Researchers have found a way to make long-context diffusion models significantly faster by skipping attention work they were already repeating.
The paper introduces FlashBlock, a technique that targets a specific inefficiency in block diffusion — a popular inference method used in diffusion language models and video generators. In block diffusion, the model processes content in chunks and caches key-value pairs to avoid recomputing everything from scratch. The problem: as context grows, attention over that cache gets recomputed at every diffusion step anyway. FlashBlock's insight is that attention outputs from tokens outside the current block stay nearly constant across steps, while attention within the block changes meaningfully. By caching and reusing the stable external attention outputs, the system cuts redundant computation without touching the underlying diffusion process.
The gains are concrete: up to 1.44x higher token throughput and up to 1.6x faster attention time, with what the authors describe as negligible quality loss. The technique also stacks with sparse attention methods — rather than competing with existing efficiency tricks, it acts as a complementary layer, recovering accuracy that aggressive sparsification would otherwise sacrifice.
Diffusion-based text and video generation has been gaining ground as an alternative to autoregressive models, partly on the promise of parallelism and speed. FlashBlock addresses a bottleneck that grows worse at exactly the scales — long videos, extended documents — where diffusion is most appealing. Whether the "negligible" quality caveat holds up outside controlled benchmarks is the question practitioners will want answered before shipping this.