLMs) without compromising performance, has been introduced. This method utilizes CLIP metric for token reduction, aiming to enhance the computational efficiency of models. Key features: - Token Reduction using CLIP Metric - Extensive testing across 12 datasets - Significant reduction in computational overhead while maintaining performance Tech stack: Python, PyTorch, transformers, accelerate Why it matters: This research offers a fresh perspective on token reduction, particularly in MLLMs, and has been validated across various datasets, demonstrating its effectiveness in improving model efficiency without performance degradation.