GRNEdit

Efficient General Video Editing from a New Binary-Evidence Perspective in Generative Refinement Networks

GitHub & inference code · Project page & video results · Paper

GRNEdit edits a source video using a text instruction. It supports global stylization, background replacement, local changes, object removal and addition, creative edits, and subtitle removal.

🎉 Meet GRNEdit-2B Stage-LoRA! LoRA fine-tuning improves object addition and raises the OpenVE-Bench overall score to 4.21. With a 2B backbone and only ~1% extra conditioning parameters, it scores above the 14B open-source editors in our comparison. The parameter overhead counts the added conditioning branches, not all parameters involved in fine-tuning.

Released checkpoints

All weights below are public and can be downloaded without login or approval.

Model	Download	Size	Contents	Inference code
GRNEdit-2B	Checkpoint	16.56 GiB	Model + EMA	Stage I
GRNEdit-2B Stage-LoRA	Checkpoint	8.87 GiB	Full backbone + LoRA + source branches; no EMA	Stage-LoRA
GRNEdit-8B	Seven parts + checksum	65.82 GiB total	Model + EMA, split into byte parts	Not released yet

The public inference code currently supports 2B only. Publishing the 8B weights does not enable 8B inference in the current code.

These are FP32 PyTorch checkpoints with inference configuration and model state, not full optimizer/RNG training-resume checkpoints. Stage-LoRA is a full model checkpoint, not an adapter-only download: it does not require a separate GRN base checkpoint. The VAE and UMT5 resources are still required.

File sizes, included states, and the 8B split layout are recorded in release_manifest.json.

Quick start: Stage-LoRA (2B)

Clone the GitHub repository and follow its environment setup, then run:

pip install -r requirements-stage-lora.txt
pip install huggingface_hub
hf download debugg/GRNEdit GRNEdit-2B-Stage-LoRA.pth --local-dir checkpoints

Prepare the matching HBQ_image_video_tokenizer_64dim_M4_20260626.ckpt VAE, UMT5 encoder/tokenizer, and source-video JSONL metadata as described in the README. From the repository root:

CHECKPOINT_PATH=checkpoints/GRNEdit-2B-Stage-LoRA.pth \
WEIGHTS_DIR=weights \
VAE_PATH=weights/HBQ_image_video_tokenizer_64dim_M4_20260626.ckpt \
T5_PATH=weights/umt5-xxl \
EDIT_OFFICIAL_META_ROOT=data/metadata \
OUTPUT_DIR=outputs/stage_lora \
GPUS=1 \
GUIDANCE_MODE=none \
MAX_INFER_STEPS=50 \
bash scripts/infer_stage_lora.sh

Source-branch positions/count, LoRA rank, alpha, and text-conditioning settings are restored from checkpoint metadata. Inference requires the source video and instruction, not a target video. The default is conditional-only inference, without a classifier-free-guidance branch.

For the original 2B model, download GRNEdit-2B.pth and follow GRNEdit inference. Use the VAE and runtime configuration matching the selected checkpoint; the two variants' settings should not be interchanged.

Split weights

GRNEdit-8B/ contains seven sequential byte parts and SHA256SUMS. They are parts of one .pth file, not independently loadable model shards. Download all seven parts and the checksum file together.

The public 2B launchers also support this packaging: set CHECKPOINT_PATH to a split directory and the loader will assemble and verify the checkpoint once, then reuse it from ~/.cache/grnedit/checkpoints. Set GRNEDIT_CHECKPOINT_CACHE to choose another cache disk. Assembly requires additional disk space equal to the complete checkpoint size. This loader feature does not enable 8B inference.

OpenVE-Bench results

Model	Overall	Global style	Local change	Background	Removal	Addition	Creative	Camera	Subtitles
GRNEdit-2B	4.03	4.36	3.93	3.80	4.56	3.84	—	—	4.72
GRNEdit-8B	4.18	4.37	4.09	4.12	4.73	3.86	—	—	4.86
GRNEdit-2B Stage-LoRA	4.21	4.19	4.33	4.11	4.55	4.40	3.88	3.45	4.37

Higher is better. Original 2B/8B scores are from the paper's main table; Stage-LoRA scores are our subsequent evaluation. Dashes indicate categories not listed separately in that table. See the full baseline comparison and visual results for context.

Usage and limitations

Intended for research and creative, instruction-based video editing. Edit quality and preservation of unedited content vary with the source video and instruction; temporal artifacts and incomplete edits can occur. Use videos you are authorized to edit, and avoid misleading or non-consensual uses.

The release manifest records inference_smoke_tested: false; it does not certify an end-to-end GPU run of these exported files. The code has small-model CPU equivalence tests, but those do not replace full-weight CUDA/FA4 validation. Training code is a later release milestone.

Authors and acknowledgements

Feng Xie, Jiagao Hu, Fuhao Li, Zepeng Wang, Yuxuan Chen, Dahua Gao, Fei Wang, and Daiguo Zhou — Xidian University and MiLM Plus, Xiaomi Inc.

This work was completed during Feng Xie's internship at Xiaomi. We thank Xiaomi for its support. GRNEdit builds on GRN and its paper; we thank the authors for their code and exciting work.

Contact: fengx@stu.xidian.edu.cn.

Citation

@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}
}

License

This weight repository is marked MIT. See the project license. External GRN, VAE, and UMT5 resources remain subject to their respective licenses.