EffectLearner: World-Aware Object-Effect Reasoning for Real-World Video Object Removal
The paper targets the leftover effects objects leave behind, not just the objects themselves.
EffectLearner pairs a VLM-based reasoner with a DiT video eraser so removal is guided by compact context about object-induced effects. The authors say this helps with detached, weakly correlated, compositional, and evolving effects in real-world video. They also introduce EffectWorld, a paired dataset for complex object-effect cases, plus a progressive training curriculum. On ROSE-Bench and EffectWorld evaluations, they report stronger results than existing baselines on most metrics. HF Daily Papers' note
EffectLearner pairs a VLM-based reasoner with a DiT video eraser so removal is guided by compact context about object-induced effects. The authors say this helps with detached, weakly correlated, compositional, and evolving effects in real-world video. They also introduce EffectWorld, a paired dataset for complex object-effect cases, plus a progressive training curriculum. On ROSE-Bench and EffectWorld evaluations, they report stronger results than existing baselines on most metrics. HF Daily Papers' note
score 4