Projects

Overview: The repository contains 8 tested FLUX 2 Pro prompts, each featuring an example render, aspect ratio, and the cost per image. These prompts are designed to help developers create high-quality images efficiently. Key features: - 8 FLUX 2 Pro prompts - Example renders, aspect ratios, and cost per image - Model: `flux-2-pro` - Effective price used: $0.024 per image - $1 minimum, pay-as-you-go pricing - Asynchronous processing Tech stack: Python Why it matters: These prompts provide a straightforward way for developers to create professional-looking images quickly and efficiently without the need for expensive subscriptions or upfront payments. Repository: a60au8/flux2-prompts-fr Primary language: Python GitHub stars: 47

Overview: This repository contains the official implementation of the NeurIPS 2023 paper "Language-driven Scene Synthesis using Multi-conditional Diffusion Model." The project includes datasets, model checkpoints, and training/testing scripts for reproducing the results described in the paper. Key features: - Language-driven Scene Synthesis using Multi-conditional Diffusion Model - PRO-teXt dataset for indoor scene synthesis - HUMANISE dataset for training and testing - PyTorch-based implementation - GitHub stars: 47 Tech stack: Python Why it matters: This project provides a comprehensive resource for researchers and developers interested in language-driven scene synthesis, enabling the replication and extension of the NeurIPS 2023 paper's findings.

Overview: 听记 (Tingji) 是一个本地会议录音转写与纪要生成工具，支持 FunASR 和大型语言模型。它能够自动转写录音、整理会议纪要，并且所有数据都存储在用户的本地电脑上。用户可以手动调整说话人归属，查看发言时间轴和时间戳，以及对会议纪要进行笔记和回放。 Key features: - 说话人自动区分 - 发言人时间轴 - 逐句时间戳 - 纪要时间戳回跳 - 个人笔记层 - 议题时间轴 - 原文↔整理对照 - 批量上传 - 一键整理 - 总结模板 - 实时日志 - LLM 自由选 - 全 Web 配置 - 导出多种格式 - 实时流式转写 - 历史会议管理 - 纪要二次编辑 - 卡死任务恢复 - 本地运行 Tech stack: Python Why it matters: 听记 (Tingji) 是一个本地会议录音转写与纪要生成工具，它能够自动转写录音、整理会议纪要，并且所有数据都存储在用户的本地电脑上。用户可以手动调整说话人归属，查看发言时间轴和时间戳，以及对会议纪要进行笔记和回放。 Repository: baigong-ai/Tingji

Overview: ImageNet-D is a novel benchmark designed to test the robustness of neural networks on synthetic images, surpassing previous synthetic test sets with more varied and realistic images generated using diffusion models. This benchmark is crucial for evaluating the effectiveness of vision and vision-language models in real-world scenarios. Key features: - Utilizes diffusion models to generate realistic synthetic images. - Benchmarks the accuracy of various vision and vision-language models on ImageNet-D. - Provides a comprehensive dataset with three categories: background, texture, and material. - Includes evaluation scripts for large VQA models like LLaVa. Tech stack: Python Why it matters: ImageNet-D significantly enhances the evaluation of neural network robustness by offering more realistic and varied synthetic images, thereby providing a more accurate and relevant benchmark for real-world applications.

### Overview: This course explores Large Language Models (LLMs), from the basics to cutting-edge research. It covers topics such as motivation and context, AI goals and definitions, LLM architectures, training paradigms, test-time computation, reasoning, safety and robustness, representations, and interpretability. ### Key features: - **LLM Architectures**: Explains the basics of LLMs, including attention and transformers. - **Training Paradigms**: Discusses pre- and post-training, supervised fine-tuning, and learning from preferences. - **Test-Time Computation**: Covers sampling, prompting, and reasoning. - **Representations**: Discusses the representations in deep neural networks and the intrinsic representation of LLM hallucinations. - **AI Control**: Talks about protocols and methods for deploying untrusted AI models. - **Adversarial Reasoning**: Explains adversarial reasoning in LLMs. ### Tech stack: Python, Markdown, LaTeX. ### Why it matters: This course provides a comprehensive understanding of LLMs, from their foundational concepts to advanced research, making it essential for anyone interested in AI and machine learning.

