Researchers have released a new paper on Hugging Face titled SGF+: Decoupling Gradient Flows for Autoregressive Video Generation, addressing a core tension in autoregressive video models. In these systems, the current frames must be denoised while their key-value representations are written as context for future predictions. The paper finds that these two roles typically share parameters, and their gradients exhibit distinct patterns and systematic negative alignment, which hinders joint optimization of visual quality and temporal consistency.
To address this, the authors 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.
According to the paper, 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. Notably, trained on only 5-second rollouts, SGF+ supports continuous generation for up to 24 hours without long-video fine-tuning.
The authors highlight role-specific parameterization as an effective design principle for high-quality autoregressive video generation and native long-horizon extrapolation. The paper, project page, code, and model are publicly available.