SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer

TL;DR

SANA-WM is a 2.6B-parameter open-source NVLabs world model natively trained for one-minute, 720p, camera-controllable video generation. It uses Hybrid Linear Attention — frame-wise Gated DeltaNet interleaved with softmax attention — to keep long-context memory cheap, a Dual-Branch Camera Control module for precise 6-DoF trajectory adherence, and a two-stage generation pipeline that refines stage-1 outputs with a long-video refiner. The headline operating points are aggressive: trained in 15 days on 64 H100s with only ~213K public clips, 60s clip generation on a single GPU, and 34s NVFP4 denoising of a 60s 720p clip on a single RTX 5090; it claims 36× higher throughput than open baselines (LingBot-World, HY-WorldPlay) at comparable visual quality with stronger action-following accuracy.

Key claims

A 2.6B-parameter open-source WFM natively trained for one-minute generation achieves visual quality comparable to large-scale industrial baselines (LingBot-World, HY-WorldPlay) at much lower training and inference cost [§Abstract].
Hybrid Linear Attention — frame-wise Gated DeltaNet interleaved with softmax attention — provides memory-efficient long-context modeling at minute-scale [§Abstract, Design 1].
Dual-Branch Camera Control delivers precise 6-DoF trajectory adherence on top of the hybrid-attention backbone [§Abstract, Design 2].
A two-stage generation pipeline applies a dedicated long-video refiner to stage-1 outputs, improving quality and consistency across the full 60s sequence [§Abstract, Design 3].
A robust annotation pipeline extracts accurate metric-scale 6-DoF camera poses from public videos, yielding high-quality spatiotemporally consistent action labels without paired-action capture [§Abstract, Design 4].
Training uses only ~213K public video clips with metric-scale pose supervision and completes in 15 days on 64 H100s [§Abstract].
Each 60s clip is generated on a single GPU at inference [§Abstract].
The distilled variant denoises a 60s 720p clip in 34s on a single RTX 5090 with NVFP4 quantization [§Abstract].
On the authors’ one-minute world-model benchmark, SANA-WM reports stronger action-following accuracy than prior open-source baselines at comparable visual quality and 36× higher throughput [§Abstract].

Method

SANA-WM is a 2.6B diffusion transformer for action-conditioned video generation, with four load-bearing design choices. (1) The attention backbone is hybrid linear: frame-wise Gated DeltaNet handles temporal aggregation cheaply across the full 60s window, while softmax attention is interleaved to preserve the per-frame fidelity that pure linear attention is known to lose. (2) Camera control is realized as a dual-branch module that injects 6-DoF trajectory information separately from the main generation pathway, enabling precise viewpoint adherence without retraining the backbone. (3) Generation is split into two stages — a base stage produces a coarse minute-long clip, and a dedicated long-video refiner improves cross-frame consistency and visual quality in a second pass. (4) The training data is built from ~213K public clips with a robust annotation pipeline that recovers metric-scale 6-DoF camera poses (not just relative ones), so the model learns trajectory adherence in physically meaningful units. The base model and a distilled, NVFP4-quantized inference variant are released.

Results

60s 720p video on a single GPU, 15 days of training on 64 H100s, ~213K video clips [§Abstract].
Distilled + NVFP4 variant: 34s denoise for a 60s 720p clip on a single RTX 5090 [§Abstract].
36× higher throughput vs prior open-source baselines (LingBot-World, HY-WorldPlay) at comparable visual quality on the authors’ one-minute world-model benchmark [§Abstract].
Stronger action-following accuracy than prior open-source baselines on the same benchmark [§Abstract].

Why it’s interesting

SANA-WM is the cleanest counterpoint on file to Advancing Open-source World Models (LingBot-World) (LingBot-World): both are open-source minute-scale WFMs built around camera/action control, but LingBot-World is a 28B-total/14B-active MoE built by adapter-tuning a frozen Wan2.2, while SANA-WM is a 2.6B from-scratch hybrid-linear DiT — and claims 36× the throughput at parity. It is the first filed datapoint in the World Foundation Models cluster where the attention substrate itself (Gated DeltaNet + softmax) is the load-bearing efficiency story, complementing the cache-partitioning recipes filed under Context Length / Quality Trade-off in Video Generation (Context Forcing: Consistent Autoregressive Video Generation with Long Context, PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference) and the recurrent-memory-vector recipe in MALT Diffusion: Memory-Augmented Latent Transformers for Any-Length Video Generation — those keep softmax attention and engineer the cache; SANA-WM changes the attention. The Dual-Branch Camera Control design also slots into the Camera-Controlled Video Diffusion taxonomy as a new point: not a Wan-adapter, not a 4DGS reconstructor, but a from-scratch dual-branch DiT trained natively with metric-scale 6-DoF pose supervision.