Crimson AI NewsA CrimsonLingua Network service
EN ع
← Back to news
Research paper Hugging Face

SANA-Video 2.0: Hybrid Linear Attention Cuts Video Generation Cost, Matches Softmax Quality

AI By Crimson AI Hugging Face Papers 24 July 2026 · 00:00 17 views
Share: X Telegram

NVIDIA and Hugging Face introduce SANA-Video 2.0, a hybrid video diffusion transformer that combines linear and softmax attention to generate high-quality 720p video on a single GPU, achieving up to 120x speedup over prior models.

SANA-Video 2.0: Hybrid Linear Attention Cuts Video Generation Cost, Matches Softmax Quality

Key points

Researchers from NVIDIA and Hugging Face have unveiled SANA-Video 2.0, a hybrid video diffusion transformer designed for efficient, high-resolution video generation. The model, available in 5B and 14B parameter scales, can produce 720p video on a single GPU while matching the quality of full-softmax video diffusion transformers (DiTs) at a fraction of the computational cost.

The key innovation is Hybrid Linear-Softmax Attention, which combines gated linear attention (O(N) complexity) with periodic gated-softmax anchors at a 3:1 ratio. This hybrid approach restores full-rank token interactions that pure linear attention lacks, avoiding quadratic complexity while maintaining expressiveness. Additionally, Block Attention Residuals (AttnRes) propagate completed block summaries into later linear layers, boosting deep-layer effective rank by ~12%.

In benchmarks, SANA-Video 2.0 achieves a VBench score of 84.30 at 480p in just 13.2 seconds on a single H100 GPU with 40-step sampling. Its compiled DiT forward pass is 3.2x faster than a matched full-softmax baseline at 720p/60s, and this gap widens with longer video durations. With full-stack Sol-Engine optimization (kernel fusion, caching, sparse attention), the 5B pipeline runs in 13.06s at 720p/5s, making it 120x faster than Wan 2.2-A14B on one H100.

The model is trained from scratch, learning the hybrid architecture directly rather than linearizing pretrained models. Proxy studies at reduced resolution established 25% softmax as the optimal quality-efficiency trade-off. The code is available on GitHub.

MetricSANA-Video 2.0 (5B)Full-Softmax BaselineWan 2.2-A14B
VBench Score (480p)84.30
Latency (480p, 40 steps)13.2s
DiT Forward Pass Speedup (720p/60s)3.2x1x
Pipeline Latency (720p/5s, with Sol-Engine)13.06s~1567s (est.)
Speedup vs. Wan 2.2-A14B120x1x
Source
Hugging Face · Hugging Face Papers
Related news
Research paper
Hugging Face 31 Aug 2026

Hugging Face Unveils StepGuard: Step-Level Guardrails for Safer AI Agents

StepGuard, a new step-level guard model from Hugging Face, audits agent actions before execution, reducing attack success rates by...

1
Research paper
Hugging Face 31 Aug 2026

Hugging Face Researchers Unveil ABot-Recon for Stable Long-Horizon 3D Reconstruction

ABot-Recon, a new streaming 3D reconstruction model from Hugging Face, achieves stable long-horizon performance using only local t...

1
Research paper
Hugging Face 31 Aug 2026

ContextPilot: Teaching Agents Proactive Context Management via Fine-Grained RL

Hugging Face researchers introduce ContextPilot, a framework that enhances long-horizon agent reasoning by expanding context-editi...

1