NVIDIA Introduces SANA-WM: A 2.6B-Parameter Open-Supply World Mannequin That Generates Minute-Scale 720p Video on a Single GPU

World fashions (techniques that synthesize reasonable video sequences from an preliminary picture and a set of actions) have gotten central to embodied AI, simulation, and robotics analysis. The core problem is scaling these techniques to generate minute-long, high-resolution video with out requiring prohibitively massive clusters for each coaching and inference. Best open-source baselines both require multi-GPU inference or sacrifice decision to remain inside compute budgets.

NVIDIA’s SANA-WM immediately targets these bottlenecks. Constructed on the SANA-Video codebase and accessible via the NVlabs/Sana GitHub repository, it’s a 2.6B-parameter Diffusion Transformer (DiT) educated natively for one-minute era at 720p with metric-scale 6-DoF digicam management. It helps three single-GPU inference variants: a bidirectional generator for high-quality offline synthesis, a chunk-causal autoregressive generator for sequential rollout, and a few-step distilled autoregressive generator for quicker deployment. The distilled variant denoises a 60-second 720p clip in 34 seconds on a single RTX 5090 with NVFP4 quantization.

The Structure: 4 Core Design Choices

1. Hybrid Linear Consideration with Gated DeltaNet (GDN)

Customary softmax consideration has reminiscence and compute complexity that grows quadratically with sequence size — a significant issue when producing 961 latent frames for a 60-second video at 720p. SANA-Video, the predecessor, used cumulative ReLU-based linear consideration, which maintains a constant-size recurrent state. Nonetheless, this has no decay mechanism: all previous frames accumulate with equal weight, inflicting drift over minute-scale sequences.

SANA-WM replaces most consideration blocks with frame-wise Gated DeltaNet (GDN). Not like token-wise GDN utilized in language fashions, SANA-WM’s frame-wise variant processes one total latent body per recurrent step. The GDN replace rule incorporates a decay gate γ (which down-weights stale previous frames) and a delta-rule correction (which updates solely the residual between the goal worth and the present state prediction), preserving the recurrent state at a continuing D×D measurement no matter video size.

To stabilize coaching, the analysis crew introduces an algebraic key-scaling method: keys are scaled by 1/√(D·S), the place D is the pinnacle dimension and S is the variety of spatial tokens per body. This ensures the spectral norm of the transition matrix stays bounded and eliminates the NaN divergence occasions noticed with normal L2 key normalization (1/√D) or no scaling in any respect, each of which triggered NaN occasions at steps 16 and 1, respectively.

The ultimate spine interleaves 15 frame-wise GDN blocks with 5 softmax consideration blocks (at layers 3, 7, 11, 15, and 19) throughout 20 whole transformer blocks. The softmax blocks present precise long-range recall the place GDN’s recurrence alone is inadequate.

2. Twin-Department Digicam Management

Digicam-controlled world modeling requires the mannequin to faithfully comply with a steady 6-DoF trajectory, not simply align with a textual content description of movement. SANA-WM makes use of two complementary branches that function at completely different temporal charges:

• Coarse department (UCPE consideration): Operates on the latent-frame price. For every latent token, it computes a ray-local digicam foundation from the camera-to-world pose and intrinsics, then applies a Unified Digicam Positional Encoding (UCPE) to the geometric channels of every consideration head. This captures world trajectory construction throughout the complete sequence.
• Wonderful department (Plücker mixing): Addresses a compression mismatch. Every latent token summarizes eight uncooked frames, every with its personal distinct digicam pose. The high-quality department computes pixel-wise Plücker raymaps (a 6D illustration: ray course d and second o×d) from all eight uncooked frames inside one VAE temporal stride, packs them right into a 48-channel tensor, and injects this embedding after every self-attention output through a zero-initialized projection. This restores intra-stride digicam movement that the coarse department can not see at latent-frame decision.

Ablations on OmniWorld present that neither department alone matches the twin method: UCPE-only achieves a Digicam Movement Consistency (CamMC) of 0.2453, whereas UCPE + Plücker mixing reaches 0.2047.

3. Two-Stage Technology Pipeline

Stage-1 SANA-WM outputs, whereas spatiotemporally constant, can include structural artifacts over lengthy sequences. A second-stage refiner, initialized from the 17B LTX-2 mannequin with rank-384 LoRA adapters fine-tuned on paired artificial and actual video information, corrects these artifacts. It makes use of truncated-σ move matching: stage-1 latents are perturbed with a big beginning noise (σ_start = 0.9), and the refiner learns to map this noisy enter towards the high-fidelity goal. Solely three Euler denoising steps are wanted at inference. The refiner reduces long-horizon visible drift (ΔIQ) from 3.79 to 1.17 on the Easy-Trajectory break up, and from 3.09 to 0.31 on the Exhausting-Trajectory break up.

4. Sturdy Information Annotation Pipeline

Coaching camera-controlled video era requires metric-scale 6-DoF pose annotations, the data not accessible in normal video datasets. The analysis crew modified VIPE (a camera-pose annotation engine) by changing its depth backend with Pi3X (for long-sequence-consistent depth) fused with MoGe-2 (for correct per-frame metric scale). In addition they prolonged the bundle adjustment stage to deal with focal lengths and principal factors as per-frame variables fairly than shared world intrinsics, enabling extra strong annotation on web video with various focal lengths.

