Projects

### Overview: Langue is a modern platform designed for creating and learning conlangs, focusing on the unity, liveness, and collaboration of diverse language communities. It aims to be the most common platform for online and offline conlanging, leveraging the power of JavaScript engines and isomorphic interfaces. ### Key Features: - **Modules**: Offers a flexible framework for developers to choose specific features. - **Distributed**: Works independently and synchronizes edits in real-time. - **Distinctive Channels**: Supports phonetic, gestural, and luminance-based languages. - **Unity**: Represents languages as a single graph, linking synonyms and translations. - **Liveness**: Synchronizes real-time language usage and user activities. ### Tech Stack: TypeScript ### Why It Matters: Langue is a platform that aims to be the most common and accessible platform for conlanging, leveraging the power of JavaScript engines and isomorphic interfaces. It is designed to be the most common platform for online and offline conlanging, making it a valuable resource for language enthusiasts and creators. ### GitHub Stars: 45 ### Repository: yuhr/langue ### Primary Language: TypeScript ### Plans: Currently in the planning stage, with the official web service (langue.link) set to go live once the project is finalized. The repository is likely to be squashed periodically. ### Logo: The logo features a repetitive looping symbol modeled after the IPA transcription /læŋ/, a knot symbolizing the project's goal of linking diverse language communities, and an underline symbolizing infrastructure. The colors from

Overview: Alith is a developer-friendly AI agent framework designed for the Web3 and Crypto, providing a modular and extensible platform for building on-chain AI agents. Key features: - LazAI Gateway support for wallet management, transactions, and privacy features. - Support for various AI models, including small and large language models. - Highly extensible architecture with customizable roles, goals, tools, and behaviors. - Workflow support for implementing any workflow pattern. - Cross-language SDKs for Rust, Python, and Node.js. - High-performance AI training and inference leveraging Rust and GPU optimizations. - Web3 friendly and secure integration with blockchain capabilities. Tech stack: Rust. Why it matters: Alith offers a robust and secure platform for developers to build, deploy, and manage AI agents on the Web3 and Crypto, making it ideal for projects in DeFi, NFTs, and DAOs. Repository: 0xLazAI/alith GitHub stars: 44 GitHub description: Simple, Composable, High-Performance, Safe and Web3 Friendly AI Agents and LazAI Gateway for Everyone.

Overview: This project is an educational CNN-LSTM image captioning baseline that uses a ResNet-50 image encoder and an LSTM language decoder. It includes vocabulary building, preprocessing, training with BLEU evaluation, and inference. The project generates natural language captions for images and includes metrics, model checkpoints, and visualization outputs. Key features: - Load image-caption pairs from a CSV file - Build a vocabulary from training captions - Extract visual features from images using a CNN encoder - Train an LSTM decoder to generate captions - Train an optional attention decoder (Show, Attend and Tell) - Evaluate generated captions using BLEU scores - Save trained model checkpoints - Save vocabulary files - Generate training and validation loss curves - Generate BLEU score plots - Run inference on new images - Support greedy decoding - Support beam search decoding Tech stack: Python, PyTorch, ResNet-50, LSTM, BLEU evaluation. Why it matters: This project serves as a clean, educational example of a CNN-LSTM image captioning baseline, demonstrating the workflow, architecture, and evaluation metrics commonly used in such projects.

### Overview: Bria 3.2 is a state-of-the-art commercial-ready text-to-image model with 4 billion parameters, providing exceptional aesthetics and text rendering. It outperforms leading open-source models and is fully licensed, ensuring legal and ethical compliance. ### Key Features: - **4 Billion Parameters**: Superior to open-source models. - **4B Parameters vs 12B Other Models**: 3.2 is X3 smaller, making it highly efficient. - **Same Architecture as 3.1**: Ideal for upgrading without disruption. - **2x Faster Fine-Tuning**: Significantly reduces time for model fine-tuning on L40S and A100. - **Commercial Ready**: Built on licensed data, ensuring compliance and security. ### Tech Stack: Bria AI uses Python, PyTorch, and Hugging Face for development and deployment. ### Why It Matters: Bria 3.2 offers unparalleled performance, aesthetics, and text rendering capabilities, setting a new benchmark in the AI industry. Its commercial licensing ensures legal and ethical compliance, making it ideal for enterprise and commercial use. ### Repository: Bria-AI/BRIA-3.2 ### GitHub Stars: 44 ### GitHub Description: GitHub repository for the Bria 3.2 pipeline ### GitHub URL: <https://github.com/Bria-AI/BRIA-3.2> ### TL;DR: Bria 3.2 is the next-generation commercial-ready text-to-image model with 4 billion parameters, providing exceptional aesthetics and text rendering.

