1

CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration

arxiv.org

CacheScope identifies high-risk reuse steps online and selectively falls back to full denoising computation while leaving the underlying cache reconstructor unchanged, requiring no additional training, gradients, auxiliary networks, or prompt-specific offline schedule optimization.

Content history (1)
  • 2026-10-12 · Summary · vi · Replaced stored text
    CacheScope đưa ra phương pháp phân bổ exact-compute cho việc tăng tốc diffusion mà không …
    Việc tái sử dụng đặc trưng không cần huấn luyện tăng tốc suy diễn mô hình transformer lan…
Sign in to suggest edits

Key sources

  1. SOURCEarxiv.org
Markdown