Projects

### Overview: Captain is a tool designed to give your computer an AI brain, making it easy to use AI applications directly without needing an internet connection. It allows you to start applications quickly, generate images, search data, and build your own AI apps. ### Key Features: - **Start applications quickly**: Similar to using Spotlight on a Mac, you can start different AI apps with a few keystrokes. - **Use core apps**: Generate images and turn them into stories using AI. - **Use AI without the internet**: Run AI models on your computer without needing to connect to the internet. - **Build your own AI apps**: If you know how to make a website, you can easily create AI applications. - **Search your data easily**: Find documents and data quickly with natural language search commands. ### Tech Stack: TypeScript ### Why It Matters: - **Privacy**: 100% of your data stays on your computer. We don’t see or store any of your data. - **Easy to use**: If you can use a computer, you can use Captain. It’s that simple. - **Free to start**: Get started for free and explore what you can do with AI. ### Repository: blib-la/captain ### GitHub Stars: 61 ### GitHub Description: Give your computer an AI brain ### Screenshots: ![](https://github.com/blib-la/captain/assets/1148334/5413a0fb-d051-4ef9-b11e-935

### Overview: This repository contains scripts and instructions for training acoustic models using Kaldi in Brazilian Portuguese (BrPort). It includes scripts for speech recognition, phonetic alignment, and speaker diarization, as well as text and speech datasets. ### Key Features: - **Speech Recognition (Vosk + LapsBM):** Scripts for training Kaldi models on Brazilian Portuguese datasets. - **Speech Recognition (Vosk + Datasets):** Scripts for training Kaldi models on public speech datasets in Portuguese, including Brazilian and general Portuguese. - **Phonetic Alignment (Gentle):** Scripts for training Kaldi models for phonetic alignment using the Gentle system. - **Phonetic Alignment (UFPAlign):** Scripts for training Kaldi models for phonetic alignment using the UFPAlign system. - **Speaker Diarization (CallHome):** Scripts for speaker diarization using the CallHome system, which is pre-trained on English data for inference only. ### Tech Stack: - Shell Scripts ### Why It Matters: These scripts and instructions are essential for researchers and developers working on speech recognition, phonetic alignment, and speaker diarization tasks in Brazilian Portuguese. They provide a comprehensive set of tools and resources to facilitate the development and deployment of speech recognition models in the Brazilian language. ### Repository: Repository: falabrasil/kaldi-br GitHub Stars: 61 GitHub Description: ☕🇧🇷 Scripts para o Kaldi em Português Brasileiro README Excerpt: This repo contains instructions and scripts to train acoustic models using K

Overview: Este projeto é uma implementação de uma ferramenta semelhante ao ChatPDF utilizando o modelo open-source LLAMA V2 (Meta) para conversar com PDFs, com base na primeira versão do projeto. Key features: - Conversação com PDFs - Utilização do LLAMA V2 como LLM - Diagrama de solução Tech stack: Langchain, ChromaDB, Gradio, LLAMA V2 Why it matters: Este projeto serve como um exemplo de como se pode criar uma ferramenta de conversação com PDFs utilizando uma LLM open-source, oferecendo uma maneira interessante de interagir com documentos digitais.

Overview: This project is a GPT Image 2.5 API with support for 4K and 2K resolutions, priced at $0.021 per image for 4K, $0.014 for 2K, and $0.0085 for 1K. It offers OpenAI-compatible endpoints and is compatible with billing models. Key features: - Support for 4K and 2K resolutions - Pricing: $0.021 per image for 4K, $0.014 for 2K, $0.0085 for 1K - OpenAI-compatible endpoint - Billing model: $0.0085 per image (flare@1K, sunburst@1K) Tech stack: Python Why it matters: This API provides a cost-effective way to generate high-resolution images, making it ideal for businesses looking to enhance their visual content without the need for a subscription. Repository: gkdh0uepxud4b/gptimage2.5-gpt-image-2.5-4k Primary language: Python GitHub stars: 61 GitHub description: gpt-image-2.5-ext API (gptimage2.5 / gpt image 2.5 ext): 4K $0.021 per image; 2K $0.014; 1K $0.0085. Model id, 4k settings, curl and Python examples over an Open

Overview: This repository contains covering grammars for English and Russian text normalization, as detailed in Gorman and Sproat's 2016 and Ng et al.'s 2017 publications. Key features: - Covering grammars for English text normalization - Covering grammars for Russian text normalization Tech stack: Thrax Why it matters: These grammars provide a minimally supervised approach to text normalization, making them useful for applications requiring efficient and accurate text processing.

