📦 doby-xu/witheveryone

WithEveryone: Group Photos That Actually Know Who Is Who

WithEveryone can generate a single, coherent photo of five to ten people while making sure every specific face stays exactly who it is. Most group generators today blur identities together, creating a jumbled mess where you cannot tell who is who, but this new research project solves that by assigning a dedicated token to each person and planning their layout before drawing a single pixel. It actively supervises the right face in the right spot, so the result looks like a real group shot rather than a distorted collage. An open source version is currently in training, so the full tools are on the way.

🆔 @hackernewsgithubprojects

📦 awslabs/keys_values

How Keys Values Beats Memory Limits

Keys Values lets large AI models process massive amounts of text without running out of memory. This AWS open-source tool acts like a smart filter for the model's working memory, keeping only the most important details while discarding the rest. Instead of forcing your computer to hold every single word it has ever read, it dynamically decides which pieces of information matter for the current task. This means you can fine-tune powerful models on long documents using standard hardware that would otherwise crash under the weight of data. It turns a massive memory bottleneck into a manageable calculation, making advanced AI training accessible to more researchers and developers.

🆔 @hackernewsgithubprojects

📦 solerezhang/gc-opd

GC-OPD: Making AI Reasoning Sharper on Long Tasks

GC-OPD makes small AI models reason better on long documents by fixing how they learn from mistakes. Instead of just copying a teacher’s answers, it uses a clever scoring system that weighs the teacher’s advice against a strict checker. This helps the model understand exactly which words in a long answer were wrong and why. By balancing these two signals, the model learns to be more accurate without needing massive amounts of compute. It’s a smart, efficient way to boost performance for complex tasks like answering questions from huge documents. If you’re working with AI on long text, this method could save you time and improve your results.

🆔 @hackernewsgithubprojects

📦 bawadou/ai-data-extractor

Grab Your AI Coding Chat History with AI Data Extractor

Pull every conversation from your AI coding assistants into one clean file, right on your machine. This open-source tool grabs your chat history from tools like Claude Code, Cursor, and Aider, handling all the messy local storage differences for you. It saves you from losing years of code reviews and debug sessions when an app clears its database or updates. You get normalized, easy-to-read data perfect for personal backups or fine-tuning models, without ever sending your private code to a cloud server. If you live in your IDE, this is your safety net. Grab it before it's gone.

🆔 @hackernewsgithubprojects

📦 donkeycut/donkey

Edit Videos With Chat? Donkey Cut Is Real

Donkey Cut is a free, open source video editor that runs entirely in your browser, but it lets you generate video, images, voiceovers, and music just by typing a description in a chat. It’s like having a creative director in the sidebar who can shoot the scenes you can’t actually film. You can create projects that stay local on your Mac or save them to the cloud, and the interface works the same either way. If you have a Mac companion app, it even uses your local hardware for faster transcription and screen recording.

📰 https://news.ycombinator.com/item?id=49384105

🆔 @hackernewsgithubprojects

📦 warpfront/hipfire

Run Big AI Models on Your Gaming GPU

Run large language models directly on your AMD Radeon graphics card without needing a dedicated AI chip. Hipfire is a native inference engine written in Rust that squeezes massive AI capabilities out of the GPU hardware you already own. It bypasses the usual middleman software layers, talking straight to the card to process text and images with speed that surprised even the developers. This is huge for anyone who wants powerful, private, local AI without buying expensive, power-hungry enterprise equipment. If you have a modern AMD graphics card, this project proves it is more than just a gaming tool, turning your desktop into a serious AI workstation today.

🆔 @hackernewsgithubprojects

📦 addyosmani/factory

Factory: The AI Software Factory Setup

Factory turns your GitHub issues into a self-driving code assembly line using stock Claude Code and Codex. Instead of constantly steering an agent, you install a version-controlled set of rules that lets the system triage new tickets, pick safe tasks, write the code, and open a draft pull request on its own. The real surprise is that the machine never merges anything; it stops right before the finish line to wait for your human approval. This setup uses your existing test scripts and a separate verifier to prove the code actually works before it even asks for your attention.

