ContextMaster: Interactive Multi-Shot Video Creation via Fixed-Budget Sparse Context Routing
The paper frames video generation as an ongoing, multi-shot workflow with shared memory, not isolated prompts.
ContextMaster is built to generate from text, follow references, and edit footage while keeping prior shots available. Its sparse context routing keeps the model’s per-step context cost fixed as history grows. The authors say distillation from a dense teacher helps preserve full-context behavior under that sparse setup. They report stronger task fulfillment and shot-to-shot consistency than specialized baselines, with 16 FPS on a single GPU. HF Daily Papers' note
ContextMaster is built to generate from text, follow references, and edit footage while keeping prior shots available. Its sparse context routing keeps the model’s per-step context cost fixed as history grows. The authors say distillation from a dense teacher helps preserve full-context behavior under that sparse setup. They report stronger task fulfillment and shot-to-shot consistency than specialized baselines, with 16 FPS on a single GPU. HF Daily Papers' note
score 5