World models for video have been one of those capabilities that everyone agrees matters — autonomous vehicles, robotics simulation, game engine replacement — but where the compute and memory requirements have kept serious open-source work firmly in the “nice demo, impractical to run” tier. NVIDIA’s SANA-WM, released this week, is a notable step out of that tier.
The model is a 2.6B-parameter Diffusion Transformer that takes a single image and a 6-DoF camera trajectory as input and produces a 720p, 60-second video that follows the trajectory faithfully. What’s notable isn’t just the resolution or the duration (prior open-source world models maxed out well below both) but the inference cost: a distilled variant generates a full 60-second clip in 34 seconds on a single RTX 5090 with NVFP4 quantization. Training ran on 64 H100s for 15 days using around 213,000 public video clips — meaningful but not hyperscale. The model is Apache 2.0 licensed with weights and code available.
The throughput number — 36× higher than prior open-source baselines — is striking, and it comes primarily from the architecture choice. SANA-WM uses a hybrid linear attention design it calls Gated DeltaNet (GDN): within each frame, it applies the efficient delta-rule linear attention from the DeltaNet line of work; across frames it layers in standard softmax attention to handle long-range temporal dependencies without blowing up memory. The frame-wise linear attention means that generating longer sequences doesn’t hit the quadratic wall at the intra-frame level, and the two-stage pipeline — a coarse generation pass followed by a refinement stage — improves temporal consistency without requiring a separately trained upsampler.
Camera control is handled by a dual-branch design: a control branch that processes the camera trajectory signal runs in parallel with the main generation branch, then injects the camera conditioning through cross-attention at each transformer block. The paper reports strong 6-DoF trajectory following on standard driving benchmarks. The annotation pipeline that extracts camera poses from public videos — allowing training without proprietary labeled data — is probably the part that gets underappreciated in coverage of this kind of work. Having a reproducible recipe for building your own training set is genuinely useful.
What this practically enables: researchers building simulation environments for robotics or autonomous systems no longer need a closed, proprietary video foundation model to get minute-scale, controlled video generation. The single-GPU inference cost puts it within reach of anyone with a recent high-end GPU.
The remaining limitations are real. SANA-WM is designed for forward-facing camera trajectories in the driving/outdoor domain; the training data distribution reflects that. It doesn’t handle arbitrary scene composition or indoor environments with the same fidelity. And 34 seconds on an RTX 5090 is still 34 seconds — real-time generation of 720p minute-scale video is not on the table here. But as a research artifact that’s actually runnable on one GPU, ships under a permissive license, and closes most of the quality gap with closed-source alternatives, it earns the attention it’s getting.