### Overview: This collection features 252 curated prompts for GPT Image 2, organized into 9 categories, with preview images for each. ### Key Features: - **Portrait & Avatar** - **No. 1:** Editorial Studio Portrait - **No. 2:** Candid Street Portrait - **No. 3:** Corporate Headshot - **Product & E-commerce** - **No. 4:** Product Detail - **No. 5:** Product Close-Up - **No. 6:** Product Label - **Poster & Typography** - **No. 7:** Poster Text - **No. 8:** Typography Illustration - **Infographic & Diagram** - **No. 9:** Infographic Text - **No. 10:** Diagram Caption - **Anime & Illustration** - **No. 11:** Anime Character - **No. 12:** Illustration Text - **3D & Material** - **No. 13:** 3D Text - **No. 14:** Material Illustration - **Food & Lifestyle** - **No. 15:** Food Label - **No. 16:** Lifestyle Illustration - **Architecture & Interior** - **No. 17:** Interior Text - **No. 18:** Architecture Illustration - **Nature & Landscape** - **No. 19

Overview: This project utilizes Baidu's ASR and TTS APIs to provide a comprehensive solution for automatic speech recognition and text-to-speech functionalities. It includes a class diagram, interface and class structure, and a user-friendly interface for file recognition and text-to-speech conversion. Key features: - ASR (Automatic Speech Recognition) - TTS (Text To Speech) Tech stack: C# Why it matters: This project offers a practical tool for developers looking to integrate speech recognition and synthesis functionalities into their applications, leveraging Baidu's powerful APIs for enhanced user experience.

Overview: This repository contains the source code for the book "Make Python Talk," providing a comprehensive guide to Python programming, including various modules and packages for speech recognition, text-to-speech, and more. Key features: - Speech recognition and text-to-speech functionalities - Custom package mptpkg for consistent code across chapters - Modules for audio playback and stock market information - Custom-made package mptpkg for easy integration and consistency Tech stack: Python Why it matters: This repository serves as a practical resource for Python developers, offering a structured approach to integrating various functionalities into their projects, making it easier to maintain and update code across different chapters.

Overview: This project involves the development of a hashtag prediction model using PyTorch, specifically utilizing ALBERT for text processing and VGG16 for image feature extraction. The model is trained on Instagram data and evaluated on a second-place finish in the OpenResource Hackathon 2019. The project includes a Dockerized server for running the model locally and a Google Drive integration for uploading images. Key features: - Multimodal hashtag prediction - Utilization of ALBERT and VGG16 - Prediction of 100 labels Tech stack: Python Why it matters: This project showcases the potential of combining different machine learning models to enhance predictive capabilities in natural language processing and image analysis, demonstrating the effectiveness of PyTorch and its integration with other deep learning frameworks.

Overview: This project is a powerful forced aligner designed specifically for conversational speech, providing accurate timing for natural speech with minimal errors. Key features: - Handles vocal sounds and filler words like `um` and `laughter` as distinct units. - Aligns transcripts from CrisperWhisper, ensuring precise timing. - Achieves a 19.9 ms mean word-boundary error on conversational English, significantly better than state-of-the-art systems. - Robust against noise and maintains alignment even at low signal-to-noise ratios. - Fast and efficient, aligning up to eight minutes of audio in just three seconds. Tech stack: Python Why it matters: This tool is essential for accurately timing speech data from real conversations, making it invaluable for speech recognition, language modeling, and other applications that require precise timing of words and phonemes. Repository: nyrahealth/nyra-forced-aligner GitHub stars: 47

### Overview: PRG (Pretrained Reversible Generation) is a novel approach to unsupervised visual representation learning that leverages pretrained continuous time flow diffusion generators to produce multi-level features. These features are then fine-tuned to serve as an unsupervised representation extractor, offering a unique method for unsupervised representation learning. ### Key Features: - **Practical Application**: PRG is designed to be easily integrated into existing projects, making it accessible for developers. - **Unsupervised Representation Learning**: By turning a pretrained continuous time flow diffusion generator upside down, PRG generates multi-level features that can be fine-tuned for unsupervised representation learning. - **Versatile Usage**: The codebase is adaptable to various training stages, allowing for flexibility in the training process. ### Tech Stack: - **Python 3.10**: The primary programming language used for development. - **Conda**: Recommended for managing Python environments. - **Accelerate**: Enables efficient execution of the codebase. ### Why It Matters: PRG offers a novel and practical solution for unsupervised representation learning, providing a unique method for generating features that can be fine-tuned for unsupervised representation learning. This approach can be particularly useful in applications where labeled data is scarce or unavailable. ### Repository: opendilab/PRG ### GitHub Stars: 47 ### GitHub Description: [ICCV 2025] Pretrained Reversible Generation as Unsupervised Visual Representation Learning

