Breaking01 / AI
RAWATTA processes 200 images without additional training
The method uses 8-bit precision for attention calculations and retains 16-bit precision only in layers that require it after a calibration check.

Researchers reported that RAWATTA processed 200 images without additional training in their published tests.
The researchers reported the result on a graphics processing unit (GPU) with 12 GB of memory.
The paper's submission date on arXiv is 13 Nov 2026. RAWATTA uses 8-bit precision by default for attention, the calculation that weighs relationships between inputs. It retains 16-bit precision only in layers where a calibration check requires it.
The researchers explained that RAWATTA-Fast replaces calculations in some layers with approximations. These include some layers that average information across the whole scene or use the first image as a reference.
According to the paper's abstract, VGGT reconstructs a scene in three dimensions from multiple photographs. In the researchers' published tests, its existing public code ran out of memory at 100 images.
The researchers did not disclose a specific processing speed improvement factor in the abstract.