Showing 2026Show all
2 papers · 1 filter
cs.LG2026
RIPPLE: Generating Multi-Channel Phase, Not Recovering It
Jaehyuk Lee, Yeajin Lee, Dayeon Shin +1
Generative models synthesize magnitude spectra with high fidelity, while phase is delegated to a recovery module---Griffin--Lim, a vocoder, or a latent decoder---applied independen…
cs.LG2026
ASAP: Attention Sink Anchored Pruning
Jaehyuk Lee, Hanyoung Kim, Yanggee Kim +1
Vision Transformers (ViTs) face severe computational bottlenecks due to the quadratic complexity of self-attention at high resolutions. Existing token reduction methods rely on loc…