Researchers have developed CALM (Calibrated Allocation for Layer-Head Merging), a novel method to reduce KV-cache memory usage in autoregressive video generation. Unlike previous approaches that applied uniform merging ratios, CALM calibrates merging ratios based on the sensitivity of different attention layers and heads to token merging. This offline calibration process uses attention-output error as a proxy for video quality, allowing for efficient determination of optimal layer-head risk maps. Experiments show CALM outperforms uniform merging, reducing LPIPS by approximately 21% while maintaining higher video quality. AI
IMPACT This research could lead to more efficient video generation models, reducing computational costs and improving accessibility.
RANK_REASON The item describes a new research paper detailing a novel method for optimizing video generation models. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
