Video generation has an awkward piece of arithmetic: a ten-plus-second 768p clip takes diffusion models minutes to render — an order of magnitude slower than playback, which rules out livestream and interactive use. Video DeltaNet (VDN), released this week by a team from UC Berkeley, Impossible, and UT Austin, narrows that gap decisively: VDN-H3, deployed on MiniMax H3, returns a finished 14.4-second 768p video in about 9.0 seconds on 8 B200 GPUs after warm-up, with the denoising loop itself taking 6.9 seconds — generation now runs faster than playback (project page: openvdn.github.io).

Attention is the dominant bill

Video diffusion models repeatedly process long spatiotemporal token sequences during denoising, and softmax attention cost grows quadratically with sequence length. The project page puts a stark number on it: on a frontier omni-model like H3, softmax attention accounts for more than 85% of total runtime. LLMs already answered this question — Gated DeltaNet, Kimi Linear and other linear attention families cut the cost to linear — but porting them straight into video models breaks quality: global features like subject identity, scene layout, and temporal dependencies degrade.

The VDN split: local Softmax plus long-range linear memory

VDN refuses the either/or. Video-to-video attention is split into two complementary branches: a sliding-window softmax branch handles local detail — chunks of five consecutive latent frames, each attending to itself and its neighboring chunks — while bidirectional linear attention summarizes distant context, with forward and reverse states covering frames before and after the local window. A 4-way boundary anchor lets every frame see the first and last frames (and vice versa), adding only 3.57% attention density while stabilizing long-range consistency. Interactions involving text or audio keep softmax untouched.

The genuinely new part is Video Delta Attention (VDA): it upgrades the token-wise delta rule from LLMs into a frame-wise one. All spatial tokens within a frame jointly solve a regularized objective instead of issuing independent reads and writes — mathematically, an (I+A) inverse solve on the frame-level Gram matrix. Two consequences: the inherited-state transition is non-expansive by construction, with no frame-size rescaling needed (unlike SANA-WM), and repeated evidence directions damp each other while independent directions keep full write strength — a hundred patches pointing at one memory direction grow the evidence weight toward, but never past, one.

The speedup ledger and a three-stage recipe

Inserting a randomly initialized linear branch into a pretrained H3 cannot be brute-forced; it perturbs the calibrated softmax path. The team's recipe: layer-wise alignment first, then end-to-end branch adaptation (softmax branch, QKVO projections, and FFN all frozen), then LoRA co-adaptation. The softmax gate stays frozen until the final stage, so the optimizer cannot take the shortcut of simply suppressing the softmax branch.

By the numbers: on a single B200, 50-NFE dense H3 takes 13.95 minutes to denoise a 14.4-second video; VDN-H3 drops that to 5.34 minutes, with a single block going from 332.5ms to 125.3ms (2.65x). Across 8 GPUs, Ulysses parallelism cuts per-step latency from 6.46s to 1.62s; assigning 5 GPUs to the softmax branch and 3 to VDA (a profiling-driven split) reaches 1.405s, 13.3% below standard Ulysses. With 8-step distillation (deliberately avoiding an aggressive 4-NFE setting), the DiT loop completes in 11.23 seconds — 74.5x versus the dense single-GPU baseline and 10.7x versus the dense 8-GPU baseline. The paper abstract's matching numbers: 6.70 seconds of denoising for a 14.3-second 768p video, a 14.5x speedup over the 50-step dense baseline on the same GPU count (arXiv:2609.20744).

So what

"Faster than playback" is not just a vanity metric: once denoising beats playback duration, streaming and livestream generation becomes economically writable — the livestream in the paper's title is not rhetoric. Note also that VDN and last month's FastH3 (4-step distillation) accelerate on different axes — attention architecture versus step count — and should in principle stack. The full migration path of the DeltaNet family from LLMs into video diffusion, including the three-stage training recipe and SGLang serving, is now open (code and weights: OpenVDN/vdn-minimax-h3 on GitHub and Hugging Face). The open question for whoever runs next: the official "near-lossless quality" claim still awaits third-party replication.