### Overview: D²PPO: Diffusion Policy Policy Optimization with Dispersive Loss is an enhanced version of Diffusion Policy Policy Optimization with dispersive loss regularization designed for improved representation learning and policy robustness in continuous control and robot learning tasks. This project was accepted by AAAI 2026, showcasing its effectiveness in enhancing diffusion-based policy learning and improving sample efficiency. ### Key Features: - **Dispersive Loss**: A regularization technique that prevents representation collapse in diffusion policy networks, enhancing diffusion-based policy learning. - **Improved Sample Efficiency**: 22.7% improvement in pre-training, 26.1% after fine-tuning. - **Enhanced Policy Performance**: Achieves 94% average success rate on manipulation tasks. - **Vision-based Robotics**: Validated on Robomimic benchmark and real Franka Emika Panda robot. ### Tech Stack: - Python ### Why It Matters: D²PPO addresses a fundamental challenge in diffusion policy networks—representation collapse, which can lead to redundant or collapsed representations. By preventing representation collapse, D²PPO enhances diffusion-based policy learning, improves sample efficiency, and boosts policy performance, making it a significant advancement in continuous control and robot learning tasks.

### Overview: This project introduces a novel method called FARMS (Fixed-Aspect-Ratio Matrix Subsampling) for analyzing the eigen-spectrum of neural network weight matrices. By subsampling submatrices with a fixed aspect ratio, FARMS mitigates the issue of aspect ratio bias, ensuring a more accurate and fair analysis of neural network performance. ### Key features: - FARMS normalizes weight matrices by subsampling submatrices with a fixed aspect ratio. - It measures the average eigen-spectrum of these subsampled submatrices instead of the heavy-tailed distribution of the original eigen-spectrum. - The project includes code for both Image Classification and Language Model Pruning, demonstrating its versatility across different domains. - The code is open-source and available on GitHub, with a high number of stars indicating its popularity and usefulness. ### Tech stack: Python ### Why it matters: This method addresses a significant issue in the analysis of neural network performance, ensuring fair and accurate evaluation across different aspects of the network's structure and function.

* Model Site: https://github.com/QwenLM/Qwen ## Qwen-Chat Prompt Template ```python template = """<|system|> You are a helpful assistant. <|user|> {query} <|assistant|> """ ``` * References: * https://github.com/QwenLM/Qwen/blob/5aa84bdfd3237 * Model Site: https://github.com/QwenLM/Qwen ## Qwen-Chat-2 Prompt Template ```python template = """<|system|> You are a helpful assistant. <|user|> {query} <|assistant|> """ ``` * References: * https://github.com/QwenLM/Qwen/blob/5aa84bdfd3237 * Model Site: https://github.com/QwenLM/Qwen ## Qwen-Chat-3 Prompt Template ```python template = """<|system|> You are a helpful assistant. <|user|> {query} <|assistant|> """ ``` * References: * https://github.com/QwenLM/Qwen/blob/5aa84bdfd3237 * Model Site: https://github.com/QwenLM/Qwen ## Qwen-Chat-4 Prompt Template ```python template = """<|system|> You are a helpful assistant. <|user|> {query} <|assistant|> """ ``` * References: * https://github.com/QwenLM/Q

