Existing video object removal methods are slow and non-causal, often taking minutes per video and preventing real-time interaction. StreamErase aims to change that by enabling sub-second latency and up to 26 FPS streaming generation on a single GPU.

The approach first augments a text-to-video model with mask input to generate high-quality videos that erase target objects and synthesize coherent backgrounds. The erasing capability is then progressively transferred from a powerful bidirectional foundation model to an efficient unidirectional streaming eraser, allowing long-horizon erasing without sacrificing visual fidelity.

To support long or potentially infinite horizons, the authors designed sliding-window causal attention with attention sinks. By incorporating self-rollout with attention sinks and KV cache rolling during training, they simulate inference-time extrapolations with a fixed context window, keeping generation speed constant as video length grows.

The paper also introduces a dedicated benchmark, LongErase, for streaming long-video object removal. The model reportedly performs strongly against existing video object erasing models and generalizes well to real-world scenarios, uniquely enabling infinite-length streaming.