📦 krispinwandel/vit-up

ViT-Up: Dense Feature Upsampling for Vision Transformers

ViT-Up provides an implicit feature upsampler that predicts backbone-aligned features at any continuous image coordinate for Vision Transformers. By leveraging self-supervised feature distillation across over one million images, this project enables downstream models to work directly with dense features from DINOv3 architectures. It solves the challenge of extracting fine-grained, high-resolution information from transformers, making it useful for tasks like dense prediction and semantic correspondence. The repository includes specialized toolkits for evaluation, probing, and even experimental video encoding, allowing you to project features into visual formats easily. This framework effectively bridges the gap between transformer backbones and detailed spatial analysis.

🆔 @hackernewsgithubprojects

📦 nilskikonjizdunava/fprm

Understanding Fixed-Point Reasoners in Looped Transformers

Enhance neural network reasoning by utilizing fixed-point convergence as a dynamic halting mechanism within looped Transformer architectures. This project provides an implementation of models that address signal propagation challenges common in deep recurrent systems by incorporating pre-norm layers and specialized residual scaling. By decoupling parameter counts from effective computational depth, this approach allows models to naturally adapt their compute usage based on the specific difficulty of a task. These optimizations eliminate the need for complex, manual halting strategies, providing a more stable and efficient way for systems to solve demanding reasoning benchmarks like Sudoku and maze navigation tasks.

🆔 @hackernewsgithubprojects

📦 aigeeksgroup/motionvla

MotionVLA: Generating Realistic Humanoid Motion with Dual-Stream AI

MotionVLA is an autoregressive model that generates realistic humanoid motion from scene images and text instructions by leveraging a specialized Dual-Stream Frequency-domain Tokenizer. By splitting motion data into a Base stream for pose semantics and a Phys stream for physical dynamics, the model effectively decouples low-frequency information from high-frequency velocity components, solving the common issue where shared codebooks cause geometric drift. This approach uses a Qwen3.5 backbone to arrange these streams into a unified, phase-aware sequence, significantly improving motion consistency and diversity. It is a powerful tool for researchers and developers looking to enhance animation and robotic humanoid control.

🆔 @hackernewsgithubprojects

📦 oweixx/gast

Geometry-Aware Style Transfer in 3D Gaussian Splatting

Geometry-aware style transfer allows for the artistic transformation of 3D Gaussian Splatting scenes while preserving their underlying spatial structure. This project introduces a batch pipeline that applies reference styles to pre-trained 3D Gaussian models, enabling the creation of stylized animations. By utilizing a specific training script to process point clouds and style images, users can generate depth-aware, visually consistent results across different scenes and artistic styles. This tool is particularly useful for researchers or developers looking to integrate creative aesthetics into 3D environments without compromising the integrity of the scene geometry. It offers a structured approach for rendering stylized video outputs.

🆔 @hackernewsgithubprojects

📦 furiosa-ai/async-opd

AsyncOPD: Speeding Up Large Language Model Training

AsyncOPD is the multi-process training pipeline that finally keeps your GPUs fully utilized during complex language model distillation and reinforcement learning. By decoupling rollout generation from learner updates, it allows simultaneous processing for on-policy distillation, GRPO, and supervised fine-tuning. The framework integrates vLLM for rollout, FSDP or Megatron for training, and NCCL for efficient weight synchronization, offering configurable staleness to balance throughput against training accuracy. With curated examples for various research workloads, this project provides a robust, asynchronous architecture that lets you scale large model training jobs efficiently. Adopt AsyncOPD to keep your expensive hardware productive today.

🆔 @hackernewsgithubprojects

📦 ai-builder-club/skills

Build Compounding AI Agents with this Plugin

Automate your codebase by implementing compounding agent loops that handle work independently. This plugin marketplace provides the necessary scaffolding to transform a standard repository into an agent-ready environment. You can deploy an isolated cloud development box for each agent, generate one-command local launchers for your stack, and build robust end-to-end test gates that verify features before shipping code. By utilizing shared file-based memory and specialized plugins, your agents can investigate issues, test their own changes, and log results to create a self-improving system. Leverage these tools to design smarter workflows that save you from constant manual prompting.

🆔 @hackernewsgithubprojects

📦 lpynow/mma-82

MMA-82: Advancing Micro-Action Recognition

MMA-82 is the comprehensive benchmark that moves micro-action analysis beyond controlled labs into diverse real-world domains. This project provides a massive dataset of nearly eighty thousand annotated instances, covering eighty-two fine-grained categories like head and hand movements across psychiatric, street, and television settings. It enables researchers to tackle challenging tasks including micro-action recognition and multi-label temporal detection, supporting in-domain and cross-domain evaluation protocols. By capturing subtle body cues, this dataset reveals how actions reflect affective states, offering a powerful resource for building smarter video understanding models. Explore this benchmark to sharpen your computer vision research today.

