Daily Paper Briefing

Abstract Overview

This paper studies long-horizon autoregressive video diffusion and argues that existing Self Forcing training still leaves a missing supervision path for how self-generated history is written into the model’s key-value memory. The authors call this the historical context-gradient gap: future losses can train later tokens to read cached history, but they do not backpropagate into the clean-timestep computation that wrote earlier self-generated latents into the cache. To address this, the paper proposes Self Gradient Forcing (SGF), a two-pass training procedure with a no-gradient serial rollout followed by a parallel reconstruction step that restores gradients through reconstructed context K/V representations. Experiments cover frame-wise and chunk-wise generation at 5s, 60s, and 240s horizons under multiple initializations, with emphasis on native extrapolation far beyond the 5-second training window.

Novelty

The main novelty is the identification of the historical context-gradient gap in frozen-cache Self Forcing and the introduction of SGF to recover that supervision without full backpropagation through the entire rollout. The method is distinctive in using a bounded two-pass reconstruction that keeps the sampled rollout fixed while allowing future losses to supervise memory writing for self-generated context.

Results

Across controlled comparisons, SGF is reported to be broadly comparable to Self Forcing at 5 seconds while improving most long-horizon metrics at 60 and 240 seconds, especially subject consistency, background consistency, imaging quality, motion smoothness, and flickering. Blind human preference scores are positive for SGF in all reported long-horizon matched comparisons, and the training analysis shows modest overhead relative to frozen-cache Self Forcing while avoiding the out-of-memory behavior of a directly differentiable cache.

Key Points

1. SGF targets a specific training deficiency in autoregressive video diffusion: future losses do not normally supervise how self-generated history is encoded into causal K/V memory.
2. The method uses a no-gradient rollout pass plus a gradient-enabled parallel reconstruction pass, restoring context-memory gradients without retaining a full recurrent autograd graph.
3. Empirical evaluation suggests the approach mainly benefits long-horizon extrapolation rather than short-horizon quality, with stronger temporal stability and consistency over 60-second and 240-second videos.

References

• arXiv: https://arxiv.org/abs/2607.20368v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2607.20368v1
• Hugging Face Papers: https://huggingface.co/papers/2607.20368
• Project: https://zhuang2002.github.io/SelfGradientForcing