Projects

### Overview: This repository contains a set of nodes designed to prepare noise predictions before the CFG function in ComfyUI, enhancing the quality and compatibility of the generation process. ### Key features: - **Automatic CFG Scaling**: Applies predictable scaling to improve consistency across different models. - **Perp-Neg Logic**: Implements the perp-neg logic for better noise reduction. - **Sharpening**: Subtracts the previous step's noise to make images sharper. - **Exponentiation**: Exponentiates the noise to enhance contrast. ### Tech stack: Python ### Why it matters: These nodes help streamline the workflow by preparing noise predictions correctly, ensuring better results and compatibility with various models.

Overview: This project introduces FantasyTalking2, a novel framework for aligning audio-driven portrait animation models with fine-grained, multidimensional preferences. The key features include the introduction of Talking-Critic, a multimodal reward model, and the creation of Talking-NSQ, a large-scale multidimensional human preference dataset. The proposed TLPO framework decouples preferences into specialized expert modules, enabling comprehensive, fine-grained enhancement across all dimensions without mutual interference. Key features: - Talking-Critic: A multimodal reward model for quantifying human-aligned reward functions. - Talking-NSQ: A large-scale multidimensional human preference dataset containing 410K preference pairs. - TLPO: A novel framework for aligning diffusion-based portrait animation models with fine-grained, multidimensional preferences. Tech stack: Python, PyTorch, TensorFlow, and other related libraries. Why it matters: This project addresses the challenges of optimizing across competing preference objectives in audio-driven portrait animation, offering significant improvements in lip-sync accuracy, motion naturalness, and visual quality.

### Overview: The `genQC` repository is a project focused on generating quantum circuits using diffusion models. It provides a comprehensive library for sampling pre-trained diffusion models and fine-tuning them for various quantum tasks. ### Key features: - **Generative Quantum Circuits**: The repository allows the creation of quantum circuits using diffusion models. - **Minimal Example**: A simple example demonstrating how to compile and sample the 4-qubit Quantum Fourier Transform (QFT) unitary using parameterized circuits. - **Documentation**: The project page provides detailed documentation and examples for using the library. ### Tech stack: The repository uses Jupyter Notebook as the primary language for development and documentation. ### Why it matters: This project is significant as it leverages diffusion models to generate quantum circuits, which can be useful for various quantum computing applications, including quantum chemistry, quantum cryptography, and quantum simulation. The ability to fine-tune models from scratch and sample pre-trained weights makes it a valuable tool for researchers and developers in the field of quantum computing. ### Repository: The code repository is maintained on GitHub, with 66 GitHub stars. It includes a full release of the used diffusion pipeline, examples and tutorials, and source notebooks for further exploration.

Overview: This is a major update of my project JARVIS-The-Ultimate-Project, featuring improved speech recognition, a custom Launchpad interface, and enhanced multimedia player functionality. The project aims to provide a more efficient and user-friendly virtual assistant experience. Key features: - Custom Launchpad interface - Improved multimedia player - Speech recognition module for voice control Tech stack: Python Why it matters: This project offers a more advanced and user-friendly virtual assistant, making it a valuable tool for managing tasks and controlling devices efficiently. Repository: JoelShine/Jarvis-v2.0 GitHub stars: 65 GitHub description: A virtual assistant project that includes a custom Launchpad interface, improved multimedia player functionality, and speech recognition capabilities. It aims to enhance the efficiency and ease of use of a virtual assistant.

### Overview: 在ComfyUI中进行LoRA微调的节点，依赖于Kohya-SS/sd-scripts等训练工具。这些节点简化了LoRA微调的流程，使得开发者能够更高效地进行模型训练。 ### Key features: - 支持混元LoRA训练，脚本来自https://github.com/KohakuBlueleaf/sd-scripts/tree/HunYuanDiT 分支。 - 初始化训练文件夹，文件夹位于output目录。 - 复制图片到训练文件夹中和一些数据集配置。 - 生成相同标注文件或单一标签内容。 - 基础的训练配置和更多的训练配置。 - 训练主线程，包括加载LoRA模型和示例图片生成。 ### Tech stack: Python ### Why it matters: 简化了LoRA微调的流程，提高了开发效率，使得LoRA微调变得更加容易和快捷。