🆔 @hackernewsgithubprojects

📦 microsoft/megadetector-overhead

Detect Wildlife from Above with Microsoft's Overhead AI

Automate wildlife monitoring in drone and aerial imagery using a specialized collection of point-based detection models. This project addresses the unique challenges of overhead perspectives, such as small, distant targets and image distortion, by providing a robust training, evaluation, and inference stack. It includes the versatile OWL model family, ranging from lightweight baseline versions to high-quality models built on DINOv3 foundation encoders. Whether you are tracking populations or surveying habitats, this repository offers the end-to-end tooling you need to process aerial data efficiently, making it an essential resource for researchers and conservationists working to protect biodiversity from above.

🆔 @hackernewsgithubprojects

📦 hammergpt/hyper-alpha-arena

Hyper Alpha Arena: The AI-Powered Crypto Trading Platform

Hyper Alpha Arena is the multi-exchange AI trading platform that lets you automate sophisticated crypto strategies using institutional-grade quantitative factors. This project enables Large Language Models like Deepseek and GPT to monitor market flow, such as order imbalances and funding rates, to trigger trades automatically on Hyperliquid and Binance Futures. Whether you use natural language for AI-driven strategies or Python for custom rule-based programs, the platform provides deep performance analytics and factor mining tools to refine your edge. Start leveraging AI agents to analyze, validate, and execute your trades efficiently today, even with no prior coding experience.

🆔 @hackernewsgithubprojects

📦 vicharak-in/shrike

Getting Started with Shrike FPGA Development Boards

A compact development board sits on a breadboard with LEDs flashing as an FPGA communicates directly with an RP2040 microcontroller. The Shrike project provides a family of low-cost, open-source FPGA development boards designed specifically for makers, students, and hobbyists. By pairing an FPGA with popular microcontrollers like the RP2040, RP2350, or ESP32-S3, these boards make it easy to explore hardware design and reconfigurable computing. With robust documentation, breadboard compatibility, and a supportive community, it acts as a perfect stepping stone for those wanting to bridge the gap between traditional software programming and complex FPGA logic. Start your hardware journey today.

🆔 @hackernewsgithubprojects

📦 archisvaze/arcdlp

Easily Download Videos From Anywhere With ArcDLP

Download videos and audio from YouTube, Instagram, and thousands of other sites using a simple, private desktop interface. This open-source tool runs locally on your machine, so you never have to worry about tracking or cloud accounts. You can fetch multiple URLs back-to-back, select your preferred quality, and manage downloads through a clean queue system. It even handles playlist batches and private content via secure, direct sign-in options. Whether you need to save clips or extract audio, this lightweight application provides a straightforward way to keep your favorite media accessible whenever you need it.

🆔 @hackernewsgithubprojects

📦 nova452/comfyui-conditioningkrea2rebalance

Unlock Krea 2 Potential with Conditioning Rebalance

This ComfyUI tool provides fine-grained control over Krea 2 model generation by allowing users to optimize conditioning through precise per layer weighting. It effectively introduces functionality similar to IP-Adapter for this model, while simultaneously bypassing the built-in quality dilution caused by the default safety filter. By working as an unfiltering solution, this implementation restores original model performance and offers significant flexibility for creators looking to refine their output beyond factory constraints. It is a straightforward way to reclaim control over your model generation process, ensuring the final results meet your exact standards without unnecessary interference from internal filters.

🆔 @hackernewsgithubprojects

📦 krea-ai/krea-2

Understanding Krea 2: A New Open Image Generation Model

Krea 2 stands out as a highly aesthetic open-source image generation model built from the ground up for creative exploration. The repository provides official inference code for two primary variants: the RAW base model, which is ideal for fine-tuning and training custom LoRAs, and the Turbo model, which is a distilled version optimized for fast, high-quality generation in just eight steps. By training on the flexible RAW model and deploying on the efficient Turbo variant, users can achieve high-resolution results tailored to their specific stylistic needs. This release provides a powerful, versatile toolkit for advancing AI-driven image generation workflows.

🆔 @hackernewsgithubprojects

📦 lei-kun/rl-100

RL-100: Real-World Reinforcement Learning for Robotic Manipulation

RL-100 provides a unified framework for training performant robotic manipulation policies through reinforcement learning. This codebase enables developers to bridge the gap between initial human teleoperation data and reliable real-world robot deployment. It supports a versatile range of policy backbones including diffusion and flow models, while allowing for both offline and online reinforcement learning training stages. By facilitating iterative data collection and policy refinement, the library helps improve the efficiency and robustness of robot control. This comprehensive toolkit offers a structured way to handle complex visual inputs and control modes, ultimately simplifying the path toward more capable autonomous robotic systems.