### Overview: A real-time language translator tool that uses voice recognition to translate spoken words into the target language, providing a convenient and efficient way to bridge language gaps. ### Key features: - Real-time language translation - Voice recognition integration - Streamlit UI for easy setup and usage ### Tech stack: Python (v3.8.5), GTTS Module, SpeechRecognition Module, Streamlit UI Module, Pygame Module, Googletrans (v3.1.0a0) ### Why it matters: This tool offers a natural and fluid translation experience, replicating the accuracy and efficiency of human translators, making it a valuable asset for bridging language barriers in various applications. ### Repository: - **Repository:** gunarakulangunaretnam/real-time-language-translator - **GitHub stars:** 61 - **GitHub description:** A voice recognition-based tool for translating languages in real-time. - **README excerpt:** - A voice recognition-based tool for translating languages in real-time. This tool serves as a virtual interpreter, offering users a convenient and efficient way to bridge language gaps. Inspired by the natural process of human translation, the tool listens to spoken words and converts them into the target language, replicating the fluidity and accuracy of a human translator. -  - **System Architeture Diagram:**  - **Setup Steps:** - Navigate to the following directory

### Inference #### Download Pre-trained Weights The pre-trained weights are stored on [InstaDeepAI HuggingFace](https://huggingface.co/InstaDeepAI). Two models are provided: `denovo.pth` and `inpainting.pth`. Please download them using the following command: ```bash # Make sure you have git-lfs installed git lfs clone https://github.com/InstaDeepAI/FrameDiPT.git cd FrameDiPT python download_weights.py ``` #### Inference Config The inference configuration is provided in the `config` directory. This includes parameters for the diffusion process, such as the number of steps and the learning rate. #### Inference Outputs The inference outputs are stored in the `outputs` directory. The results include the predicted protein structure and any additional information generated during the inference process. #### Generate Full-atom Model The full-atom model is generated by combining the predicted backbone structure with the atomic coordinates from the original protein structure. This model can be further refined and validated using additional tools and techniques. ## Evaluation ### Inpainting Model Evaluation on TCR The model has been evaluated on TCR (T-cell receptor) datasets, demonstrating its effectiveness in inpainting missing or damaged parts of protein structures. ### De novo Protein Design Evaluation The model has also been evaluated for de novo protein design, where it has shown promising results in generating new protein structures from scratch. ## Step-by-step Tutorial for Paper Reproduction The tutorial provides a detailed guide on how to reproduce the results presented

### Overview: GPIC is a Giant Permissive Image Corpus consisting of approximately 28 trillion pixels, providing a benchmark for scalable visual generative modeling. It includes diverse internet images captioned by a state-of-the-art vision-language model, with 100M training, 200K validation, and 1M test examples, and is permissively licensed for research and commercial use. The dataset is safety-filtered, deduplicated, and hosted on Hugging Face, making it accessible and reproducible for open research. ### Key Features: - **GPIC Dataset Overview**: Contains approximately 28 trillion pixels, with diverse internet images captioned by a state-of-the-art vision-language model. - **Training Data**: 100M training examples. - **Validation Data**: 200K validation examples. - **Testing Data**: 1M test examples. - **Permissive Licensing**: Permissively licensed for both research and commercial use. - **Safety Filtering**: Safety-filtered to ensure high-quality images. - **Central Hosting**: Centrally hosted on Hugging Face for easy access and reproducibility. ### Tech Stack: Python ### Why it Matters: GPIC serves as a crucial benchmark for scalable visual generative modeling, providing researchers with a large and stable dataset to test and improve their models. Its diverse and high-quality images make it an essential resource for advancing the field of visual generation.

### Overview: This project allows developers to pre-train Wav2vec 2.0 on their own datasets, making it easy to integrate into their workflows. ### Key features: - Train on your own local datasets. - Visualization with Tensorboard. - Resume training from the latest checkpoint. - No data caching. Data is loaded on the fly. - Separate checkpoint saving at each epoch. ### Tech stack: Python ### Why it matters: This project streamlines the pre-training process, making it accessible and efficient for developers to integrate into their workflows without requiring extensive setup or data caching.

