Search papers, labs, and topics across Lattice.
1
0
3
Naively quantizing autoregressive video diffusion models tanks performance due to exponentially increasing error accumulation across frames and heterogeneous outlier patterns, but Q-ARVD solves it.