GRNEdit: Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks
Feng Xie, Jiagao Hu, Fuhao Li, Zepeng Wang, Yuxuan Chen, Dahua Gao, Fei Wang, Daiguo Zhou
cs.CV
2026-08-17
GRNEdit recasts video editing as per-bit retain-or-flip decisions on a binary backbone; with 0.6M pairs and under 3% conditioning params, its 2B model outscores several 14B editors.
Instruction-based video editing aims to unify style transfer, background swaps, object removal and local edits under one text command. The bottleneck in current editors is how source-video information reaches the generative backbone: ControlNet- and VACE-style designs copy whole backbone blocks as conditioning branches (billions of parameters), while source-latent concatenation pushes the cost into inference. The conditioning channel scales with the backbone. A Xiaomi MiLM Plus team, working with Xidian University, asked the opposite question: leave content synthesis to the backbone, and let a tiny branch decide only where to edit and where to keep still.
They swap the substrate. GRN (Generative Refinement Network) is neither diffusion nor autoregression: HBQ, a hierarchical binary quantizer, compresses video latents into stacks of 0/1 codes, and GRN refines the whole bit map with global random refinement at every step. The key observation is that quantized source and target share aligned bit coordinates, so an edit reduces to a per-coordinate retain-or-flip decision. The source is no longer regressed as continuous features; it acts as evidence biasing binary decisions, supervised only through the final bit likelihood, while the backbone handles content synthesis.
Training runs in two stages. Stage I injects evidence: small per-site projectors (about 3M parameters for the 2B model, 7M for 8B) turn source bits into instruction-modulated signals, added only at the first and last Transformer chunks; injecting into all seven chunks scores worse (3.95 versus 4.03). An identity-aligned null condition borrows the condition dropout of classifier-free guidance but defines the empty instruction as no edit, supervised by source reconstruction; mixing in 10% identity samples lifts the overall score from 3.97 to 4.03 and yields a source-preserving reference state in the same representation space. Stage II freezes Stage I and trains a 30M-parameter Bit-Margin Router that compares edited and reference states per bit and revises undecided positions with signed coefficients, raising PSNR on non-edited regions from 22.0 to 23.8 dB. Total conditioning is 37M (2B) and 45M (8B) parameters, under 3% of the backbone, trained on a task-balanced 0.6M pairs from OpenVE-3M for 60K steps on 16 H200 GPUs.
OpenVE-Bench (eight editing tasks, VLM-scored), selected Overall results: Runway Aleph commercial reference 4.49; GRNEdit-8B 4.18 with 45M conditioning parameters at 84 s per video, tying UniVideo (14B, over 7B conditioning, 893 s) and edging Kiwi-Edit (4.17); GRNEdit-2B 4.03 with 37M at 39 s, ahead of DITTO 14B (3.44, 611 s) and VACE 14B (3.01, 545 s). By category, the 8B model posts the best open-source scores on background change (4.12) and local removal (4.73), plus 4.86 on subject replacement; its weakest category is object addition (3.86, versus UniVideo's 4.41). On the out-of-distribution ReCo-Bench it ranks first on all four Remove components (aggregate 7.99) and scores 9.08 on Style, second to ReCo's 9.17 but with the best motion-naturalness and stability components. The mechanism has direct support: retained bits receive 6.2x stronger source-aligned responses than flipped bits, and the same evidence conditioning beats VACE-style branches on both GRN and Infinity backbones with a roughly 100x smaller branch, while the ordering reverses on the continuous-latent Wan backbone. The trick is tied to binary representations, not a renamed conventional branch.
It is a concrete counterexample to the assumption that conditioning branches must scale with the backbone: reduce control to per-bit binary choices and about 2% of backbone parameters suffice. The 39-second and 84-second inference times sit in the interactive range for editing products. For binary generative backbones like GRN and Infinity, which are still fighting for mainstream adoption, this is the first credible answer on editability.
The authors concede two: object addition, where the source offers little evidence, stays weak; and the method is tied to binary backbones. OpenVE-Bench relies on VLM scoring, which the authors admit is insensitive to pixel drift in non-edited regions, hence the PSNR side-evidence; whether a 4.18-versus-4.17 gap is statistically meaningful goes untested. Training and evaluation data (OpenVE-3M, OpenVE-Bench, ReCo-Bench) all come from one synthetic-instruction ecosystem, with no validation on real footage. The word evidence is doing marketing work; mechanically this is source conditioning recast as bias on a binary decision axis, though the visualizations do support its directionality.