Scaling backbone for complexity

Background: Generating long, coherent, multi-shot video sequences for narrative storytelling presents challenges related to maintaining visual consistency across scene transitions and managing error accumulation in autoregressive generation.

Question / Future Work: While the current model addresses error accumulation via a two-stage self-forcing distillation strategy and maintains consistency using a dual-cache mechanism, visual artifacts and inconsistencies are still observed when scenes and text prompts are highly complex. Scaling up the backbone model size is suggested as a potential remedy for improving performance and stability in these challenging, complex scenarios.