The ensuing pipeline processes seven coaching corpus entries drawn from a number of open-source sources: SpatialVID-HQ (actual, 10s clips), DL3DV actual clips (10s), DL3DV GS Refined artificial clips (60s, rendered through 3D Gaussian Splatting), OmniWorld (artificial, 60s), Sekai Sport (artificial, 60s), Sekai Strolling-HQ (actual, 60s), and MiraData (actual, 60s). This yields a complete of 212,975 clips with metric-scale pose annotations. The LTX2-VAE used for compression is 2.0× smaller than ST-DC-AE and eight.0× smaller than Wan2.1-VAE, which immediately improves coaching and inference effectivity.

For DL3DV, which comprises static 3D scene captures fairly than native one-minute movies, the analysis crew match one FCGS 3D Gaussian Splatting reconstruction per scene, designed various one-minute digicam paths, rendered lengthy movies with recognized intrinsics and extrinsics, after which refined the rendered outputs with DiFix3D to cut back splatting artifacts.

Coaching Technique and Infrastructure

SANA-WM’s compute includes two phases on 64 H100 GPUs. First, earlier than DiT coaching, the crew adapts the LTX2 VAE to the SANA-Video SFT coaching information in roughly 50K steps, taking roughly 3.5 days. The principle DiT coaching then follows a four-stage progressive schedule lasting roughly 15 days:

• Stage 1 (~2.75 days): Adapt the pre-trained SANA-Video mannequin to the frame-wise GDN structure on quick (5s) video clips. This replaces cumulative linear consideration with the recurrent GDN blocks on a less expensive, short-horizon coaching regime the place failure modes are simpler to diagnose.
• Stage 2 (~2 days): Introduce hybrid consideration by changing each fourth GDN block with a regular softmax consideration block on the identical short-clip setting, bettering the effectivity–high quality trade-off.
• Stage 3 (~8 days): Prolong coaching to 961-frame (60-second) sequences and incorporate Twin-Department Digicam Management. Context-Parallel (CP=2) sharding distributes the latent sequence throughout GPUs utilizing prefix-sum composition of GDN transition matrices — a mathematically precise parallelization technique requiring minimal communication overhead.
• Stage 4 (~2.5 days): Wonderful-tune a chunk-causal variant for autoregressive rollout, then apply self-forcing distillation to cut back sampling to 4 denoising steps. Consideration-sink tokens and native temporal home windows are added to the softmax consideration layers to maintain reminiscence and per-chunk latency fixed throughout lengthy rollouts.

Customized fused Triton kernels for GDN scan and gate operations contribute roughly 1.5× to 2× effectivity good points all through coaching.

Benchmark Outcomes

The analysis crew introduces a purpose-built 60-second world-model benchmark with 80 preliminary scenes generated by Nano Banana Professional throughout 4 scene classes sport, indoor, outdoor-city, and outdoor-nature (20 per class). Every paired with Easy and Exhausting digicam trajectory splits. The principle analysis makes use of every mannequin’s multi-step, undistilled autoregressive setting.

On this benchmark, SANA-WM with the second-stage refiner achieves the next throughout each splits:

• Digicam accuracy (Easy / Exhausting): Rotation error (RotErr) of 4.50° / 8.34°; Translation error (TransErr) of 1.39 / 1.39; CamMC of 1.41 / 1.44 — one of the best amongst all in contrast strategies, together with LingBot-World (14B+14B parameters, 8 GPUs) and HY-WorldPlay (8B parameters, 8 GPUs).
• Visible high quality: 80.62 / 81.89 VBench General on Easy / Exhausting splits, similar to LingBot-World (81.82 / 81.89) whereas producing 720p outputs on a single GPU per clip.
• Throughput: 22.0 movies/hour on 8 H100s with the complete pipeline (refiner included), in comparison with 0.6 movies/hour for LingBot-World — a 36× throughput benefit.
• Reminiscence: The complete pipeline suits in 74.7 GB, throughout the 80 GB H100 finances. Stage-1-only inference suits in 51.1 GB.
• Temporal stability: After refinement, ΔIQ (imaging high quality degradation from first to final 10-second window) drops to 1.17 on Easy and 0.31 on Exhausting, in comparison with 23.59 and 25.88 for HY-WorldPlay.

Marktechpost’s Visible Explainer

# Clone the repo
git clone https://github.com/NVlabs/Sana.git
cd Sana && ./environment_setup.sh sana

Key Takeaways

• NVIDIA’s SANA-WM generates 60-second, 720p, camera-controlled movies on a single GPU — educated in ~18.5 days on 64 H100s with solely 212,975 public video clips.
• A hybrid Gated DeltaNet + softmax consideration spine retains the recurrent state at a continuing D×D measurement no matter video size, fixing the reminiscence explosion that makes minute-scale era impractical with normal softmax consideration.
• Twin-branch digicam management — UCPE on the latent-frame price and Plücker mixing on the raw-frame price — brings CamMC all the way down to 0.2047, one of the best amongst all in contrast strategies together with fashions 5× bigger.
• A second-stage refiner initialized from 17B LTX-2 with rank-384 LoRA cuts long-horizon visible drift (ΔIQ) from 3.09 to 0.31 on Exhausting trajectories utilizing simply 3 Euler denoising steps.
• At 22.0 movies/hour on 8 H100s, SANA-WM + refiner delivers 36× greater throughput than LingBot-World (14B+14B, 8 GPUs) at comparable VBench visible high quality scores.

Try the Paper, GitHub Repo and Project Page. Additionally, be happy to comply with us on Twitter and don’t overlook to hitch our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Have to associate with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and many others.? Connect with us