SGF+

SGF+ builds on Self Forcing (SF) and Self Gradient Forcing (SGF). SF blocks gradients from context writing, SGF restores them on shared parameters, and SGF+ separates the two roles into independent parameter sets that are trained jointly with the original generation objective and no auxiliary losses. Training uses a two-pass procedure: a no-gradient autoregressive rollout is recorded, then the forward computation is rebuilt from detached context and noisy target latents so future-generation losses supervise context writing through differentiable KV states. The project reports that across 128 prompts and four denoising timesteps, all 512 paired cosine similarities between context-writing and denoising gradients are negative in each module family. It is trained on the Wan2.1-T2V-1.3B base model with Wan2.1-T2V-14B as teacher; checkpoints are 11.4 GB. Evaluated baselines are SF and SGF, in framewise and chunkwise generation.