NVIDIA proposes VFlash visual planning decoding framework
Understand this story
NVIDIA, together with Johns Hopkins University and other institutions, proposed VFlash, a general decoding framework for Physical AI. Its approach is to use a lightweight drafter to generate complete visual primitives in parallel at once, including boxes, points, depth, and 2D/3D trajectories, while keeping the target VLM frozen. According to the research, VFlash adapts speculative decoding for VLMs; the core idea is to use the known structure of visual primitives to achieve spatial precision without having to reproduce the exact tokens of the autoregressive target. Currently only the method itself has been made public; actual speedup and accuracy performance have not yet been provided.
AI overview based on source reports · Updated 30 min ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AIVFlash: Enabling System-1 VLMs for Physical AI
VFlash is a generalist decoding framework for Physical AI that adapts VLM speculative decoding by exploiting the known structure of visual primitives, using a lightweight drafter trained on annotated geometry to generate complete primitives in one parallel pass while the target VLM stays frozen. It reduces mean 3D displacement error by 42.6% relative to the autoregressive target across three robot datasets and three target variants, and reaches 70.2% average success on MolmoSpaces versus 62.6% for the baseline. On Jetson Thor, DGX Spark, and H100, VFlash cuts decoding latency by 12–42×, yielding 5× faster robot inference and over 3× faster NAVSIM driving inference on Jetson Thor and DGX Spark.
Activity for this story
Not enough continuous observations to show a trend yet.