Overview: PlantDreamer is a new framework designed to generate realistic 3D plant models using diffusion-guided Gaussian splatting. It supports the creation of bean, kale, and mint plants. Key features: - Generates 3D plant models based on synthetic L-System meshes or real plant point clouds. - Supports the generation of bean, kale, and mint plants. - Uses Gaussian diffusion for splatting to achieve high-quality textures and geometry. Tech stack: Python Why it matters: This framework addresses the challenge of generating realistic 3D plant models, which are difficult to achieve with existing text-to-3D models. PlantDreamer provides a solution for creating detailed and textured plant models, making it valuable for applications in virtual environments, education, and more. Repository: Lewis-Stuart-11/PlantDreamer GitHub stars: 44 GitHub description: PlantDreamer: Achieving Realistic 3D Plant Models with Diffusion-Guided Gaussian Splatting [CVPPA: ICCVW 2025] Supporting material: Full explanation of the process can be found on the paper [arxiv.org/abs/2505.15528], and supporting material can be accessed via the dataset [plantimages.nottingham.ac.uk/datasets.html#plantdreamer]. The project uses the following repositories: [GaussianDreamer](https://github.com/hustvl/GaussianDreamer) and the GaussainDreamer [threestudio extension](https://github.com/cxh0

### Overview: This project is a PHP wrapper for the Imagick library, designed to generate document images. It supports various features such as element count, structure, padding, margins, borders, backgrounds, and rotations, making it versatile for creating professional-looking images. ### Key Features: - **Customizable Elements**: Supports any element count and structure. - **Flexibility in Styles**: Handles padding, margins, borders, backgrounds, and rotations. - **Custom Elements**: Allows creating custom elements. - **Computed Properties**: Automatically applies computed properties. - **Simple Markup**: Uses HTML-like markup for easy integration. - **Custom Canvas**: Allows custom canvas manipulation using Imagick methods. - **Optimized Performance**: Faster and more efficient than HTML->PDF->Image conversion. ### Tech Stack: - **PHP**: Version 5.6+ required. - **Imagick Extension**: Essential for image manipulation. ### Why It Matters: This tool simplifies the process of generating professional-looking document images, making it an invaluable asset for developers looking to enhance their projects with visually appealing and easily customizable images.

### Overview: MMM-WhisperGPT is a MagicMirror² module designed to integrate Whisper and ChatGPT for interactive voice transcription and response. It's built using JavaScript and is self-hosted for privacy and security. ### Key Features: - **Voice-to-Text Transcription**: Uses Whisper for real-time transcription of voice commands. - **ChatGPT Integration**: Passes transcribed text to ChatGPT for intelligent responses. - **Privacy and Security**: Utilizes Picovoice for offline trigger word processing and Whisper for self-hosted transcription. - **Customizable Options**: Configurable through a configuration block in `config/config.js`. ### Tech Stack: - JavaScript ### Why It Matters: This module allows MagicMirror² users to interact with AI-powered voice commands, enhancing the user experience with personalized and intelligent responses.

### Overview: Marlin is a diagnostic tool designed to quickly categorize dog behavior incidents into medical, behavioral, or urgent categories. It uses a lens ensemble approach, combining independent diagnostic lenses to provide a ranked differential and a handling protocol for owners to discuss with veterinarians or trainers. ### Key features: - **Lens Ensemble**: Scores incidents independently by medical, behavior, environmental, and breed lenses, then combines weights into a ranked differential. - **Red-flag Gate**: Bypasses the diagnostic process for reports mentioning seizure-like signs, sudden aggression, self-injury, or acute pain. - **Drift Detection**: Tracks the acceptance rate of each lens over time to detect if a lens has become stale. - **Autotuner**: Adjusts lens weights based on owner-confirmed outcomes over time. - **Breed Lookup**: Incorporates known breed-linked tendencies into the differential. ### Tech stack: Go ### Why it matters: Marlin provides a structured and evidence-based approach to diagnosing dog behavior issues, helping owners make informed decisions with veterinary or behaviorist input.

### Overview: ComfyUI-HunyuanImage3 is a project that allows the native running of Tencent's HunyuanImage-3.0 model on ComfyUI, utilizing the KSampler, VAE Decode, and ComfyUI's own memory management techniques. This model is optimized for a single 24 GB GPU, providing efficient and fast performance. ### Key Features: - **Instruct-Distil**: The fast 8-step Instruct-Distil model, ideal for everyday use with a 26-second per image generation time. - **Instruct**: The full 50-step Instruct model, suitable for detailed editing and multi-image fusion. - **Base**: The pretrained Base model, providing strong photographic detail from the pretrained model. - **Prompt Rewriting**: The model can expand short prompts into detailed descriptions before drawing. - **Spectrum**: An optional speed-up that makes the 50-step models about 3.4 times faster, at the cost of some drift from the full render. - **Ready-Made Weights**: Available in three sizes (W4A8, int8, bf16) on HuggingFace, including scripts for making them. ### Tech Stack: The project uses Python and ComfyUI to run the HunyuanImage-3.0 model. ### Why It Matters: This project is significant because it allows users to leverage Tencent's high-performance model directly within ComfyUI, without the need for additional Python wrappers, thus providing a seamless and efficient workflow.