🆔 @hackernewsgithubprojects

📦 spade-rl/spade

SPADE: The AI That Writes Its Own Homework

SPADE is a framework where one language model does two jobs at once: it designs its own practice problems and then solves them. Usually, AI training gets stuck because we hand it a fixed list of puzzles that become too easy to figure out. This project flips that script. The model generates executable coding environments in real time, adjusting the difficulty to match its current skill level. It’s like a student writing their own test questions and then grading their own work. The result is a self-updating curriculum that keeps the learner challenged instead of letting them plateau.

🆔 @hackernewsgithubprojects

📦 milkdargon/omnialign

OmniAlign: One Model for All Language Pairs

OmniAlign does something rare: it figures out exactly which word or sentence in one language matches which in another, even when the grammar is totally different. This open source tool uses a single model to handle both word and sentence alignment across dozens of languages like English, Chinese, and German. It solves a major translation problem by finding the precise links between texts without needing a separate tool for every pair. That means you can trace exactly where a meaning shifted or how a phrase was translated, which is huge for checking quality or studying language patterns. It is built on a strong foundation and works on long documents.

🆔 @hackernewsgithubprojects

📦 udpsendtofailed/ninfer-4090

Run Qwen on a 4090

A single consumer graphics card can now run a massive twenty-seven billion parameter AI model. The project called NInfer ports the Qwen 3.8 architecture onto an RTX 4090, solving a critical problem: usually, models this large need expensive data center hardware to run smoothly. By optimizing the code specifically for this specific GPU, NInfer makes it possible to handle complex text generation and vision tasks right on your desktop without needing a cluster of servers. This is huge for developers because it puts serious, high-performance AI capabilities within reach of anyone who owns a modern graphics card.

🆔 @hackernewsgithubprojects

📦 battlewen/jailbreakskill

JailbreakSkill: The Self-Evolving AI Red Team

JailbreakSkill is the automated red-teaming framework that turns failed attacks into permanent, reusable tools. It works by first trying a standard library of prompt tweaks to stress-test an AI model. If those don't work, the system doesn't just give up. Instead, it analyzes exactly why the attack failed and automatically writes new, specialized techniques to fix it. This means the system gets smarter with every mistake it makes, creating a growing library of evolved skills. It is a brilliant way to see how AI safety gaps can be found and patched without a human manually guessing every single variation.

🆔 @hackernewsgithubprojects

📦 joeahkim/instaaddict

InstaAddict: The Self-Healing Instagram Bot

InstaAddict keeps your Instagram automation alive by actively patching the tool whenever Instagram updates its app interface. Most bots fail silently when the app changes, but this project tracks those shifts and fixes the code to match. It works by driving the real app on an Android device, mimicking human taps instead of using risky APIs. The key benefit is reliability; you get a free, open-source tool that doesn't break every week. If you need consistent engagement, this is the maintained version to use.

🆔 @hackernewsgithubprojects

📦 zeyu-zheng/far

FAR Automates Mathematical Conjecture Review

Save hours of tedious literature digging with FAR, a new tool from CMU and Anysphere. It acts as an automated research assistant that scans massive paper archives to hunt down unsolved math problems matching your specific interests. Once it finds those open questions, it uses advanced AI agents to attempt proofs and strictly grades their validity. The system filters out the noise and only hands you the most promising, verified results for your expert eye. It turns a chaotic search process into a clean, reliable pipeline that highlights high-impact mathematical discoveries ready for human review.

🆔 @hackernewsgithubprojects

📦 francois141/upal

UPAL Finds Points and Lines at Once

UPAL is the computer vision tool that finally detects points and lines in a single pass. Most image matching systems handle corners or straight edges as separate problems, but this new research from ECCV 2026 combines them into one unified network. It identifies precise sub-pixel keypoints and their descriptors while simultaneously predicting a distance field that highlights straight lines. This allows the system to match features across different photos of the same scene, linking both individual points and long segments with high accuracy. By treating the whole scene as a connected map rather than isolated parts, it helps computers understand spatial relationships much better.