### Overview: GreatSage 是一款Windows桌面语音秘书，能够听你说话并记住跨会话的信息。它能够将会议和资料整理成带出处的纪要、待办事项和文档，并在操作前确认用户输入。此外，它还支持云和本地AI，具有可审计和可审阅的记忆功能。 ### Key Features: - **跨会话记忆**：保留原文、显式记忆和三层摘要，支持关键词检索和语义检索。 - **带出处的纪要**：导入 `.md` 资料，整理成带出处的纪要和待办事项。 - **操作先确认**：写入、打开和删除操作需要在控制台确认。 - **云或本地模型**：支持语音识别、语言模型和语音合成。 - **可审计**：每次交互都有 trace ID，记录实际使用的来源、Skill 版本和耗时。 - **桌宠**：Q版猫头鹰显示运行状态、任务进度和待确认提示。 ### Tech Stack: - **GitHub**: 用于存储代码和项目信息。 - **Python**: 主要编程语言。 - **Electron**: 开发桌面应用。 - **LangChain**: 连接云端/本地AI。 - **LangGraph**: 编排任务与审批。 - **MCP HTTP**: 外部工具显式发现和确认。 ### Why It Matters: GreatSage 提供了一种高效、可审计的桌面语音秘书解决方案，能够帮助用户整理会议

### Overview: RESTORE-DiT is a novel Diffusion-based framework designed for Satellite Image Time Series (SITS) reconstruction. It promotes sequence-level optical-SAR fusion through a diffusion framework, achieving superior reconstruction performance for highly-dynamic land surfaces under persist cloud cover. ### Key Features: - Promotes sequence-level optical-SAR fusion using a diffusion framework. - Achieves superior reconstruction performance for highly-dynamic land surfaces under persist cloud cover. - Firstly, inspired by the success of diffusion models in image and video generation, approaches the time series reconstruction problem from the perspective of conditional generation. - Conditioned on SAR image time series and date information, RESTORE-DiT achieves high reconstruction accuracy even under highly dynamic and cloud-covered conditions. ### Tech Stack: Python ### Why It Matters: RESTORE-DiT offers a novel and effective solution for time series reconstruction in satellite imagery, particularly in areas with highly dynamic land surfaces and persist cloud cover. Its performance is superior to existing methods, making it a valuable tool for remote sensing and environmental monitoring applications. ### Repository: SQD1/RESTORE-DiT - Primary language: Python - GitHub stars: 66 - GitHub description: [RSE 2025] RESTORE-DiT: Reliable satellite image time series reconstruction by multimodal sequential diffusion transformer ### Data Preparation: 1. **Dataset Download**: Download the original PASTIS-R dataset from [here](https://zenodo.org/records/5735646). 2. **Generate Cloud Masks**: Use CloudSEN12 to

### Overview: Gemini Nano Banana 2.1 Prompts: 8 Tested Recipes with Outputs ### Key Features: - **Model ID:** `gemini-nano-banana-2.1` - **Image Prompt Recipes:** 8 different prompts with example renders, aspect ratios, and the cost per image. - **Effective Price:** $0.0137 per image, estimated cost for the entire set: $0.1096. ### Tech Stack: Python ### Why It Matters: - **Single Endpoint:** Access to multiple models and APIs with a single endpoint, compatible with OpenAI. - **Dynamic Pricing:** No need for monthly subscriptions or upfront fees, pay as you go. - **Real-Time Monitoring:** Cost is reflected in the response, and tasks can be retried or batched as needed. - **Asynchronous Execution:** Tasks are executed in batches and can be monitored through `GET /v1/tasks/{id}`. ### Repository: - **Repository:** a-kiecshv/gemini-nano-banana-2-1-prompts-id - **Stars:** 49 - **Description:** Prompts for Gemini Nano Banana 2.1, including example renders, aspect ratios, and the cost per image. ### Primary Language: - **Primary Language:** Python ### GitHub Description: - **Description:** Gemini Nano Banana 2.1 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`gemini-nano-banana-2.1`) and the cost

### Overview: Seedream 5.0 Flash Prompts offers tested image prompt recipes, including examples, aspect ratios, and the model ID (`seedream-5-0-flash`) along with the cost per image. This resource is useful for developers looking to optimize their image prompts for better results and cost efficiency. ### Key Features: - **Model ID:** `seedream-5-0-flash` - **Aspect Ratios:** 1:1, 4:3, 8:5, etc. - **Cost per Image:** $0.0137 - **Prompt Examples:** Recipes for various subjects, light setups, and framing techniques. ### Tech Stack: - Python ### Why It Matters: - **Efficient Pricing:** Access to the Seedream 5.0 Flash model without the need for monthly fees. - **Versatile Light Setup:** Explicitly setting the light direction and quality to guide the model. - **Clear Framing:** Using specific ratios and lens language to ensure the output remains unchanged. - **Cost Transparency:** The cost is explicitly returned in the response, making billing straightforward. ### Repository: - **Repository:** `ac0y48cgee/seedream-5-0-flash-prompts-ru` - **GitHub Stars:** 50 - **Primary Language:** Python - **GitHub Description:** Seedream 5.0 Flash Prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedream-5-0-flash`) and the cost per image.