Overview: This project provides scripts and tools for generating outpainting images using DALL-E 2, focusing on 1912x1912 and 2868x2868 square outpainting layers. Key features: - Scripts for DALL-E 2 outpainting - 1912x1912 and 2868x2868 square outpainting layers - Shell script for iterative outpainting process - Illustration of outpainting steps - Output files: stage1.png and final.png - Jupyter notebook for DALL-E 2 zoom outpainting Tech stack: Python, Jupyter Notebook, Docker, OpenAI API, Imagemagick Why it matters: This project offers a practical solution for generating high-quality outpainting images using DALL-E 2, which can be particularly useful in creative and artistic applications. Repository: mdbecker/dalle2_outpaint Primary language: Jupyter Notebook GitHub stars: 61 GitHub description: Scripts/tools for dalle2 outpainting GitHub URL: https://github.com/mdbecker/dalle2_outpaint

Overview: This repository contains the source code for "Learning to Modulate pre-trained Models in RL," which was accepted at NeurIPS 2023. It supports training Decision Transformer models on multiple domains including Meta-World, Continual-World, Atari, Gym-MuJoCo, ProcGen, and DMControl. Key features: - Support for training Decision Transformer models online or from offline datasets on various domains. - Open-source dependencies including PyTorch, Huggingface transformers, stable-baselines3, wandb, and Hydra. Tech stack: Python Why it matters: This project offers a practical approach to modulating pre-trained models in reinforcement learning, demonstrating its utility in various domains through its implementation on diverse environments. GitHub stars: 61 Repository: ml-jku/L2M

### Overview: This repository is a curated list of active submissions on OpenReview related to diffusion models for ICLR 2025. It covers a wide range of applications such as 3D reconstruction, image-to-3D conversion, and more. ### Key Features: - **3D Diffusion**: Improving Multi-view Diffusion Model with Synthetic Data - **Personalized T2I**: View Interpolation using Multi-view Diffusion for 3D Mesh Reconstruction - **Super-Resolution**: Enhancing Image Quality with Diffusion Models - **Concept Eraser**: Removing Conceptual Elements from Images - **Adversarial Diffusion**: Generating Realistic Images with Adversarial Training - **Video Generation**: Creating Videos from Textual Descriptions - **Distilled Learning**: Simplifying Models for Better Performance - **3D-Adapter**: Enhancing Geometry Consistency in Multi-View Diffusion - **HouseCrafter**: Transforming Floorplans into 3D Scenes - **ComPC**: Completing 3D Point Clouds with 2D Diffusion Priors - **HOI-Diff**: Synthesizing 3D Human-Object Interactions - **NEXT-MOL**: Generating 3D Molecules from 1D Language Models - **SeaLion**: Semantic Part-Aware Latent Point Diffusion Models for 3D Generation - **PSHead**: 3D Head Reconstruction from a Single Image - **MVLight**: Relightable Text-to-3D Generation - **

### Overview: OrcaS 文档预览插件是一款纯本地版的多模态文档预览工具，适用于Office、WPS、iWork、PDF、CAD、代码文档、图片、视频、音频、压缩包等多种文件格式的预览。它提供了跨平台支持，效果和兼容性更好，可自行适配存储。 ### Key features: - 支持Office(docx/xlsx/pptx)、WPS、iWork、PDF、CAD、代码文档、图片、视频、音频、压缩包等多种文件格式的预览。 - 跨平台支持，适用于多种操作系统。 - 支持自动适配存储，提高预览效果和兼容性。 ### Tech stack: Go ### Why it matters: OrcaS 文档预览插件提供了一个高效、本地化的文档预览解决方案，适用于多种文件格式和操作系统，特别适合需要离线预览和跨平台兼容性的场景。

Overview: SphereDiff is an official PyTorch implementation that enables the generation of 360° panoramic images and videos without the need for tuning, using pretrained diffusion models. Key features: - Spherical latent representation for distortion-free 360° generation - Supports both static and dynamic panoramas - Plug-and-play with existing pretrained diffusion models - No additional fine-tuning required Tech stack: Python Why it matters: SphereDiff simplifies the process of generating 360° panoramic images and videos by eliminating the need for tuning, making it more accessible and efficient for developers and researchers alike.

