GRNEdit: Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks
GRNEdit frames video editing as bit-level retain-or-flip decisions, aiming to cut conditioning cost without giving up edit quality.
The paper proposes a lightweight two-stage framework built on Generative Refinement Networks. Its first stage turns source video codes into evidence signals and trains an empty instruction as a no-edit reconstruction path. Its second stage compares the edited state with that source-preserving state to revise unresolved target bits. The authors report that GRNEdit-2B and GRNEdit-8B use less than 3% conditioning parameters, train on 0.6M pairs, and score 4.03 and 4.18 on OpenVE-Bench. HF Daily Papers' note
The paper proposes a lightweight two-stage framework built on Generative Refinement Networks. Its first stage turns source video codes into evidence signals and trains an empty instruction as a no-edit reconstruction path. Its second stage compares the edited state with that source-preserving state to revise unresolved target bits. The authors report that GRNEdit-2B and GRNEdit-8B use less than 3% conditioning parameters, train on 0.6M pairs, and score 4.03 and 4.18 on OpenVE-Bench. HF Daily Papers' note
score 4