Overview: Cortex is an open-source AI backend control plane designed for modern AI applications, offering model routing, agent tools, OpenAI-compatible APIs, and private workspaces. Key features: - One API for many providers - Routing that can adapt - An agent harness you can own - Private workspaces for real work - Simple extension points Tech stack: JavaScript Why it matters: Cortex provides a modern AI stack without the constraints of a single provider SDK, brittle prompt chain, or proprietary agent loop, making it ideal for teams looking to build AI applications without being locked into a specific technology stack.

### Overview: This repository provides sample code for implementing a question answering application using the Retrieval Augmented Generation (RAG) technique with Amazon Bedrock. A RAG implementation consists of two parts: a data pipeline that ingests data from documents into a knowledge base, and an application that retrieves relevant information from the knowledge base and generates a response using an LLM. ### Key Features: - **Data Pipeline:** Ingests documents from S3 into an Amazon OpenSearch Service Serverless (AOSS) knowledge base. - **Application:** Receives user questions, retrieves relevant information from the knowledge base, and generates a response using an LLM. - **Managed Knowledge Bases:** Simplifies the setup by eliminating the need for a separate vector store, supports agentic retrieval, and manages embedding, storage, and retrieval automatically. ### Tech Stack: - Amazon Bedrock - Amazon OpenSearch Service Serverless (AOSS) - Amazon SageMaker Notebook ### Why It Matters: This workshop provides a practical guide for implementing a question answering application using Amazon Bedrock's RAG technique, making it easier to integrate and manage a knowledge base without the need for complex data pipelines.

### Overview: Discover free online AI certifications, courses, and digital badges from top universities and platforms. ### Key features: - **Interactive Dashboard**: Visualize the AI learning landscape with dynamic charts. - **Filterable Certifications**: Find courses by AI domain or credential type. - **Detailed Course Information**: View key topics, duration, and direct course links. - **Learning Journey Guide**: Actionable advice on strategic learning, portfolio building, and leveraging credentials. - **Responsive Design**: Optimized for desktop, tablet, and mobile devices. ### Tech stack: HTML ### Why it matters: Empower learners to enhance their AI skills and earn verifiable credentials without financial barriers.

### Overview: GarmentDreamer is a project that introduces 3DGS Guided Garment Synthesis with Diverse Geometry and Texture Details, aiming to enhance the quality of 3D garment synthesis. This innovative approach leverages diverse geometry and texture details to create more realistic and varied garments. ### Key Features: - **3DGS Guided Synthesis**: Utilizes 3D Geometry and Structure (3DGS) to guide the synthesis process, ensuring more accurate and diverse results. - **Diverse Geometry and Texture Details**: Incorporates various geometric and texture details to produce more realistic and varied garments. - **Improved Deformation**: Enhances the deformation process to achieve more significant changes in the garment. ### Tech Stack: - **Python**: The primary programming language used for implementing the project. ### Why It Matters: GarmentDreamer offers a significant improvement in the quality of 3D garment synthesis by incorporating diverse geometry and texture details, making it a valuable tool for designers and developers in the field of 3D garment design and simulation.

Overview: OpenStudio is an open-source AI image generation studio featuring a justified gallery, node-based workflows, and a variety of models including Gemini 3.1 Flash, Gemini 3 Pro, GPT Image 1, FLUX, and more. It is built with React, Tailwind CSS, and Hono, and supports SQLite as a database. Key features: - Quick Generate: Prompt bar with model, aspect ratio, resolution, and batch count selectors - Justified gallery: Generated images displayed in a pixel-perfect justified layout - Image-to-image: Drag a generated image or upload a file as source for the next generation - Node-based workflows: React Flow editor for building reusable multi-step generation pipelines - Subject lock: Save a character or product once and keep it identical across a whole batch - Style presets: Save a look once and reuse it across every workflow - Prompt variables: Type `/` in a prompt node to reference another prompt's content - Auto-naming: Prompt nodes automatically get a descriptive title via Gemini 3.1 Flash Lite on blur - Editable titles: Double-click any node title to rename it - 10+ models: Gemini 3.1 Flash, Gemini 3 Pro, GPT Image 1, FLUX, and more via OpenRouter - Live model list: Every image model OpenRouter currently offers, fetched at request time - Cost visibility: Each model is tagged with a price range, and every generation records what the provider actually billed - Generation history: All generations persisted with prompt,

