WorldAttention pairs hierarchical KV caching with hybrid sparse attention to reach 22 FPS on a single H100 and 0.9472 subject consistency on VBench-Long

WorldAttention leads on generation quality across VBench-Long and InterVBench, with subject consistency of 0.9472 and 0.9668, background consistency of 0.9691 and 0.9614, and motion smoothness of 0.9915 and 0.9972. Prior interactive long-video methods trade off between sliding-window caches that discard early visual context and sparse-retrieval caches whose KV store grows linearly with length; this work co-designs the attention structure with KV management to keep full history while improving consistency metrics. Evaluated on VBench-Long using LongLive's 160 interactive 60-second prompts (each with six successive 10-second prompts) and on InterVBench with 200 videos longer than 50 seconds, reporting five VDE drift metrics and five VBench metrics; Tables 1, 2, and 12 give per-metric comparisons against baselines including MAGI-1, Self Forcing, SkyReels-V2, LongLive, and BIFE.

The Hierarchical KV Cache (HKV) partitions historical KV pairs into semantic pages of 8 frames each, stores them across L0 GPU SRAM, L1 GPU HBM, L2 CPU DRAM, and L3 NVMe, and uses two-stage prompt-level and page-level retrieval to assemble a top-Kp active cache. Existing schemes either retrieve at chunk granularity, loading whole chunks and creating redundancy, or keep only a recent window; HKV refines retrieval to the page level and automatically demotes unselected pages to cheaper memory so GPU occupancy does not grow with video length. Table 5 shows HKV (Psize = 8, Kp = 4) reaching a quality score of 85.53 at 22.0 FPS versus 82.52 for sliding window and 84.82 for re-caching 32 frames (which runs at only 7.10 FPS); Table 6 compares page-only, prompt-only, chunk-level, recent-page, random-page, and mismatched-page retrieval under the same four-page budget, with HKV scoring 25.01 CLIP on the 50-60s segment versus 21.48, 20.12, and 17.36.

Hybrid Sparse Attention (HSA) fuses a linear global branch with a head-adaptive block-sparse branch through learned gates, where the linear branch uses Linformer-style low-rank projection to cut complexity from O(L^2) to O(Lr) and the sparse branch selects the minimal block set per head whose cumulative attention mass exceeds a threshold tau_h. Earlier sparse and linear attention work targets inference-time acceleration without adapting to chunk-wise long-horizon video generation or co-designing with KV management and hardware kernels; HSA lets each head set its own sparsity from its attention distribution and distinguishes inter-chunk (Binter = 128) from intra-chunk (Bintra = 64) granularity. Table 4 ablations show linear-only at 0.9021 subject consistency and sparse-only at 0.9365 with tau_max = 1.00 and tau_min = 0.35, while combining both reaches 0.9472; Figure 6 shows the linear branch tracks full-attention global attention mass and output cosine similarity more closely than the SLA and SANA-Video linear attention variants.

Hardware-oriented kernel customization turns HSA's theoretical efficiency into throughput: the sparse attention kernel peaks at 14.02x over FlashAttention-3, combining with HKV for a 2.21x end-to-end speedup that persists on B200, 14B models, 720p, and 90-second horizons. The work reports that the sparse block attention kernel accounts for only 1.86% of total latency while preprocessing such as KV cache reorganization and contiguous memory copies dominates, motivating optimization of the whole system rather than a single attention kernel. Table 9 gives the incremental end-to-end speedups: baseline 1.00x, plus HKV 1.15x, plus HSA 1.91x, plus kernel customization 2.21x; Table 10 breaks down HSA execution time; Table 11 reports 2.06x to 2.46x across H100 and B200, 1.3B and 14B, 480p and 720p, and 60s and 90s settings.

The result targets text-conditioned interactive video world models within a chunk-wise autoregressive diffusion framework for minute-long generation, with evaluation on VBench-Long (LongLive's 160 interactive 60-second prompts) and InterVBench (200 videos longer than 50 seconds), and efficiency results spanning H100 and B200, 1.3B and 14B, 480p and 720p, and 60s and 90s configurations. In the limitation and future work section, the authors state that the current framework mainly uses textual prompts as the control signal and does not yet model action signals, control trajectories, or embodied interaction cues; extending to action-conditioned generation would let the model serve as a video world model for controllable environment simulation and decision-driven video prediction. For a reader, this means the architecture's value lies first in interactive generation settings that need long-range memory and low latency rather than in every video generation task.

Open questions include how two-stage retrieval behaves on genuinely multi-event prompts, since the authors note that top-k chunk retrieval could be more robust for multi-event prompts but that retrieving more chunks dilutes the fixed page budget under the current setting, which is why Top-1 is used; how the page size and retrieved-page count (Psize, Kp) trade off at longer horizons or higher resolutions; and whether the HKV and HSA designs still hold when the control signal extends from text to actions or trajectories. In addition, this summary is based on the full paper text and the homepage evidence bundle and does not include implementation details from the code repository or project website, so reproduction-level specifics still require the original text and appendices.