SGF+: Decoupling Gradient Flows for Autoregressive Video Generation

2026-10-07

AI Summary

This study addresses the gradient conflict in autoregressive video generation arising from shared parameters between denoising and context writing, which constrains generation quality. To overcome this limitation, we propose SGF+, a method that decouples gradient flows by assigning independent parameters to these distinct functional roles. The modules are interconnected via a causal attention mechanism, enabling joint optimization under the original objective without requiring auxiliary losses. Our approach significantly enhances long-horizon temporal consistency. Notably, when trained on merely five seconds of data, the model is capable of continuously generating high-quality videos spanning up to 24 hours, demonstrating substantial improvements in both efficiency and output fidelity for extended video synthesis.

Abstract

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 original generation objective without auxiliary losses, with context writing supervised through its contribution to future predictions. This simple change improves visual quality and long-horizon consistency over the evaluated baselines in both framewise and chunkwise generation, without additional video training data or a longer training horizon. Trained on only 5s rollouts, SGF+ supports continuous generation for up to 24 hours without long-video fine-tuning. These results highlight role-specific parameterization as an effective design principle for high-quality autoregressive video generation and native long-horizon extrapolation.

Problem

• Autoregressive video generation
• Gradient conflict
• Parameter sharing
• Temporal consistency
• Joint optimization

Innovation

• Autoregressive Video Generation
• Gradient Decoupling
• Self Gradient Forcing
• Role-specific Parameterization
• Long-horizon Extrapolation