### Overview: This project showcases a variety of AI technologies including Voice Recognition, Text to Speech, Language Translation, OAuth2, Image Generation, Face Detection, and a Voice Chatbot. Delphi developers can experiment with different AI models and providers, enhancing their applications with AI capabilities without vendor lock-in. ### Key features: - Voice Recognition - Text to Speech - Language Translation - OAuth2 - Image Generation - Face Detection - Voice Chatbot ### Tech stack: Delphi with Pascal programming language. ### Why it matters: This project aims to make AI accessible to Delphi developers by providing a generic interface to various AI models. It demonstrates the versatility of AI in applications like language translation, image generation, and real-time speech-to-text, making it easier for developers to integrate AI capabilities into their projects.

Overview: This Unity package allows developers to run pretrained diffusion models with Unity Sentis, providing a simplified interface for creating and applying diffusion-based effects in their projects. Key features: - Support for various diffusion models such as SD 1.5, SD 2.1, SDXL, and SD-Turbo. - Compatibility with multiple pipeline/model types including txt2img, img2img, inpaint, upscale, depth2img, and marigold. - Async pipeline methods for efficient slicing of inference. - Support for multiple images per prompt. Tech stack: C# Why it matters: This package simplifies the integration of advanced diffusion models into Unity projects, making it easier for developers to experiment with and implement cutting-edge image generation techniques in their applications.

### Overview: This repository is a comprehensive survey of Multimodal Recommender Systems (MRS), focusing on unified processes and techniques across four branches: Modality Encoder, Feature Interaction, Feature Enhancement, and Optimization. It includes resources and papers from the CSUR'24 conference, providing a valuable resource for researchers and developers in the field. ### Key Features: - **Unified Process Overview**: The survey outlines the general process of MRS, including Raw Feature Representation, Feature Interaction, and Recommend Model. - **Branch Classification**: Works are classified into four branches: Modality Encoder, Feature Interaction, Feature Enhancement, and Optimization. - **Open-Source Repositories**: Two repositories are provided for implementing MRS models: MMRec and Cornac. - **Datasets**: A detailed dataset list is included, covering various domains such as micro-video, movie, and news, with specific modality types and scales. ### Tech Stack: - **Programming Languages**: Unknown, but the primary language is not specified. ### Why it Matters: This survey provides a clear understanding of MRS, covering essential techniques and resources, making it a valuable resource for researchers and developers looking to build or improve multimodal recommender systems.

Overview: Syn.Speech is a flexible speaker-independent continuous speech recognition engine designed for Mono and .NET framework, developed by Leslie Timmy for a digital assistant interface in Linux environments. Key features: - Speaker-independent continuous speech recognition - Flexible and customizable engine - Compatible with Mono and .NET framework Tech stack: C# Why it matters: Syn.Speech is essential for creating robust and user-friendly digital assistants that can understand and respond to spoken commands in various environments, making it a valuable tool for developers working on AI and voice technology projects.

Overview: Automated Course Content Generator (ACCG) is an AI-powered Streamlit app designed to assist SMEs, Content Creators, and Educators in creating detailed and comprehensive course content effortlessly. Users input course details to generate a dynamically generated course outline, full course content, and quizzes. Key features: - Interactive Course Customization 🎨 - Detailed Content Generation 📝 - Quiz Generation 🧩 - PDF Export 📄 - PPT Export (New!) 📄 - LATEX Tech stack: Python, Streamlit, OpenAI's GPT-3.5 Turbo/4 Turbo, Meta’s LLama-3-70B-Instruct, Prompt Engineering, GitHub, VS-Code. Why it matters: Streamlit simplifies the process of creating educational content, making it more efficient and user-friendly. The AI models used ensure that the generated course content is tailored and engaging, enhancing the learning experience for users.

Overview: HAI Code Generator is a VS Code extension designed to streamline development workflows by automating task execution, file management, and code generation. Leveraging AI-powered capabilities, it integrates with various LLMs to enhance productivity and code quality. Key features: - AI-assisted code generation - Context-aware code generation with built-in and custom domain experts - Inline editing for quick code modifications - Intelligent file identification and management - Configurable settings for LLMs and models - Correlation Matrix Integration for tracking code patterns and origin analysis Tech stack: TypeScript Why it matters: HAI Code Generator simplifies complex development tasks by automating repetitive and time-consuming processes, enhancing code quality, and streamlining workflows for both developers and teams aiming to boost productivity and precision.