跳到正文
DEV Community· aimodels-fyi·· 10 天前AI 評分41

Wave Forcing 以交錯去噪每輪串流生成一個影片區塊

Wave Forcing Streams Video at One Chunk per Round with Staggered Denoising

AI 導讀

Wave Forcing 結合區塊式因果注意力與交錯去噪,在管線填滿後,每個並行前向輪次輸出一個影片區塊。於 Wan2.1-T2V-1.3B 上,五步版本的 VBench Total 為 81.64,高於 Rolling Forcing 五步版本的 81.15,測得的動態幅度亦大幅增加。方法透過重疊去噪階段與快取更新提升生成效率,同時保留每次去噪更新。

正文

Cover image for Wave Forcing Streams Video at One Chunk per Round with Staggered Denoising

aimodels-fyi

This is a Plain English Papers summary of a research paper called Wave Forcing Streams Video at One Chunk per Round with Staggered Denoising. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.

Wave Forcing targets one chunk per concurrent round

Wave Forcing (WF) combines chunk-wise causal attention with a staggered denoising pipeline to generate streaming video at one output chunk per concurrent forward round after pipeline fill. On Wan2.1-T2V-1.3B, its five-step version reaches VBench Total 81.64, compared with 81.15 for five-step Rolling Forcing, while substantially increasing measured motion.

The efficiency problem WF addresses comes from the clean-history caching convention used in the study: each chunk needs denoising passes and a separate store pass to refresh its key/value history. That makes even one-step generation require two forward rounds per chunk when executed serially. WF overlaps denoising stages and cache refresh across full-model replicas while retaining every denoising update.

Causal wavefront attention and three-stage training

WF lets a newer, noisier chunk attend to earlier chunks that are already cleaner, even when those predecessors have not finished denoising. Its chunk-wise causal mask keeps full attention within each chunk and permits access to cached history, live store context, and eligible earlier chunks; later, noisier chunks cannot condition earlier ones....


Continue reading the full paper summary on AIModels.fyi →

來源:DEV Community · dev.to