LANCE boosts image compression by 2.99% with adaptive neural context

LANCE (Locally Adaptive Neural Context Estimation) addresses a key limitation in overfitted image compression (OIC) frameworks like Cool-Chic: their globally signaled autoregressive models struggle with non-stationary image statistics. The authors introduce a forward-signaled spatial hyperprior that tells the decoder how to adapt the entropy model per region, without costly backward signaling. To keep overhead low, they combine a classic Median Edge Detector (MED) with a lightweight learned context model—a predictive coding scheme that balances efficiency and adaptability.

Experimental results show clear gains over Cool-Chic 4.0: at the high end of decoder complexity (606–1481 MAC/pixel), LANCE achieves 1.40% BD-rate reduction on Kodak and 1.97% on CLIC 2020. At the low end, improvements jump to 2.41% and 2.99%, respectively. Qualitative analysis reveals the learned hyperprior naturally segments images into regions of similar statistics, creating an automated, content-aware adaptation layer. This work was submitted to IEEE TCSVT and represents a practical step toward smarter, more efficient learned image compression.