# VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

**Authors:** Yu-Cheng Chou, Shihao Wang, Jin Huang, Chi-Pin Huang, Chiao-An Yang, Xianzhe Fan, Jing Wang, Zhiqi Li, Yuanguo Kuang, Ruijie Zheng, Johan Bjorck, Sifei Liu, Hang Su, Abhishek Badki, Lei Zhang, Alan Yuille, Linxi "Jim'' Fan, Jan Kautz, Zhiding Yu  
**Lab:** NVIDIA · **Published:** 5 Nov 2026 · **Upvotes:** 100 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2611.08449) · [PDF](https://arxiv.org/pdf/2611.08449) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.08449.html)

## Curator's notes

**TL;DR:** VFlash jointly decodes visual primitives for fast, accurate planning, achieving ~25× speedup and higher task success.

- **What's new:** It replaces token‑by‑token autoregressive decoding with a single‑pass drafter that predicts whole primitives jointly.
- **Key result:** A world action model using VFlash plans reaches 70.2% average success across four MolmoSpaces tasks, versus 62.6% baseline.
- **Why it matters:** Builders get much quicker and more precise visual goal predictions, enabling responsive robotics and autonomous systems.

## Key numbers

- **25×**: Preliminary single-rollout measurements indicate approximately 25x faster generation for 3D point motion forecasting
- **70.2%**: …world action model conditioned on VFlash plans achieves 70.2% average success across four MolmoSpaces task suites
- **62.6%**: compared with 62.6% for the baseline

## Abstract

Models for physical AI tasks, such as robotics and autonomous driving, need not reason solely in language. They can use visual plans--points, bounding boxes, and trajectories--to represent spatial goals and anticipate motion. However, autoregressive vision-language models generate these plans one coordinate token at a time, making richer predictions increasingly costly. We introduce VFlash, a method for fast, accurate visual planning through Flash Primitive Decoding. VFlash exploits the known structure of visual primitives to predict their coordinates jointly, using a lightweight drafter that generates an entire primitive in one forward pass. The layout adapts to the primitive's type, dimensionality, and length, supporting compact boxes and long, variable-length trajectories. Crucially, spatial accuracy need not require reproducing an autoregressive model's exact token sequence. VFlash therefore trains on annotated primitives and directly accepts structurally valid predictions without target-token agreement. Across the evaluated 2D and 3D robot trajectory tasks, VFlash consistently reduces prediction errors relative to autoregressive decoding. Preliminary single-rollout measurements indicate approximately 25x faster generation for 3D point motion forecasting. A world action model conditioned on VFlash plans achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline.

## What you could build

1. Generate real‑time 3D trajectory forecasts for drone navigation
2. Create an interactive robot arm planner that outputs bounding‑box goals instantly
3. Build a visual‑language assistant that suggests motion plans for self‑driving cars

## Quiz yourself

How does VFlash generate visual primitives compared to autoregressive models?

**Answer:** VFlash uses a lightweight drafter to predict all coordinates of a primitive jointly in one forward pass, instead of generating tokens sequentially.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.