The Research & Innovation Magazine
SIGGRAPH, CHI, Ars Electronica, and MIT Media Lab — academic research translated for a creative audience.
DELUGE — Decomposed Entropy-coded Live Unstructured Geometry Exchange for Real-time Particle Streaming — is accepted to IEEE ISMAR 2026 and TVCG. It targets the transmission of live, unstructured particle and point-cloud geometry for immersive visualisation, which is the bottleneck for remote volumetric capture and telepresence work.
PrismML's Ternary-Bonsai-2-27B ships GGUF builds at roughly 1.72 bits per weight, putting a 27B model in 5.95GB against a ~54GB FP16 baseline — a 9x reduction while retaining a claimed 98.2% of FP16 benchmark performance. It entered our tracker at trending rank #4 with 405,609 downloads, and requires PrismML's llama.cpp fork.
Ampixa Labs' sanoTTS runs a full phonemes-to-audio neural TTS pipeline on a ~$3 ESP32-S3 microcontroller, generating 4.54 seconds of speech in 1.02 seconds — 0.22× real time, without a neural accelerator. Ten new languages landed on September 8, and the models are GPLv3 with an arXiv paper behind them.
A new paper introduces a 3D-printed, conformal electronic skin for humanoid robots that maps touch and pressure across a whole body using electrical impedance tomography and just 16 electrodes, rather than a dense embedded sensor grid — a geometry-scalable approach to giving robots a real sense of touch.
Sakana AI, NYU, and MIT rebuilt Picbreeder — the classic site where humans evolved abstract images through selection — with vision-language model agents doing the choosing. The result won Best Paper at GECCO 2026, and it's a genuinely honest measurement of where AI creativity still falls short.
Korea's answer to the AI compute gap isn't a national cloud — it's a joint lab with NVIDIA, embedded directly inside one of the country's top research universities.
Merging 20-30 community LoRAs by hand means dragging that many sliders and hoping. GimmBO turns it into a conversation: show it what you like, it narrows the search.
Instead of treating every video frame as equally important, Mage-VL borrows the logic of video compression itself — predict what's redundant, spend compute only on what's new.
Gemini 3.5 Flash-Lite wasn't pitched as an image tool. A week later it's already pulling nearly 12% share of OpenRouter's image-generation usage, mostly at the expense of Google's own older Flash-Lite.
Twelve-tone equal temperament is a default, not a law. A research project from the American University of Beirut has made the alternative both readable and playable — an open corpus of Arabic maqām theory plus a plugin that retunes MIDI while you play.
NVIDIA used its SIGGRAPH keynote to show where AI-driven graphics is heading: DLSS 5 that swaps rendering models per object, and a world model small enough to run on-device. Both matter well beyond gaming.
Photorealistic path tracing looks incredible and runs slowly; real-time relighting is fast but approximate. Disney Research's new 'neural render proxy' tries to get both — interactive relighting that still looks path-traced.
No screen, no headset — just a speck of matter held aloft by sound, whipped through the air fast enough to draw a glowing 3D shape you can walk around. Acoustic levitation displays are inching from lab marvel toward real medium.
Generative video gives you a clip to watch. Genie 3 gives you a place to walk around in — a world that keeps generating, coherently, as you move through it. For artists, that's not a longer video; it's a new kind of space.
Smell is the sense most directly wired to memory. This project takes that fact seriously as design material, rather than treating AI-and-photos as a purely visual problem.
Converting a 3D scene into accurate 2D line art has been unreliable and slow for years. This year's award-winning approach fixes it by changing which math problem it's solving.
No tool reviews in Research & Innovation yet.