SANA-WM: NVIDIA Packs a 'One-Minute World Model' onto a Single GPU

Forum topic · 小凯 · 2026-05-31

Summary

SANA-WM is NVIDIA's newly announced 2.6B-parameter world model that turns a single image plus a camera trajectory into 720p, 60-second explorable video. Trained on 64 H100 GPUs for roughly 15 days, it runs inference on one H100, and a distilled NVFP4-quantized variant generates a full minute in about 34 seconds on an RTX 5090. The design combines hybrid attention—15 Gated DeltaNet (GDN) blocks interleaved with 5 softmax blocks across 20 transformer layers—to avoid the quadratic memory cost of full softmax attention while preventing state drift; a dual-branch camera controller (UCPE global pose plus Plücker raymaps) for accurate 6-DoF control; and a two-stage pipeline where a 17B LTX-2 refiner with rank-384 LoRA adapters fixes long-sequence artifacts in 3 denoising steps. Compared with the 14B+14B LingBot-World, SANA-WM achieves 22.0 versus 0.6 videos/hour (about 36x throughput), better rotation error (4.50° vs 10.47° simple setting), 74.7 GB vs 454 GB memory, and comparable VBench scores. Training data comes from ~213K public video clips with metric-scale camera poses. Project page: https://nvlabs.github.io/Sana/WM/, code: https://github.com/NVlabs/Sana.

Overview

A Chinese forum post discusses SANA-WM, NVIDIA's efficient minute-scale world model. The core claim: a 2.6B-parameter model, trained on 64 H100s for ~15 days, generates 720p 60-second controllable video on a single H100 — previously requiring 14B+14B dual models on 8-GPU clusters.

Key points

Why one minute is hard: standard softmax attention scales quadratically with sequence length; 720p @ 16fps for 60s = 961 latent frames, which OOMs with full attention. Prior systems dropped to 480p and/or used 8 GPUs.
Hybrid Linear Attention: 20-block DiT with 15 frame-wise Gated DeltaNet (GDN) blocks plus 5 softmax attention blocks (layers 3, 7, 11, 15, 19). GDN adds a decay gate (γ) and delta-rule correction (β) to prevent state drift seen in pure cumulative linear attention; the recurrent state stays constant-sized (D×D).
Training stability: standard L2 key normalization caused NaNs at step 1; the team used algebraic key scaling (keys scaled by 1/√(D·S)) to bound the transition matrix spectral norm.
Dual-Branch Camera Control: a coarse branch (Unified Camera Pose Embedding at latent-frame rate) plus a fine branch (Plücker raymaps at raw-frame rate) yields rotation error of 4.50°/8.34° (Simple/Hard) and CamMC 1.41/1.44 — more accurate than 14B+14B LingBot-World.
Two-stage pipeline: the SANA-WM backbone produces the draft; a 17B LTX-2 refiner with rank-384 LoRA adapters (Q/K/V/O + FFN), fine-tuned via truncated-σ flow matching, repairs long-sequence artifacts in 3 Euler steps. Quality gap between first and last 10 seconds drops from ΔIQ 3.79/3.09 to 1.17/0.31 (HY-WorldPlay: 23.59/25.88).
Data: ~213K clips from public video with metric-scale 6-DoF camera poses — no synthetic data.

Benchmark comparison (from the post)

| Metric | SANA-WM | LingBot-World | HY-WorldPlay | Matrix-Game 3.0 | Infinite-World | |---|---|---|---|---|---| | Params | 2.6B | 14B+14B | 8B | 5B | 1.3B | | Resolution | 720p | 480p | 480p | 720p | 480p | | Inference GPUs | 1 | 8 | 8 | 8 | 1 | | Rot. err. Simple (°) | 4.50 | 10.47 | 17.89 | 12.96 | 16.55 | | Rot. err. Hard (°) | 8.34 | 18.99 | 35.46 | 18.79 | 41.31 | | VBench Simple | 80.62 | 81.82 | 68.82 | 78.53 | 79.18 | | VBench Hard | 81.89 | 81.89 | 70.46 | 78.79 | 79.51 | | Throughput (videos/hr) | 22.0 | 0.6 | 1.1 | 3.1 | 5.9 | | Memory (GB) | 74.7 | 454.1 | 215.5 | 106.2 | 53.5 |

Key takeaway: with ~9% of the parameters and 1/8 the inference GPUs, SANA-WM matches or beats camera-control accuracy and visual quality, with ~36x throughput and ~1/6 the memory.

Training pipeline (four stages, ~15 days on 64 H100s)

1. Stage 1 (~2.75d): adapt pretrained SANA-Video to frame-wise GDN on 5s clips. 2. Stage 2 (~2d): introduce hybrid attention (every 4th GDN block replaced by softmax). 3. Stage 3 (~8d): scale to 961-frame (60s) sequences with camera control; Context-Parallel (CP=2) with prefix-sum composition of GDN transition matrices. 4. Stage 4 (~2.5d): chunk-causal autoregressive variant with self-forcing distillation down to 4 denoising steps; attention-sink tokens and local temporal windows for constant memory on long rollouts.

VAE pre-adaptation adds ~3.5 days; custom fused Triton kernels give 1.5–2× throughput for GDN scan and gate ops.

Applications and details

Autonomous-driving simulation, robot training, game content generation, film previsualization.
Paper: *SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer* (arXiv 2605.15178)
Project page: https://nvlabs.github.io/Sana/WM/ — Code: https://github.com/NVlabs/Sana (SANA-WM code not yet released)
Architecture: 20-block DiT, 15 GDN + 5 softmax, d_model=2240, 20 heads; VAE: LTX2-VAE with C=128 latent channels.
Distilled + NVFP4-quantized variant runs on an RTX 5090, generating one minute in ~34 seconds.

The post's conclusion: world models don't need 28B parameters and 8-GPU clusters — a smart attention hybrid, a pragmatic two-stage pipeline, and efficient use of public video data make 2.6B sufficient.