2026-10-08 · America/Los_Angeles · 论文 · #9

SGF+: Decoupling Gradient Flows for Autoregressive Video Generation

Autoregressive video generation requires denoising the current frames while writing their key-value representations as context for future predictions. However, these two roles typically share parameters, and we find that their gradients exhibit distinct patterns and systematic negative alignment, hindering the joint optimization of visual quality and temporal consistency. We introduce Self Gradient Forcing Plus (SGF+), which assigns separate parameters to context writing and denoising while preserving their interaction through causal attention. Both roles are jointly optimized using the…

热度 56.9 / 100；排名与评分保留该期记录。