GRNEdit: Efficient General Video Editing from a NewBinary-Evidence Perspective in Generative Refinement Networks

Feng Xie1,2,*Jiagao Hu2Fuhao Li2Zepeng Wang2Yuxuan Chen2Dahua Gao1,†Fei Wang2Daiguo Zhou2
Xidian University1 Xidian University
Xiaomi2 MiLM Plus, Xiaomi Inc.

* This work was completed during Feng Xie's internship at Xiaomi. We thank Xiaomi for its support. † Corresponding author.

GRNEdit is an efficient general video editing framework from a binary-evidence perspective, offering performance competitive with leading video editing models while adding under 3% extra parameters.

Explore the codeRead the paper
Open-source plan
  1. Inference codeAvailable
  2. PreprintAvailable
  3. Training codePlanned
  4. Model weightsPlanned

01 · Video results

One framework.
Many kinds of edits.

01

Global style

Watercolor animation

“Apply the watercolor animation style.”

A global appearance edit that keeps the source composition and motion legible.

02

Background

Library & fireplace

“Replace the background with a classic library and fireplace.”

A scene-level replacement that preserves foreground identity and temporal content.

03

Local removal

Remove a person

“Remove the man behind the computer.”

A localized removal that leaves the primary subject and surrounding scene intact.

04

Creative edit

Alchemical symbols

“Remove the flame and add glowing alchemical symbols.”

A compositional edit combining removal and addition in one instruction.

02 · Binary-evidence perspective

Resolve the edit,
progressively.

01

Locate with evidence

Early binary evidence separates editable scope from content that should remain stable.

02

Refine generatively

Coarse proposals evolve into target semantics, coherent structure, and fine details.

03

Adapt efficiently

The editing capability is introduced with under 3% additional parameters over the GRN backbone.

Inside the refinement process

Prediction and evidence, step by step.

Explore the trajectoriesChoose an editing type

7 cases · every sampled step updates together

Selected instruction“Transform the video into an Impressionist painting style.”

Full sampled trajectoryStep 1 → Step 50
Refinement state
Step1
Step10
Step20
Step30
Step40
Step50
01PredictionPixel space
02RMS Δ-MarginVAE latent space

Each column pairs the prediction with the RMS Δ-Margin from the same refinement step. All twelve videos share normalized time. Margin maps are enlarged proportionally from VAE latent space; brighter regions indicate a larger leave-one-branch-out change in bit margin.

03 · Qualitative gallery

Broad coverage,
consistent intent.

Qualitative GRNEdit results for style, background, and local edits
Across editing tasks

Style · Background · Local change

Qualitative GRNEdit results for removal, subtitle removal, and addition
Generalization & composition

Removal · Subtitle removal · Addition

05 · Citation

Cite GRNEdit

If you find GRNEdit useful, please cite our arXiv paper.

BibTeXarXiv · 2608.16328
@article{xie2026grnedit,
  title         = {GRNEdit: Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks},
  author        = {Xie, Feng and Hu, Jiagao and Li, Fuhao and Wang, Zepeng and Chen, Yuxuan and Gao, Dahua and Wang, Fei and Zhou, Daiguo},
  journal       = {arXiv preprint arXiv:2608.16328},
  year          = {2026},
  eprint        = {2608.16328},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

Acknowledgements

GRNEdit is built upon GRN: Generative Refinement Networks for Visual Synthesis and its official codebase. We sincerely thank the authors for releasing their code and for their exciting work.