Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation
Haocheng Xi, Yiming Xie, Hexu Zhao, Yiwen Zhang, Michael Liu, Thomas Creavin, Kurt Keutzer, Xiuyu Li, et al.
cs.LG
Quick verdict
VDN adds a frame-wise delta-rule linear branch to video diffusion; strong systems work, honest speedup decomposition, but quality evidence is proxy-VQA only.
Summary
Video DeltaNet (VDN) accelerates video diffusion by replacing most quadratic video-to-video Softmax attention with a hybrid: a 15-frame sliding-window Softmax aligned to the VAE's five-frame chunks plus first/last-frame boundary anchors, alongside bidirectional linear memory. Its core operator, Video Delta Attention (VDA), extends the delta rule from single tokens to whole latent frames, solving a joint objective in closed form via (S_t+B_t)(I+A_t)^{-1}; the authors prove the inherited-state transition is non-expansive, avoiding the frame-size key scaling SANA-WM requires. Softmax remains responsible for text and audio interactions; separate gates, RMSNorm, and output projections calibrate the branches. A staged recipe — per-layer alignment, end-to-end alignment, LoRA refinement, and DMD2-style eight-step distillation — adapts the pretrained MiniMax H3 without training from scratch, using roughly 10,000 curated private clips. On a 14.3-second, 768p workload, the optimized backbone alone runs 2.6x faster on B200 and 3.2x on H200; with distillation and eight-GPU sharding, denoising completes in 6.70 s on eight B200s (119.3x versus single-GPU dense; the abstract cites 14.5x against a same-GPU-count dense baseline not tabulated in the text). Over 103 prompts, eight-step VDN-H3 matches or slightly exceeds 50-step dense H3 on five no-reference metrics, with slightly lower first/last-frame fidelity. Limitations: headline speedups bundle distillation and sharding; no human study; the VDA-versus-batched-rule comparison is analytical rather than empirical.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Video Delta Attention performs one joint, closed-form memory update per latent frame, is provably non-expansive on inherited state, and removes the need for SANA-WM's frame-size key scaling.
- A staged teacher-alignment plus LoRA plus eight-step distillation recipe converts the pretrained Softmax-only MiniMax H3 into a hybrid whose optimized backbone alone is 2.6x faster on B200 and 3.2x on H200 at the 14.3-s 768p workload.
- With distillation and eight-GPU head-sharded inference, VDN-H3 completes DiT denoising in 6.70 s on eight B200s — 119.3x versus single-GPU 50-step dense, and 14.5x versus a same-GPU-count dense baseline per the abstract.
- Eight-step VDN-H3 matches or slightly exceeds 50-step Dense H3 on five no-reference quality metrics over 103 prompts and clearly outperforms FastH3 at matched step counts.
Strengths
Principled frame-wise delta operator with a stability guarantee
VDA is derived from a joint regularized least-squares objective (Eq. 4) with closed-form update (S_t+B_t)(I+A_t)^{-1} (Eq. 5); Proposition 1 (App. A.1) proves the inherited-state transition is non-expansive, and App. A.2 gives a concrete repeated-key analysis (Eq. A.3) showing why VDA needs no SANA-WM-style 1/sqrt(U) key scaling.
Honest, decomposed efficiency accounting
Tables 2-3 separate backbone acceleration (2.6x B200 / 3.2x H200 at 50 NFE) from few-step distillation (6.3x gain) and eight-GPU sharding (7.4x gain), cumulating to 119.3x / 141.6x against single-GPU dense — the appendix also clarifies that significance asterisks compare only against 50-step Dense H3.
Practical adaptation recipe without from-scratch pretraining
Concrete systems engineering and public release
Matched-NFE competitive positioning
Weaknesses
Headline speedup framing outruns the architectural contribution
Quality evaluation is proxy-only and narrow
No empirical ablation of the core operator
Significant small regressions at the deployed configuration
Reproducibility constrained by private data and hardware
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Peer Review: Video DeltaNet (arXiv 2609.20744)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How much faster is Video DeltaNet than standard dense attention?
The architecture alone speeds up the denoiser backbone by 2.6x on one B200 and 3.2x on one H200 for a 14.3-second 768p video. The headline 6.70 s on eight B200s additionally includes eight-step distillation (6.3x gain) and eight-GPU sharding (7.4x gain), cumulatively 119.3x versus a single-GPU 50-step dense baseline; the abstract separately cites 14.5x versus a dense baseline on the same GPU count.
Does the hybrid attention hurt video quality?
On five no-reference metrics (VQAA, VQAT, Q-Align, FAST-VQA, DOVER++) over 103 prompts, eight-step VDN-H3 matches or slightly exceeds 50-step dense H3. First/last-frame conditioning fidelity drops slightly and significantly (PSNR 28.67 vs 28.85), and World Coherence dips marginally (1.77 vs 1.84). No human preference study is reported in the paper.
Is code or model weight access available for VDN-H3?
Yes. The paper links code at github.com/OpenVDN/vdn-minimax-h3 and weights at huggingface.co/OpenVDN/vdn-minimax-h3. However, the 10,015-clip training set is private, and reproducing the headline latencies requires eight NVIDIA B200 GPUs with the SGLang serving stack.
What is Video Delta Attention (VDA) in one sentence?
VDA is a delta-rule memory update computed once per latent video frame: all spatial tokens of the frame jointly define a new memory state via the closed-form solution (S_t+B_t)(I+A_t)^{-1}, and the authors prove this update never amplifies inherited memory (non-expansive transition).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Portions of Sections 4-5 (training/distillation detail, Table 1 context) were elided in text extraction; judgments lean on the preserved Section 6, figure traces, and appendices.
- All figures are available only as text traces; visual frame comparisons (Figs. 1, 5, 8-10) could not be inspected.
- The 8-GPU dense baseline behind the abstract's 14.5x same-GPU-count claim is not visible in the extracted tables.
- Exact statistical test types and p-values are unrecoverable from the extraction (only asterisk significance flags survive).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.