Meta AI's S-Space extracts hidden 3D coordinates from GPT-4V-style models

Meta AI released S-Space, a method that reads 3D coordinates (x, y, depth) from activations of frozen multimodal models without spatial training. A 0.3% parameter adapter yields +12% on SpaCE-10 and cuts distance estimation error by 18%.

• S-Space extracts 3D coordinates from GPT-4V-style model activations without spatial training
• The adapter adds 0.3% parameters and 2 ms latency per image
• SpaCE-10 spatial inference accuracy rose from 71.4% to 83.6%
• Distance estimation error dropped from 9.8 to 8.0 cm