NVIDIA Research presents VFlash for parallel visual decoding

VFlash is described on the NVIDIA Research site as a generalist decoding framework for Physical AI.[2] Its supported visual primitives include boxes, points, depth, and 2D and 3D trajectories.[2] The central change is that VFlash predicts complete visual primitives in parallel without requiring token-level agreement.[2] For variable-length predictions, a tool call declares the output length, while the primitive type determines the coordinate layout.[2] That distinction may matter when teams assess whether a model’s output format fits their perception or planning pipeline.

Fan said VFlash enables up to 12 times faster action decoding and a 5.1-times overall robot-inference speedup on edge GPUs.[1] NVIDIA Research gives two quoted latency comparisons: 703.3 milliseconds versus 3,552.7 milliseconds for autoregressive decoding, and 649.9 milliseconds versus 3,356.5 milliseconds.[2][3] The research also reports up to 61% lower error and a 42.6% decrease in mean 3D displacement error relative to the autoregressive target.[2] These are reported results; the evidence does not establish that the same gains apply to every task or deployment.

In one reported comparison, VFlash reached 70.2% average success, compared with 62.6% for the baseline.[2] In the MolmoSpaces comparison, autoregressive 2D-plus-3D plans achieved 61.0% success, while VFlash achieved 70.2%.[2] For dense 3D motion forecasting, VFlash reduced DAVIS average displacement error by 27% and generated a single rollout approximately 25 times faster on an H100 at batch size one.[2] On COCO and Dense200, it produced 2.6 to 2.8 times more bounding boxes per second, with mean F1 close to the autoregressive target.[2] The figures describe different evaluations, so teams should compare results relevant to their own tasks.

NVIDIA Research measured VFlash on customized TensorRT-Edge-LLM runtimes and TensorRT engines on Jetson Thor, DGX Spark, and H100.[2] The reported latency is model-pipeline latency, not full sensor-to-actuator time.[2] That caveat limits what the latency figures alone can say about complete robot response time. Fan described the release as open paper and open source, and the VFlash project page includes a technical report.[1][2] The available evidence does not specify licensing terms or establish how the reported measurements compare with a particular production system.