目前釋出的 SparkWan 權重涵蓋部分 Wan 2.1、Wan 2.2 設定,各模型支援的稀疏率不一,不能假設每個版本都能達到 97%。[9][10][11][12]
How does the open-source SparkDiffusion framework from Peking University, Tsinghua University and Alibaba accelerate Wan 2.1 and Wan 2.2 vidSparkDiffusion combines sparse attention, few-step distillation and FP8 kernels to accelerate Wan video generation.
AI 提示詞
Create a landscape editorial hero image for this Studio Global article: How does the open-source SparkDiffusion framework from Peking University, Tsinghua University and Alibaba accelerate Wan 2.1 and Wan 2.2 vid. Article summary: SparkDiffusion speeds up Wan video generation by combining three changes: it computes far less attention, distils generation into a few steps, and runs the resulting model with FP8 fused kernels. The researchers report u. Topic tags: general, academic, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, char
openai.com
研究團隊提出的 SparkDiffusion,目標是縮短阿里巴巴 Wan 影片生成模型的等待時間。它不是只靠單一技巧提速,而是把注意力計算、生成步數與每一步的執行效率一起最佳化:採用稀疏注意力、透過蒸餾減少生成步數,再以 FP8 融合核心加快運算。68
團隊在特定 Wan 2.1 基準測試中報告最高約 265 倍加速。不過,這個數字取決於模型、影片規格與硬體,不能直接視為所有生成任務都能達到的速度提升。68