HeatKV method compresses visual autoregressive model KV-cache

Researchers have developed HeatKV, a novel method to compress the KV-cache memory used by visual autoregressive models. This technique tunes cache allocation for each attention head based on its focus on previously generated image scales. HeatKV achieves a 2x higher compression ratio compared to existing methods for the Infinity-2B model, while maintaining image quality and prompt alignment. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a method to significantly reduce memory requirements for visual autoregressive models, potentially enabling larger models or faster generation on constrained hardware.

RANK_REASON The cluster contains an arXiv paper detailing a new technical method for model compression. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• HeatKV
• Infinity-2B
• Jonathan Cederlund

• paper
• infra