SGF+: Decoupling Gradient Flows for Autoregressive Video Generation
SGF+ decouples video context writing from denoising, enabling 24-hour generation from 5-second training rollouts.
SGF+ (Self Gradient Forcing Plus) addresses conflicting gradients in autoregressive video generation by giving context writing and denoising separate parameters while preserving interaction through causal attention. Both roles are optimized with the original generation objective, and context writing is supervised via its effect on future predictions, without auxiliary losses or extra video data. The method improves visual quality and long-horizon consistency over the evaluated baselines in framewise and chunkwise settings. Models trained only on 5-second rollouts support continuous generation for up to 24 hours without long-video fine-tuning.
- Separates parameters for context writing and frame denoising.
- Keeps the roles coupled through causal attention without auxiliary losses.
- Improves quality and consistency in framewise and chunkwise generation.
- Models trained on 5-second rollouts generate continuously for up to 24 hours.
- AI models
- SGF+
From the original article
1 min read at huggingface.co
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.
By Zihan Su, Junhao Zhuang, Yaowei Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.