This Week’s Top 5 Github Repositories

I fed a few outfits to the AI to see how well it could automatically detect clothes, and it worked pretty well! It could tell the difference between jackets, shirts, and pants even when there were multiple objects in frame. It leverages gpt-4.5-mini for vision and gpt-image-2 to make clean cutouts, and stores everything locally on your device. What really blew me away was the ability to model an editorial layout, where detected objects are laid out on top of a reference image. The drag-and-drop UI is powered by Node 22+ + Vite, so it only took me ~3 min to get up and running.

At first I was like “lol ok” but after reading the README it looks like this is a tool for encoding a message using AES-SIV, then using a local LLM (Llama 3.2 or GPT-2) to map the encrypted bytes to tokens that make up a natural language conversation, thus hiding your message within a benign chat. This happens entirely client side, with key derivation using PBKDF2-HMAC-SHA-256 with 600,000 rounds. The wizard will automatically download the model (~500 MB to ~5 GB).

I typed npx yoinks and 30 seconds later I was downloading a YouTube video. The TUI displays all available resolutions alongside their approximate file size so you can pick your poison before downloading. It’s backed by yt-dlp downloaded as a standalone binary, so no Python needed. The UI is written in Ink (React for the terminal) and the MP3 only mode with bundled ffmpeg-static is what makes this different than just calling yt-dlp yourself.

I checked out Aether-GUI and sure enough the one click connect button does exactly what it says. It fires off the Aether binary in a pseudo terminal, automatically responds to interactive prompts, and verifies the tunnel by connecting to the local SOCKS5 proxy on 127.0.0.1:1819. In the advanced tab you can select between MASQUE, WireGuard, and WARP-in-WARP tunneling protocols, along with setting the scan mode and IP version. Written using Tauri v2 + React 19 + Rust backend, much snappier than Electron apps.

I read the README and immediately got excited about the pipeline architecture. Img2Threejs is a single reference image that generates a procedural Three.js GROUP object factory with pivots, sockets, and colliders for animating the object. This is rigorous because it has eight stages of building (blockout, structural, form refinement, material, surface, lighting, interaction, optimization) where the agent has to get approved by a vision model at each stage before proceeding. No external Python deps, 12 scripts, MIT licensed.