Overview: FlowBlock is a training-free parallel decoding framework designed for self-correcting diffusion language models, specifically the LLaDA-2.1 series. It optimizes the decoding process by allowing blocks to be processed in parallel, reducing latency and improving throughput. Key features: - Up to 2.95× higher Throughput (TPS) compared to LLaDA-2.1 under batched serving. - Up to 53.6% lower latency. - Accuracy preserved or improved across eight benchmarks. - No training required, runs the unmodified LLaDA-2.1 series weights. Tech stack: Python, PyTorch, SGLang, Wavefront Decoding. Why it matters: FlowBlock simplifies the decoding process by enabling parallel decoding, which can significantly enhance the performance and efficiency of self-correcting diffusion language models like LLaDA-2.1 series, without the need for additional training. Repository: Red-EAD/FlowBlock Primary language: Python GitHub stars: 44 GitHub description: FlowBlock: Wavefront-Parallel Decoding Framework for Self-Correcting Diffusion Language Models README excerpt: FlowBlock is a training-free parallel decoding framework for block-wise diffusion language models (dLLMs). It optimizes the decoding process by allowing blocks to be processed in parallel, reducing latency and improving throughput. It is built on two mechanisms: Gated Wavefront Decoding (GWD) and Heterogeneous Wavefront Packing (HWP). The code is designed to be easily installable and runs

### Overview: Run Kimi on your own consumer GPU, not someone's API. A single 32 GB card (RTX 5090, or any datacenter card) serves Kimi-Linear-48B at 113.83 tok/s, up 3.18x from fused INT4 kernels, and Claude Code, Codex, Cline, Aider, and opencode all connect to it unchanged. ### Key features: - **k3 Bridge**: A standalone adapter for an existing OpenAI Chat Completions backend. - **Selective INT4**: Optimized INT4 quantization for Kimi-Linear-48B. - **Fused Decode Kernels**: 3.18x performance improvement. - **Detects Client Preset**: Translates requests and responses between different protocol families. ### Tech stack: Python ### Why it matters: Runs Kimi-Linear-48B on your own consumer GPU, bypassing the need for an OpenAI API. Achieves 113.83 tok/s throughput, up to 3.18x faster than the original Kimi model. Supports Claude Code, Codex, Cline, Aider, and opencode without modification. Ships with a deployable package for production use. Repository: RightNow-AI/local-kimi GitHub stars: 44 GitHub description: Optimized local serving engine for Kimi-Linear-48B: INT4 quantizer, fused decode kernels, and an OpenAI

Overview: This project focuses on a real-time streaming Korean speech-to-text model that utilizes deep learning techniques for both speech enhancement and speech recognition. The model is designed to work on a CPU, making it suitable for deployment in various applications. Key features: - Real-time streaming Korean speech-to-text model - Uses deep learning techniques for both speech enhancement and speech recognition - Implements waveform domain processing for speech enhancement - Utilizes denoiser from Facebook and NVIDIA's Nemo framework for conformer CTC - Supports both CPU and GPU environments Tech stack: Python Why it matters: This project offers a practical solution for real-time Korean speech recognition, which is essential in applications such as voice assistants, transcription services, and more, making it valuable for developers looking to integrate advanced speech processing into their projects.

