zihan-su/self_gradient_forcing_plus is a Python repository proposing SGF+, which optimizes context history and denoising under different parameters in autoregressive video generation. The README explains that the gradients the two tasks produce on shared parameters can partially cancel each other, and presents an approach that separates their roles. The paper is linked as arXiv 2610.10429.
The central claim is that a 5-second-long training window can produce video rollouts of up to 24 hours in length. The repository analyzes gradient distributions, mean directions, and pairwise cosine similarity to explain the conflict phenomenon. On 2026-10-09 it published reproduction code in the gradient_conflict directory. The README states that this phenomenon was observed not only on Wan2.1-1.3B but also on LTX-2.5. The README also documents the procedure for downloading the Wan2.1-T2V-1.3B·14B bases and the released SGF+ inference checkpoints. So the point that parameter separation is a design aimed at long autoregressive generation is concrete.
That said, the README has no numerical benchmark table, so the size of the performance gain and any comparison with other methods cannot be confirmed. The 24 hours is a rollout claim made by the repository, not a comparative result verified here. The gradient conflict analysis supports the motivation for the approach, but it does not show how far quality is maintained over long generations. This repository gives grounds for looking at a new training design, but it is hard to treat it as proof of an advantage in long-duration generation performance.