🆔 @hackernewsgithubprojects

📦 qiisun/meshflow

MeshFlow: AI Mesh Generation via Equivariant Flow Matching

MeshFlow provides a specialized PyTorch framework for generating 3D meshes using equivariant flow matching techniques. This repository offers a complete pipeline that enables researchers and developers to train models, perform inference, and evaluate the quality of generated 3D structures. By leveraging advanced flow matching, it addresses the complexities of unconditional mesh generation, allowing users to create detailed geometric assets across various object categories. The project includes interactive demo tools for real-time visualization and post-processing, such as hole filling and vertex merging. It serves as a sophisticated resource for those exploring generative 3D modeling and geometric deep learning.

🆔 @hackernewsgithubprojects

📦 qwenlm/qwen-agentworld

Qwen-AgentWorld: A Native World Model for AI Agents

Qwen-AgentWorld is a native language world model that accurately simulates diverse environments by reasoning through ten million real-world interaction trajectories. This project introduces a model capable of navigating seven unified domains including web, android, and terminal environments, outperforming many frontier systems in agent-based tasks. By treating environment simulation as its core training objective, the model offers superior scalability and controllable performance, even in out-of-domain scenarios. It provides developers with both powerful model weights and a comprehensive evaluation benchmark to test simulation fidelity. This innovation marks a shift toward building more reliable, agentic foundation models for complex digital interactions.

📰 https://news.ycombinator.com/item?id=48655228

🆔 @hackernewsgithubprojects

📦 signerlabs/shipswift

Build Apps Faster with AI-Native SwiftUI Components

Accelerate your development workflow by using an AI-native SwiftUI component library designed to help you build real applications instantly. This repository provides production-ready code, including complex animations like shimmer effects, data-driven charts, and complete multi-file modules for authentication, camera integration, and chat features. By connecting this library via a model context protocol server, your AI assistant gains immediate access to a searchable catalog of recipes, enabling you to generate UI components or functional flows through simple commands. Leverage these self-contained, modular SwiftUI building blocks to streamline your coding process and ship polished iOS applications with confidence.

🆔 @hackernewsgithubprojects

📦 zhangsq-nju/edgerazor

EdgeRazor: Run LLMs on Edge Devices

A small demo window displays a quantized language model running efficiently on a mobile device or local edge hardware. EdgeRazor is a lightweight framework designed to make large language models smaller and faster by integrating mixed-precision quantization-aware distillation directly into your training pipeline. It helps solve the challenge of running complex AI models on resource-constrained devices by using techniques like weight, activation, and KV cache quantization with configurable bit-widths. This allows for high-performance deployments across diverse hardware while maintaining strong task accuracy. Explore this project to optimize your own models for more efficient, high-speed edge computing performance.

🆔 @hackernewsgithubprojects

📦 sums001/windows-copilot-api

Use Microsoft Copilot as a Free API

Windows Copilot API is the clever bridge that finally lets you use your personal Microsoft Copilot account as a flexible, OpenAI-compatible service. This project reverse-engineers the consumer Copilot interface, turning it into a REST API so you can tap into its powerful chat capabilities from your own Python scripts or any OpenAI-ready application. By handling authentication and Cloudflare challenges automatically, it removes the need for expensive subscriptions or developer API keys. It is a fantastic, free way to integrate intelligent chat features into your local projects while staying within your personal account limits for daily use.

🆔 @hackernewsgithubprojects

📦 attenlabs/saa-sdk

Stop Wasting Resources on Unintended Speech with SAA

Filter out background noise and unintended audio from your voice agent using this specialized addressee layer. Instead of processing every sound in a room, it uses a classifier to determine exactly when speech is directed at your device, ensuring only relevant audio ever reaches your transcription, language model, or speech synthesis services. It works across various stacks including LiveKit and Pipecat by acting as a smart gate for your pipeline. By isolating commands from side talk, media, or ambient noise, you improve performance and lower costs. Integrate this tool into your workflow to build more focused and responsive voice applications.

📰 https://news.ycombinator.com/item?id=48649105

🆔 @hackernewsgithubprojects

📦 kauravsrestha-duryodhan/digi-netra

Understanding the Digi-Netra OSINT Toolkit

Digi-Netra is a modular Python toolkit that automates open-source intelligence gathering through a straightforward command-line interface. This project simplifies digital investigations by allowing users to trace phone numbers, usernames, email addresses, and IP addresses in one centralized environment. It performs tasks like validating phone information, checking for social media footprints, and identifying potential data breaches associated with specific emails. By consolidating these recon tools into an accessible menu system, the software provides a fast and practical way to conduct technical research. This tool essentially puts powerful investigative capabilities directly into the hands of those exploring digital footprints.

🆔 @hackernewsgithubprojects