Overview: JARVIS AI Assistant is a virtual assistant project inspired by Tony Stark's JARVIS, featuring voice-based interaction, AI chat, web browsing, and more. Key features: - Voice-based interaction using speech recognition. - AI-powered chat with OpenAI's language model. - Web browsing capabilities to open websites. - Music playback functionality. - Displaying the current time. Tech stack: Python Why it matters: This project aims to provide a practical and engaging AI assistant experience, inspired by the advanced technology of JARVIS. Repository: rajkishorbgp/JARVIS-AI-Assistant GitHub stars: 61

Overview: TUFF is a local AI app built for Macs, designed to run large language models on Apple Silicon, including models far bigger than your Mac's memory. It uses a streaming MoE model to keep the experts on the SSD and read only the necessary data, making it efficient and fast. Key features: - Runs models far bigger than your Mac's memory. - Supports chat, web and file search, and background API. - Open source, no data leaves your Mac unless you enable web search. - Supports various models like Gemma, Qwen, and Qwen3.8 Flash Next. Tech stack: Swift, Metal. Why it matters: TUFF is a powerful tool for Mac users looking to run large language models efficiently, making it possible to use models that far exceed the capabilities of your Mac's memory. Repository: rexmhall09/TUFF Primary language: Swift GitHub stars: 61 GitHub description: The local AI app built for the Mac. Run LLMs like Qwen, Gemma and GPT-OSS on Apple Silicon, including models far bigger than your Mac's memory, by streaming MoE experts from the SSD. GitHub description excerpt: The local AI app built for the Mac. Runs models far bigger than your Mac's memory.

Overview: `pandas-gpt` is a Python library designed to enhance data science workflows by integrating ChatGPT or other large language models with pandas DataFrames. Key features: - Data transformation using ChatGPT - Querying and manipulating data - Plotting functionalities - Custom language models for specific requests Tech stack: Python with Jupyter Notebook Why it matters: It simplifies complex data science tasks by leveraging the power of large language models, making data analysis more accessible and efficient.

### Overview: This repository contains a validated GLM-5.3 Flash recipe tailored for two NVIDIA RTX PRO 6000 Blackwell 96GB GPUs, optimized for 262K context and 16-image prompts. The recipe includes adaptive MTP, tools, and vision, and is designed for high-throughput DFlash2 speculative decoding. ### Key features: - GLM-5.3 Flash recipe - 2x RTX PRO 6000 Blackwell 96GB GPUs - 262K context - 16-image prompts - Adaptive MTP - Tools and vision ### Tech stack: Python ### Why it matters: This recipe is specifically designed for high-throughput decoding on two RTX PRO 6000 Blackwell GPUs, providing optimal performance and resource utilization. The validated recipe ensures compatibility with OpenAI's GLM-5.3 model and supports high-resolution image prompts, making it suitable for applications requiring high-quality image generation and processing. ### GitHub stars: 61 The repository has gained 61 GitHub stars, indicating its popularity and usefulness among the developer community.

### Overview: This ComfyUI custom node allows for merging **MOSS-Image Base** and **SFT** with reproducible presets or full per-layer control. The node supports selective inheritance of SFT's strengths while retaining Base's composition and natural realism. ### Key features: - One-click presets for **Base-preserve**, **Balanced**, and **SFT-detail**. - Full per-layer control over all 42 MOSS-Image DiT blocks. - Live `merge_info` readout showing effective weights after `strength` scaling. - No extra Python dependencies beyond ComfyUI itself. - Native **int8c / convrot Base + SFT** merging support. - Compatible with the native ComfyUI MOSS-Image loader and sampling chain. - Pairs with ComfyUI's built-in **ModelSave** and model-only LoRA loader nodes. ### Tech stack: Python ### Why it matters: This node simplifies the process of combining different styles in ComfyUI, allowing for selective inheritance of strengths while maintaining the overall composition and realism of the base model. It provides a straightforward way to blend different styles without the need for complex configuration or additional dependencies.

### Overview: This repository contains the implementation of Google's DreamBooth algorithm, which is used to fine-tune the Stable Diffusion model. The repository also provides instructions on how to train a model on a vast.ai instance. ### Key features: - DreamBooth fine-tuning of Stable Diffusion model - Instructions for training on vast.ai instance - Pre-trained Stable Diffusion models available on HuggingFace - Generation of regularization images for better editability - Detailed configuration in `configs/stable-diffusion/v1-finetune_unfrozen.yaml` ### Tech stack: - Jupyter Notebook ### Why it matters: DreamBooth is a powerful tool for enhancing the ability of Stable Diffusion to generate realistic images, especially for tasks like editing and fine-tuning. The implementation provided in this repository offers a streamlined approach to fine-tuning Stable Diffusion, making it easier for developers to integrate into their projects.