How AromaGen works: multimodal AI decomposes any smell into 12 base-odor ratios
pliang279 · x · 2026-11-13
The AromaGen team explains how it generates smells it has never encountered: a multimodal AI model decomposes arbitrary inputs (photo or text description) into mixtures of the 12 base odors, which the wearable releases in real time. LLM internal knowledge provides generalization to new foods and beverages. The demo image shows the workflow: zero-shot generation outputs a 12-dimension ratio vector R, and users can give textual feedback for human-in-the-loop iterative refinement, with the AI reasoning adjusting per-odor release durations (e.g. raising onion from 0.10 to 0.15 while cutting strawberry from 0.30 to 0.05).
Related event: Stanford Team Unveils AromaGen: AI System for Real-Time Scent Transmission(6 posts)→
More from Multimodal
- Gemini 4 Hydrogen passes the pelican-on-a-bicycle test in leaked render — Main_Exchange3944 · 2026-11-13
- MiniMax H3 Image lands in HyperFrames Studio: 16MP images in under 10 seconds — MiniMax_AI · 2026-11-13
- Same AI character, three models: hands-on with Seedream 5.5, MiniMax H3, GPT v2.6 — miilesus · 2026-11-13
- Seedance 2.5 suspense short: 'Then who made the coffee this morning?' — LudovicCreator · 2026-11-13
- Seedream 5.5 drops: 100+ edits on the same image with far less drift — minchoi · 2026-11-13
- KeyframeLabs launches persona-2.0-live real-time avatar model with word-level expression control — Scobleizer · 2026-11-13