SGF+: Decoupling Gradient Flows for Autoregressive Video Generation

SGF+. Decoupling Gradient Flows for Autoregressive Video Generation. By Zihan Su and twelve others, from Tsinghua University, JD's Joy Future Academy, and one other institution. Published on the seventh of October.

An autoregressive video generator does two jobs at every step. It denoises the frames it is currently producing, and it writes their key-value representations as context that future predictions will read. Those two roles usually share the same parameters. The authors look at the gradients each role produces and find that they behave differently: distinct patterns, and a systematic negative alignment between them. One role pulls the shared weights against the other, and that tension hinders optimising visual quality and temporal consistency together.

Self Gradient Forcing Plus, written SGF Plus, answers with a separation. Context writing gets its own parameters and denoising gets its own, so neither role's gradient has to fight the other's for the same weights. What is kept is the interaction between them, which still runs through causal attention. The two roles are split in their parameters, not in what they see of each other.

Training stays close to what it was. Both roles are optimised jointly on the original generation objective, with no auxiliary losses added. The context writer is not given a target of its own. It is supervised through its contribution to future predictions, so it is judged by how useful the context it writes turns out to be for the frames that come after.

This simple change improves visual quality and long-horizon consistency over the evaluated baselines, in both framewise and chunkwise generation. It does so without additional video training data and without a longer training horizon. The headline result is about extrapolation. Trained only on five-second rollouts, SGF Plus supports continuous generation for up to twenty-four hours, with no long-video fine-tuning.

The authors draw a design principle from this: role-specific parameterization, giving each job in an autoregressive generator its own weights, as a route to high-quality video and to native long-horizon extrapolation. The gap between a five-second training window and a day of continuous output is what makes the principle concrete.