🆔 @hackernewsgithubprojects

📦 facebookresearch/moe_vie

MoE-ViE: The Efficient Vision Encoder

MoE-ViE achieves high accuracy on complex video and image tasks while using a specialized mixture of experts architecture to keep computations efficient. This new vision encoder from Meta Research handles understanding both photos and video clips by routing different parts of the visual data to specialized experts within a single transformer. It solves the problem of high computational cost in standard vision models, making it possible to process large-scale visual data without the heavy energy or hardware requirements usually needed. The repository provides the full model definition and a reproducible evaluation suite, so you can test its performance on standard benchmarks yourself.

🆔 @hackernewsgithubprojects

📦 wz-cs/vggt-align

VGGT-Align Fixes the Scale Drift Problem in 3D Maps

VGGT-Align is the 3D reconstruction framework that finally stops long videos from stretching and warping as they are mapped. When you try to build a 3D model from a long sequence of images, standard tools often suffer from scale drift. This means the map slowly gets bigger or smaller over time, causing the final 3D scene to look distorted and misaligned. The key insight here is that in places like driving scenes, certain geometric features stay the same size no matter which part of the video you look at. VGGT-Align uses these stable features to lock the scale in place.

🆔 @hackernewsgithubprojects

📦 cnulab/logico

LogiCo Finds Hidden Factory Defects

LogiCo is the anomaly detection framework that finally spots the weird stuff in factory photos. Most quality checks look for specific, known errors, like a scratch on a bottle. But LogiCo takes a different approach. It learns what a perfect object looks like, then flags anything that doesn't quite fit the pattern, whether it is a logical mismatch or a structural flaw. It uses a powerful vision model to understand complex images deeply, catching subtle defects that simpler tools miss. This is huge for manufacturing because it means you can detect problems you didn't even know were possible.

🆔 @hackernewsgithubprojects

📦 plaguekind/comfyui-plaguekind-nodes

PlagueKind Nodes: Fix Your Image Masks in ComfyUI

Stop fighting with misaligned masks in ComfyUI. PlagueKind Nodes solves the headache of resizing images and their masks separately, which usually leads to weird stretching or off-center crops. This single node ensures both the picture and the mask follow the exact same geometric rules, keeping your inpainting perfectly aligned. It handles tricky aspects like maintaining aspect ratios and ensuring dimensions fit specific model requirements without breaking your workflow. Plus, it lets you drag a crop box directly on the screen to visually select exactly where you want to cut, saving you from doing any manual pixel math. If your edits keep drifting out of sync, this is the fix.

🆔 @hackernewsgithubprojects

📦 dream-num/dsh-univer-office

Turn Natural Language Into Office Files With DSH Univer Office

Imagine asking an AI to build a complex spreadsheet or presentation deck and having it actually do the work, right in front of you. DSH Univer Office connects that power to DeepSeek Harness, letting you create, edit, and inspect spreadsheets, documents, and slides just by describing what you need in plain English. The coolest part is the live preview: as the AI builds your file, you can watch the changes happen in a window right next to your chat. It handles everything from calculating formulas to laying out slide pages, and even checks for messy text overflow so your work looks clean.

🆔 @hackernewsgithubprojects

📦 azrtydxb/procoder

Procoder: The Senior Developer Guardrail for AI

Procoder acts as a strict senior developer for your AI coding agents, refusing to let them commit code that is actually broken. It works by installing a hard commit gate that treats any unchecked tool or failed lint as a failure, ensuring no unfinished work slips through. The binary itself never touches your files, but it blocks the commit until the agent fixes every issue, from merge conflict markers to junk files. It even remembers past mistakes to prevent the same bugs from recurring, forcing the agent to act like a professional engineer instead of a guessing game.

🆔 @hackernewsgithubprojects

📦 andysalerno/r9700-serving

Run AI Locally on R9700s

Serve a large language model locally using two AMD R9700 graphics cards. This project provides the exact Docker files and configuration needed to run VLLM on this specific hardware, bypassing the usual setup headaches. It includes benchmarks proving the system can handle a twenty-seven billion parameter model with impressive speed. The repository tracks different software versions to show what actually works on your machine. If you have R9700s and want to run AI privately, this is the blueprint you need to get started.

🆔 @hackernewsgithubprojects