Research map / Video / Video generation

VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion

NeurIPS 2026 · Video › Video generation · approach: Efficient video diffusion models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces VideoMLA, an MLA-style latent KV cache for chunk-causal autoregressive video diffusion (Wan2.1-T2V-1.3B), cutting per-token per-layer KV cache by 92.7% and analyzing why low-rank latent attention works despite non-low-rank pretrained spectra.

Problem

Dense per-head KV caches dominate memory and latency in streaming long-horizon causal video diffusion, and prior work leaves the per-head KV layout unchanged.

Approach

Replace per-head keys and values with a shared low-rank content latent plus a head-shared decoupled 3D-RoPE positional key, and show the MLA bottleneck rather than the pretrained spectrum determines the effective rank.

Key results

• On VBench at 60s, VideoMLA achieves overall score 0.859, the highest among evaluated methods, with best dynamic degree 0.958.
• VideoMLA reaches 23.96 FPS throughput on a single B200 versus 18.06 for Self-Forcing, a 1.23x speedup at batch size 1 (per-token cache 224 scalars vs 3072).
• Under a fixed single-B200 memory cap, MLA supports 4.6x to at least 11.4x larger non-OOM batches; the default d_c=192 gives 8.0x batch headroom.

Surprising findings

• Pretrained Wan-1.3B [W_K; W_V] is not low-rank (99%-energy effective rank exceeds 1300 in every layer), yet the MLA-constrained model retains quality.
• Both SVD and random initialization saturate the rank budget from initialization, indicating the MLA bottleneck rather than the pretrained spectrum sets the effective rank.

Limitations

• Small latent budgets such as d_c=64 lose fine-grained detail and degrade quality, making d_c a quality-efficiency trade-off.
• Experiments are limited to Wan2.1-T2V-1.3B and minute-scale generation; larger backbones, higher resolutions, longer horizons, and prompt switching are untested.
• At 30s, VideoMLA is second-best overall (0.859 vs Reward Forcing 0.863) and lower on subject consistency and background consistency than some baselines.

Experimental setup

Datasets: OpenVid-1M

Benchmarks: VBench user study

Metrics: VBench overall aesthetic quality background consistency dynamic degree imaging quality motion smoothness subject consistency CLIP-T CLIP-F HPSv3

Baselines: CausVid Self-Forcing Rolling-Forcing Causal Forcing Reward Forcing Deep Forcing LongLive Infinity-RoPE LongSANA NOVA Pyramid Flow

Builds on: Multi-Head Latent Attention Self-Forcing Causal Forcing Wan2.1

Closest papers by meaning

• Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.92)
• Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.91)
• From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
• Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model (CVPR 2025 · similarity 0.89)
• FasterCache: Training-Free Video Diffusion Model Acceleration with High Quality (ICLR 2025 · similarity 0.89)
• Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.88)