Overview: WorkflowGenerator 是一个开源的画布设计平台，允许用户直接通过 Agent 操控无限画布，添加文本、图片、视频、音频等元素，并通过提示词和参考素材生成创作任务。它支持与 Zodiac 对话、审核计划并运行创作任务，适用于桌面和网页端。 Key features: - 无限画布组织提示词、参考素材和生成结果 - 与 Zodiac 对话、审核计划并运行创作任务 - 可以从头开始制作故事、改编小说或使用已有剧本 - 生成分集、剧本和分镜画面，支持 MP4 和字幕导出 - 保留 Skills 会话、计划和工作文件，支持备份和恢复 - 提供网页版尝鲜体验，但不支持文件和脚本操作 Tech stack: WorkflowGenerator 使用 TypeScript 编写，采用 React 和 Rust 本地服务构建，支持 macOS 和 Web 端。 Why it matters: WorkflowGenerator 提供了灵活的画布设计和创作工具，结合了无限画布和与 Zodiac 的实时对话能力，使得用户能够高效地创作和分享他们的作品。

### Overview: Rupixel is a pixel-native visual retrieval system that searches documents based on their visual content, not just text. It uses screenshots of pages to find relevant documents, especially when text is obscured by images, scans, or charts. The system is designed to be fast and efficient, running entirely in the browser without the need for installation or server setup. It includes a real-time video search feature and supports optional live captions for better understanding. ### Key Features: - **Real-time Video Search:** Searches documents by capturing and analyzing live video or screen captures. - **Visual Search:** Searches documents by their visual content, including screenshots of pages. - **Text Search:** Searches documents by their text content. - **Built-in Proxy:** A server-side proxy can be used to manage shared or managed keys, ensuring the key is never stored in the browser. ### Tech Stack: Rust ### Why It Matters: Rupixel offers a unique approach to document retrieval by focusing on visual content, making it particularly useful for documents where text is obscured or difficult to read. This system leverages the power of visual content to find relevant documents, making it a valuable tool in various fields such as archives, research, and more. ### GitHub Stars: 46 ### Repository: ruvnet/rupixel

### Overview: This project is a code repository for the paper "Finetuning Text-to-Image Diffusion Models for Fairness" presented at ICLR 2024. It includes the necessary code and checkpoints to replicate the experiments detailed in the paper. ### Key features: - Finetuning LoRA on text encoder and/or U-Net to debias binary gender to a perfectly balanced distribution. - Finetuning prompt prefix to debias binary gender to a perfectly balanced distribution. - Finetuning LoRA on text encoder and/or U-Net to jointly debias binary gender and four classes of race, to a perfectly balanced distribution. - Finetuning LoRA on text encoder and/or U-Net to jointly debias binary gender and four classes of race, to a perfectly balanced distribution while controlling age distribution. - Finetuning LoRA on text encoder and/or U-Net to jointly debias binary gender and four classes of race, to a perfectly balanced distribution. - Finetuning LoRA on text encoder and/or U-Net to jointly debias binary gender and four classes of race, to a perfectly balanced distribution. ### Tech stack: Python ### Why it matters: This project provides a practical implementation of the debiasing techniques discussed in the paper, allowing developers to evaluate the fairness of text-to-image diffusion models in a controlled environment.

### Overview: SIFT-Video is a semantic video search engine that indexes audio and visual content using natural language queries. It leverages pre-trained models for efficient similarity search at query time. ### Key features: - **Multimodal Semantic Indexing:** SIFT-Video indexes both audio and visual content for enhanced search accuracy. - **Offline Processing:** Videos are processed offline, converting audio and frames into semantic embeddings. - **Qdrant Vector Database:** Indexes are stored in a Qdrant vector database for fast similarity search. - **Natural Language Queries:** Users can search videos using natural language queries, returning relevant timestamps. ### Tech stack: Rust, Docker, Qdrant ### Why it matters: SIFT-Video provides a powerful tool for semantic video search, enabling users to find specific moments within videos using natural language queries. Its offline processing and efficient indexing make it highly scalable and effective for large video collections.

### Overview: rkvoice-stream is a Rockchip NPU speech engine designed for the OpenVoiceStream production voice AI service. It supports ASR and TTS on-device with RKNN/RKLLM acceleration, no cloud, no GPU, and no internet required. Use it as a Python library or ship it via OpenVoiceStream's one-command installer. ### Key features: - **ASR (Automatic Speech Recognition)**: Supports five backends: Qwen3-ASR (NPU), Paraformer (Hybrid), Paraformer (CPU), SenseVoice (NPU), and SenseVoice (CPU). The RK3576 RTF and RK3588 RTF times are 0.36 and 0.23 seconds, respectively, while the final ASR times are 6.5s and 3.7s, respectively. - **TTS (Text-to-Speech)**: Supports six backends: Qwen3-TTS (NPU), Matcha-TTS (NPU), Qwen3-TTS (CPU), Piper-TTS (NPU), and Kokoro-TTS (NPU). The RK3576 RTF and RK3588 RTF times are 120ms and 700ms, respectively, with final TTS times of 6.5s and 3.7s, respectively. ### Tech stack: Python ### Why it matters: rkvoice-stream simplifies the deployment of voice AI services on Rockchip NPU devices, enabling real-time speech