### Overview: Nano Banana Generator is an open-source Next.js SaaS application designed for generating and editing high-quality AI images. Built with production-ready features, it includes seamless user authentication, credit management, and billing integration. ### Key Features: - **Text-to-Image Generation**: Generate stunning visuals with text prompts. - **Multi-Image Reference Editing**: Transition smoothly to editing mode by uploading local images or referencing external images. - **Smart Search Integration**: Enhance prompts with validations or enhancements. - **Responsive UX**: Dynamic sliding dropdowns and mobile-stacked responsiveness. - **Secure Persistence**: All creations are stored in a PostgreSQL database for a customized user gallery. - **Production-Ready SaaS**: Built-in Google OAuth and Stripe Checkout workflows. ### Tech Stack: Next.js ### Why It Matters: - **Production-Ready SaaS**: Complete with Google OAuth and Stripe Checkout workflows. - **Dual-Mode Studio**: Seamless toggle between prompt-based Text-to-Image generation and Multi-Image Reference editing. - **Historical Archive**: Securely persist all creations to a PostgreSQL database. - **Responsive UX**: Dynamic sliding dropdowns and mobile-stacked responsiveness. - **Extensible API**: Easily swap out the underlying AI engine without breaking the application UI. ### Related Projects: - [awesome-ai-image-models](https://github.com/Anil-matcha/awesome-ai-image-models) — Compare AI image models by API, price, and quality. - [Nano-Banana-2.1-API](https://github.com/Anil-matcha/Nano-B

### Overview: Conformer-Athena is an open-source implementation of end-to-end speech processing engine, designed to make speech processing accessible to everyone. It includes models like Hybrid Attention/CTC based ASR, Speech-Transformer, and more. ### Key features: - **Hybrid Attention/CTC based end-to-end ASR** - **Speech-Transformer** - **Unsupervised pre-training** - **Multi-GPU training on one machine or across multiple machines with Horovod** - **End-to-end Tacotron2 based TTS with support for multi-speaker and GST** - **Transformer based TTS and FastSpeech** - **WFST creation and WFST-based decoding** - **Deployment with Tensorflow C++** ### Tech stack: Python ### Why it matters: Conformer-Athena simplifies the development and deployment of speech processing models, making them accessible to both industrial application and academic research.

Overview: SADA, or Stability-guided Adaptive Diffusion Acceleration, is an innovative project that leverages a unified stability criterion to accelerate off-the-shelf diffusion models, specifically improving the performance of SDXL, SD-2, and Flux architectures. This project is designed to streamline the process of applying stability-guided adaptive diffusion acceleration to any project built on HuggingFace Diffusers, making it easier to achieve significant performance improvements. Key features: - Accelerates SDXL, SD-2, and Flux architectures by up to 2.02x, 1.86x, and 1.80x respectively. - Integrates seamlessly with any project using HuggingFace Diffusers. - Provides a quick start guide with three demos for testing SADA with SD-2, SD-XL, and Flux. - Offers flexibility with finetuning options for LoRA checkpoints and fine-tuned models. Tech stack: Python Why it matters: SADA is crucial for anyone working with diffusion models, as it significantly enhances the performance of these models, making them more efficient and faster to train and deploy. This project addresses a common bottleneck in the diffusion model training process, offering a practical solution that can be applied across various applications, from image generation to text-to-image conversion. Repository: Ting-Justin-Jiang/sada-icml GitHub stars: 44 GitHub description: [ICML 2025] Official Repo for Stability-guided Adaptive Diffusion Acceleration. 🚀🌙Accelerating off-the-shelf diffusion model with a unified

Overview: Grok Imagine Video 1.5 API is a powerful tool for generating high-quality videos from text or images, offering flexible pricing and easy integration with OpenAI-compatible endpoints. Key features: - Supports text-to-video and image-to-video conversion - Compatible with 300+ models, including Grok images and videos - Billing by generated second, with a base price of $0.064 per second - Asynchronous processing with task ID polling - No need for monthly subscriptions, only pay for usage - Real-time billing and cost reporting - Supports multiple resolutions and aspect ratios Tech stack: Python Why it matters: This API streamlines the process of generating videos from text or images, making it ideal for content creators, marketers, and anyone looking to add dynamic visuals to their projects without the complexity of building custom solutions. Repository: a-2y29hit/grok-imagine-video-1-5-api-id Primary language: Python GitHub stars: 37 GitHub description: Grok Imagine Video 1.5 API (`grok-imagine-video-1.5`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Diffusion Handles is a training-free method that enables 3D-aware image edits using a pre-trained Diffusion Model. Key features: - 3D-aware image editing - Pre-trained Diffusion Model for training-free approach Tech stack: Python Why it matters: This method simplifies 3D image editing by leveraging a pre-trained Diffusion Model, making it accessible without the need for extensive training.