Autoregressive video generation stops sharing one weight set for context writing and denoising: 5-second training windows sustain 24 hours of continuous generation

An autoregressive video diffusion model plays two roles at once: denoising the current frames, and encoding already generated frames into key-value representations that later predictions read. SGF+ splits these roles onto two independent parameter sets so each updates separately, while causal attention keeps their forward interaction. The earlier Self Gradient Forcing already let future generation losses supervise context writing, but writing and denoising still shared one parameter set, leaving residual degradation on long sequences and, in a small number of generated videos, lower initial-frame quality than Self Forcing. When gradient contributions under the same objective are partitioned by role during SGF training, the mean gradient directions of the two roles exceed 90 degrees apart in both the attention and FFN module families, and all 512 measured prompt and denoising-timestep pairs have negative cosine similarity.

The implementation gives the context writer and the denoiser separate parameters, both initialized from the same autoregressive diffusion model, and keeps SGF's two-pass training structure. The first pass remains a no-gradient autoregressive rollout, and the second pass reconstructs a differentiable key-value path in parallel, so future generation losses supervise context writing without backpropagating through the full sampling trajectory. Training uses a Wan2.1-T2V-1.3B causal video diffusion student, a frozen Wan2.1-T2V-14B teacher, and a separately trained fake-score network, on 5s video windows with filtered and expanded VidProM prompts; the 60s and 240s evaluations use VBench-Long prompts and 128 MovieGen prompts respectively.

Parameter separation improves visual quality and long-horizon consistency at the cost of more parameters and memory, while per-step computation stays roughly the same. Generator parameters double from 1.4B to 2.8B, but each token uses only its role-specific parameters within a forward pass, so computation does not double. On the evaluated 81-frame workload, inference time moves from 4.962s to 4.969s and memory from 24.85GB to 27.96GB; training peak memory rises from 97.83GB to 98.15GB, stable memory from 70.06GB to 73.60GB, and time per step from 11.79s to 12.76s, about 8.2% more.

The report covers parameter sharing between context writing and denoising in autoregressive video diffusion, with evaluation limited to VBench quality and consistency metrics at 5s, 60s, and 240s generation horizons in framewise and chunkwise generation. The intended readers are engineers and researchers working on video generation, streaming inference, and world-model training. The source treats role-specific parameterization as complementary in mechanism to history construction and context management, and lists model merging and other ways of coordinating role optimization as future work.

Dynamic Degree is the main exception, where Self Forcing or Self Gradient Forcing scores higher; the source notes that scene jumps, object deformation, subject disappearance, or additional people appearing can produce large but incoherent apparent motion that inflates this metric, so SGF+'s lower value does not imply worse generation quality. Partial-separation variants that separate only attention or only FFN underperform the full model on most quality and consistency metrics. The source also states that the minibatch negative-alignment result requires negative alignment of the aggregated gradients, and that negative alignment of individual samples or selected modules does not establish this condition.