### Overview: Live Translator is a real-time audio translation system for macOS, enabling you to translate any audio playing on your Mac into multiple languages in real-time. It uses a live document model to maintain context and context-aware AI to produce natural translations. ### Key Features: - **Real-time translation** of any audio on your Mac. - **11 source languages** (English, German, French, Spanish, Italian, Japanese, Chinese, Korean, Russian, Arabic, Portuguese). - **12 target languages** (translate into any supported language). - **Context-aware AI** to maintain full conversation context. - **Live document model** for translation growth. - **Text-to-Speech** with Piper or OpenAI/Gemini voices. - **Multi-provider** support (OpenAI, Google Gemini, Anthropic Claude). - **Latest AI models** (GPT-5.5, GPT-5.4 mini/nano, Gemini 2.5 / 3.5 Flash, Claude Haiku 4.5 / Sonnet 5 / Opus 5). - **Bounded context** to keep cost and latency flat. - **Floating overlay** with dark-themed panel. - **In-app settings** for provider, API keys, model, TTS, languages. - **No audio drivers needed**; uses ScreenCaptureKit. - **On-device STT** (default) for speech recognition. ### Tech Stack: - Python ### Why It Matters: Live Translator offers a seamless, real-time translation experience for audio content on macOS, making it

### Overview: CityGen is a novel framework for infinite, controllable, and diverse 3D city layout generation. It uses an auto-regressive out-painting pipeline to extend local layouts to infinite city layouts and a multi-scale diffusion model to generate controllable semantic fields. ### Key Features: - Infinite and controllable 3D city layout generation. - Auto-regressive out-painting pipeline for extending local layouts. - Multi-scale diffusion model for generating controllable semantic fields. - Iterative refinement of the global semantic field to achieve nuanced and polished results. ### Tech Stack: CityGen is built using Python and utilizes libraries such as PyTorch and TensorFlow for the implementation. ### Why it Matters: CityGen offers a novel approach to 3D city layout generation, providing infinite and controllable layouts that can be used for various applications, including urban planning, architecture, and virtual reality environments.

### Overview: GISTEmbed is an innovative framework designed to dynamically select training negatives for text embedding fine-tuning, ensuring higher quality training samples. This approach addresses common challenges in contrastive learning, leading to more robust and generalizable embeddings. ### Key Features: - **Dynamic Negative Sampling**: GISTEmbed selects training negatives based on relevance, ensuring only relevant examples are used. - **In-Sample Selection**: Focuses on in-sample selection, improving the quality of training negatives. - **Fine-Tuning of Smaller Models**: Suitable for fine-tuning smaller models, leading to notable improvements across various NLP tasks. ### Tech Stack: - **Python**: The primary programming language used for implementing GISTEmbed. ### Why It Matters: - **Improved Embedding Quality**: By focusing on in-sample selection, GISTEmbed ensures higher quality training negatives, leading to more robust and generalizable embeddings. - **Efficient Fine-Tuning**: Suitable for fine-tuning smaller models, making it particularly beneficial for resource-constrained environments. - **Robustness and Generalization**: The framework addresses common challenges in contrastive learning, leading to better performance across various NLP tasks. ### Repository: - **Repository**: [worldbank/GISTEmbed](https://github.com/avsolatorio/GISTEmbed) - **GitHub Stars**: 47 - **GitHub Description**: GISTEmbed: Guided In-sample Selection of Training Negatives for Text Embeddings ### Primary Language: - **Primary Language**: Python ### Data: - **Dataset**: Compilation of the MEDI dataset and the M

### Overview: VoxFlow is a macOS, Windows, and iOS voice, screenshot, clipboard, and coding-agent workbench designed for developers and creatives. It integrates seamlessly into your existing workflows, allowing you to dictate, capture screenshots, and use local coding-agent instructions directly within your applications. ### Key Features: - **Voice Input**: Dictate text directly into your applications. - **Screenshot Capture**: Quickly capture screenshots and OCR them. - **Clipboard Integration**: Paste clipboard items directly into your work. - **Coding Agent Commands**: Execute local coding-agent instructions. - **Local Workbench**: Access a local workbench for system-wide dictation workflows. ### Tech Stack: Swift ### Why It Matters: VoxFlow simplifies the process of dictating and capturing information, making it easier and more efficient for developers and creatives to work on their applications without the distraction of traditional keyboard input. The local workbench ensures that all your work is stored locally, providing a secure and efficient environment for your projects. Repository: xingbofeng/VoxFlow Primary language: Swift GitHub stars: 47 GitHub description: A macOS voice, screenshot, clipboard, and coding-agent workbench. GitHub description excerpt: "Speak to write on all three platforms. The desktop editions add system-wide dictation workflows; macOS also includes the full local context workbench."