📦 ga1i13o/foundyou

FoundYou: One Model, Two Superpowers

FoundYou is the unified computer vision model that lets you find and trace a single object across massive datasets. You show the system just one picture of a specific item, like a blue mug, and it instantly locates every other instance in a library of one hundred million photos. It can even highlight exactly where that object sits in new images. This tool is incredibly fast and lightweight, running on a fraction of the hardware previous systems required. It bridges the gap between spotting something and cutting it out from the background. If you are building search engines or editing photos, this is the efficient engine you have been missing.

🆔 @hackernewsgithubprojects

📦 amap-ml/dreamx-creator

DreamX Creator Makes Video Sound

DreamX Creator takes a single image and a text prompt, then generates both video and sound at the same time. It is not just adding a soundtrack to a silent clip; the model builds the two streams together so the audio actually matches the on-screen action. This creates a realistic experience where what you hear is driven directly by what you see. The project is open source, meaning you can study the method or use the code to build your own synchronized media. If you want to create video that feels alive, this is the place to start.

🆔 @hackernewsgithubprojects

📦 yifanzhang-pro/fast-weight-attention

Falcon: How AI Models Learn On The Fly

Falcon changes how neural networks handle memory by treating every new piece of information as a quick, on the spot learning opportunity. Instead of trying to remember everything in a massive list, these models compress endless streams of data into a tiny, fixed size summary that updates constantly. This is called fast weight attention, and it turns the standard attention mechanism into a kind of live learning rule. You get a system that keeps adapting to what it sees right now without needing to rewind and retrain on old data. It is a clever way to make machines that learn continuously without getting bloated or forgetting what matters.

🆔 @hackernewsgithubprojects

📦 zjunlp/autoscirub

AutoSciRub: The Grader That Fixes AI Research

AutoSciRub is the research plugin that stops guessing and starts grading. Instead of letting an A I agent wander blindly through a scientific task, it first builds a specific checklist of what a good answer actually looks like. It checks the literature and the available data to define exactly which evidence proves success. Then, it uses that list to catch every gap in the draft, guiding the agent to fix only what is missing. It is the difference between hoping for a good result and verifying one. If you are building autonomous research tools, this is the missing link.

🆔 @hackernewsgithubprojects

📦 cognicore-dev/cognicore-env

CogniCore: AI Agents That Actually Remember

CogniCore gives AI agents a memory that actually sticks, solving the frustrating problem where software forgets everything the moment it restarts. Instead of relying on expensive cloud calls for every single interaction, this open-source framework stores information locally using a method that is surprisingly efficient, requiring almost no extra data to find the right answer. It even includes a safety layer that acts like an immune system, automatically blocking malicious tricks that try to trick the AI into revealing secrets or ignoring instructions. If you are building tools that need to learn from experience, this is the one project to watch.

🆔 @hackernewsgithubprojects

📦 valeoai/vatix

How Far Can 5,500 Hours of Driving Take You?

Researchers at Valeo discovered that training video models on five thousand hours of real driving footage reveals a clear rule for how these systems improve. They built a series of video generators from small to massive scale to see how data and size affect performance. The surprising finding is that letting a model train on the same data for longer improves accuracy much faster than simply building a larger model. This matters because driving video is hard to collect and expensive to create. For anyone working with limited data, this study shows that time spent training is a far better investment than just adding more parameters.

🆔 @hackernewsgithubprojects

📦 wujh2001/atgs

ATGS: The Future of 3D Video

ATGS lets you render long, high-fidelity volumetric videos without falling apart. It uses anchored temporal Gaussian splatting, a fancy way of saying it tracks moving objects over time by anchoring them to stable points in space. This solves the messy problem of keeping 3D scenes coherent when things move or change, which is hard with standard methods. It is cool because it handles long sequences, not just short clips, using a flexible encoder that models how things change frame by frame. If you want to see how 3D video tech is evolving, this is a solid starting point.

🆔 @hackernewsgithubprojects

📦 dongping-chen/game2world

Game2World: The AI Editor That Cleans Gameplay Videos

Game2World acts as a specialized video editor that strips away the clutter from real gameplay recordings, removing health bars, maps, and menu screens while leaving the actual game action completely intact. This tool solves a major hurdle in artificial intelligence research, where computer models need clean, unobstructed footage to learn how the world moves and reacts. By providing this pristine training material, the project helps researchers teach machines to understand complex environments without getting confused by the user interface. This single capability is what makes the project so valuable to the field.

🆔 @hackernewsgithubprojects

📦 yoon307/dinode

DINOde: The Tiny Head That Segments Anything

DINOde is the vision model that segments images using only image captions and no mask labels. Most tools require thousands of labeled pixels to learn what a tree or car looks like. DINOde skips that heavy lifting entirely by freezing two massive pre-trained AI brains and adding a tiny, learnable bridge between them. It uses a mathematical flow to smoothly move text concepts into visual space, letting the computer understand complex words it has never seen before. This makes the model incredibly fast to train and easy to adapt. You get powerful, general-purpose segmentation without the data grind, proving that smart math can replace massive datasets.

🆔 @hackernewsgithubprojects

📦 dobermancore/doberman-core

Doberman: The Guard Dog for AI Coding Agents

Prevent your AI coding agent from accidentally deleting your project or leaking your private keys by intercepting every tool call before it runs. Doberman acts as a runtime guard dog, sitting between the model and your actual tools to inspect each action. It instantly blocks dangerous commands like destructive file deletions or unauthorized data transfers, while letting safe routine tasks pass through without friction. If something sensitive happens, it pauses for your explicit approval, ensuring you stay in control. This project guarantees that uncertainty results in a denial, meaning a blocked action never executes.

🆔 @hackernewsgithubprojects

📦 mertcookimg/open_duck_mini_viewer

Open Duck Mini Viewer

Control a bipedal duck robot entirely from your web browser using Open Duck Mini Viewer. This open source project simulates the hardware locally, so you can walk the duck, trigger funny motions like a headbang, and paint its body parts in real time without owning the physical device. It turns a complex robotics simulation into a playful, interactive web page that runs instantly. No heavy installation or Python scripts are needed, just a browser tab. If you enjoy tinkering with robotics or want to see how a digital twin works, this is a fun, low friction way to get your hands dirty.

🆔 @hackernewsgithubprojects

📦 freeai-org/scalpel

How Scalpel Makes AI Models Faster

A vision and language model can lose nine out of twenty eight internal layers and still keep ninety nine percent of its original accuracy. This project is called Scalpel. It treats model shrinking like careful surgery. Instead of chopping off huge chunks at once, it makes tiny, precise cuts to remove one layer at a time. After each cut, it runs a quick local repair to fix the gap and keep the model smart. This prevents the usual sharp drop in quality you see with standard pruning. By the end, the model runs forty three percent faster with barely any loss in performance.

🆔 @hackernewsgithubprojects