FreeLLMVFlash introduces Flash Primitive Decoding for faster visual planning
The new VFlash method predicts visual plan coordinates jointly, cutting inference time for robotics and autonomous driving tasks.
Researchers released VFlash, a visual planning technique that replaces autoregressive token generation with joint coordinate prediction.
The method uses a lightweight drafter to output an entire visual primitive—such as a bounding box or trajectory—in a single forward pass.
VFlash adapts its layout to the primitive type, handling compact boxes and long, variable‑length trajectories without extra cost.
By training on annotated primitives, VFlash maintains spatial accuracy while avoiding the need to reproduce an autoregressive model’s token sequence.
The paper lists related works from 2026, including Jag, PDRVLA, VITAL, Forward Dynamics Models, ObjDream, DiffAdapterVLA, and Memory‑as‑Plans.
What it means for you
For AI practitioners, VFlash offers a way to speed up visual plan generation in robotics and self‑driving systems, potentially reducing compute budgets. Free‑model enthusiasts may see faster inference as a step toward more efficient local deployments, though the paper does not detail hardware requirements or open‑source releases.
Sources
- NVIDIA paper: VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding, Hugging Face Papers (official), 4 Nov 2026
Written automatically from the sources above, then checked by our code: every number and name must appear in them, and no sentence may copy a source. Spotted a mistake? Report it.
Which AI can you use for free today?
Every free AI model, ranked by blind human votes, with the memory and speed it needs on your own GPU.