Interactive Multi-Shot Video Creation

Summary & Methodology Analysis

Interactive multi-shot video creation requires combining composition, reference-guided generation, and editing into a single pipeline, but these operations typically run as isolated tasks with rapidly growing memory costs. To tackle this, the authors developed a system that manages heterogeneous context at every interaction turn while maintaining predictable computational overhead during denoising, which is the iterative process diffusion models use to clean random noise into a final image or video. The architecture relies on role-aware rotary coordinates to place reference, history, source, and target latents into a unified model while preserving distinct temporal semantics, and uses a cacheable fixed-budget context mechanism that encodes context once per round and applies asymmetric attention, a variant of the core transformer mechanism that restricts how sequence elements interact, to reuse keys and values without re-encoding. A ConstraintSink reserves mandatory budget for provided references and temporally aligned source blocks, while budgeted dynamic routing allocates residual capacity for content-relevant retrieval from source and history blocks independently. To train this efficiently, privileged context distillation uses a dense full-context teacher to supervise a sparse student through consistency distillation, a training technique that collapses multi-step generation into fewer steps, followed by distribution matching refinement on deployment-matched rollouts to recover visual details and improve consistency.

The system interacts with and builds upon several models and datasets, including Wan2.1-T2V-1.3B, ViCLIP, VBench, DINO, CLIP, Gemini 2.5 Pro, MultiShotMaster, LongLive, Infinity-RoPE, ShotStream, Phantom, VideoCoF, LucyEdit, StreamEdit, LiveEdit, HuMo, Kiwi-Edit, Ditto-1M, and Kling 3.0. In performance benchmarks, ContextMaster achieves 16 frames per second on a single GPU and reaches a Task Fulfillment score of 4.17 in text-to-multi-video compared to the next-best baseline's 4.03.

Despite these efficiency gains, the system has a notable limitation. Throughput gradually decreases as the accumulated history grows because the context branch performs bidirectional prefill over the full history. The paper does not specify any other hardware requirements, dollar costs, or training durations beyond what is listed.