Nvidia's SANA-WM turns photos into playable worlds

SANA-WM is Nvidia's new open-source, 2.6-billion-parameter interactive world model that turns a single still photo into an explorable 720p world controlled by simple keyboard arrows.

Instead of burning through multi-GPU server clusters, SANA-WM is natively designed to generate minute-scale video clips. You supply an initial picture, a text prompt, and start pressing keys, and the model synthesizes continuous footage following your movements in real time.

What it does and how it runs

The model works in both third-person and first-person perspectives. Crucially, it maintains spatial memory: if you pan the camera away and look back, the environment stays intact rather than mutating into something random.

Hardware requirements are remarkably modest. Through an NVFP4-quantized distilled variant, SANA-WM can run entirely on a single consumer-grade RTX 5090, denoising a full 60-second 720p video in just 34 seconds.

Under the hood, Nvidia combined Gated DeltaNet linear attention with standard Softmax attention to slash context memory use. They trained the entire foundation on roughly 213,000 public video clips using 64 H100 GPUs in just 15 days. On benchmarks, it delivers visual quality comparable to giant industry baselines like LingBot-World and HY-WorldPlay while boosting throughput by 36x.

The paper is currently live on arXiv, and Nvidia says open weights and code are dropping soon.

Why it matters

If you build games, interactive simulators, or visual prototypes, SANA-WM is worth bookmarking right now. Bringing real-time, minute-scale world synthesis down to a single consumer graphics card is a genuine leap in efficiency.

In our view, Nvidia's smart play here was ditching parameter bloat in favor of hybrid linear attention and distillation. They proved that responsive, coherent interactive worlds don't require an industrial server farm—just clever model engineering on accessible hardware.