Instruction-based video editing aims to unify diverse editing operations into a single intuitive interface, but existing methods often rely on resource-intensive conditioning, such as heavyweight branches or costly source concatenation. Researchers from Hugging Face propose GRNEdit, a lightweight two-stage framework that models editing intent through binary semantic decisions and source evidence, achieving strong results with minimal parameters.
The framework is inspired by Generative Refinement Networks (GRN), which encode visual semantics through combinations of bits. GRNEdit recasts editing semantics as local retain-or-flip decisions over individual bits, with source information modeled as coordinate-wise evidence supporting the observed binary states. The GRN backbone resolves the global composition into coherent generative semantics.
In Stage I, a compact encoder translates discrete source codes into continuous evidence signals that GRN assimilates during binary refinement. Inspired by null-prompt training for classifier-free guidance, the null condition is assigned an editing-specific meaning: an empty instruction denotes no edit and is supervised through source reconstruction. This identity pathway strengthens evidence utilization and content preservation, producing a source-preserving state in the same representation space as the edited state.
Stage II directly compares each edited state with its source-preserving counterpart, using their discrepancy to revise unresolved target-bit decisions. Trained on only 0.6M pairs with less than 3% conditioning parameters, GRNEdit-2B and GRNEdit-8B achieve scores of 4.03 and 4.18 on OpenVE-Bench. The 2B model outperforms multiple 14B open-source editors, while the 8B model performs on par with leading open-source editors.