VFlash Flash-Decodes Visual Primitives in One Pass, Hits 25x Speedup on 3D Motion Forecasting

VFlash replaces token-by-token coordinate generation in vision-language models with a 537M-parameter diffusion drafter that decodes complete visual primitives — trajectories, bounding boxes, points — in a single forward pass. On 3D robot trajectory prediction across InternA1 and AgiBot the method cuts ADE 28-61% while running ~25x faster on single-rollout dense motion. On MolmoSpaces manipulation it reaches 70.2% average success, 7.6pp over the autoregressive baseline.