🎬 One Editor, Many Edits: A Unified Training-Free Framework for Diverse Video Editing

Adheesh Sunil Juvekar · Onkar Kishor Susladkar · Kiet A. Nguyen · Muntasir Wahed · Nabeel Bashir · Xiaona Zhou · Tianjiao Yu · Vedant Shah · Ismini Lourentzou

University of Illinois Urbana-Champaign

✨ Overview

• 🎯 We introduce EditVid, a training-free framework that supports both instruction-guided and subject-guided video editing, leveraging MM-DiT-based image editors as strong priors across diverse editing settings.

• 🧠 We identify temporal context as a key factor in MM-DiT representation reuse and develop a RoPE-aware local–global design that uses adjacent-frame key–value memory for short-range coherence and confidence- and cycle-consistent token transfer for long-range preservation.

• 📊 Comprehensive quantitative, human, robustness, and cross-backbone evaluations demonstrate EditVid's strong performance in temporally consistent video editing without video-specific training.

Official implementation of EditVid, a training-free framework for diverse video editing, supporting various instruction-guided and reference-guided edits in one framework, including style transfer, attribute modification, object insertion, part-level editing, and subject replacement.

🛠️ Installation

The code is tested with Python 3.11, PyTorch 2.7.0, CUDA 12.8.

Access to the gated black-forest-labs/FLUX.2-klein-9B model is required.

git clone https://github.com/PLAN-Lab/editvid.git
cd editvid
conda create -n editvid python=3.11 pip -y
conda activate editvid
conda install ffmpeg -y
python -m pip install -r requirements.txt

Next, choose one of the following Diffusers installation methods. Both use the exact tested Diffusers commit 769a1f3a120dd2a483b0e99bd6b13464b8ee62fb and leave its source code unmodified.

Option A: installation helper

python scripts/install_diffusers.py

The helper fetches the pinned commit, confirms that its checkout is clean, and installs Diffusers and the local editvid package.

Option B: manual installation

mkdir -p .vendor
git clone https://github.com/huggingface/diffusers.git .vendor/diffusers
git -C .vendor/diffusers checkout --detach 769a1f3a120dd2a483b0e99bd6b13464b8ee62fb
python -m pip install --no-deps -e .vendor/diffusers
python -m pip install --no-deps -e .

Authenticate for the gated model and inspect the installed versions:

hf auth login
python --version
python -c "import torch; print('torch:', torch.__version__, 'CUDA:', torch.version.cuda, 'GPU available:', torch.cuda.is_available())"
python -c "import diffusers; print('diffusers:', diffusers.__version__, diffusers.__file__)"
python -c "from editvid import EditVidPipeline, EditVidTransformer2DModel; print(EditVidPipeline.__module__, EditVidTransformer2DModel.__module__)"
git -C .vendor/diffusers rev-parse HEAD
git -C .vendor/diffusers status --short

🎞️ Run the examples

The bundled manifest contains six full-length source videos from the project-page showcase. run_editvid.py passes each edit_prompt to the model exactly as written in the manifest.

Test the setup with one chunk

Use one source video and cap it at 16 frames with --max-frames, --chunk-size, this performs exactly one generation chunk (if memory allows):

CUDA_VISIBLE_DEVICES=0 python run_editvid.py \
  --manifest-json examples/manifest.json \
  --source-videos-dir examples/source_videos \
  --rows 0 \
  --max-frames 16 \
  --chunk-size 16 \
  --fallback-chunk-sizes 8,4 \
  --output-root outputs \
  --run-name setup-test

The equivalent convenience command is:

CUDA_VISIBLE_DEVICES=0 bash scripts/run_example.sh

Generate the full videos

Omit --max-frames to process every frame of every bundled source video. Use this version to visualize complete results or produce outputs for evaluation:

CUDA_VISIBLE_DEVICES=0 python run_editvid.py \
  --manifest-json examples/manifest.json \
  --source-videos-dir examples/source_videos \
  --chunk-size 32 \
  --fallback-chunk-sizes 16,8,4 \
  --output-root outputs \
  --run-name examples-full

Results are written under outputs/<run-name>/. For custom experiments, edit or replace examples/manifest.json; run python run_editvid.py --help for the remaining options.

📝 Citation

If you find EditVid useful for your research, please cite our paper:

@article{juvekar2026editvid,
  title   = {One Editor, Many Edits: A Unified Training-Free Framework for Diverse Video Editing},
  author  = {Adheesh Sunil Juvekar and Onkar Kishor Susladkar and Kiet A. Nguyen and Muntasir Wahed and Nabeel Bashir and Xiaona Zhou and Tianjiao Yu and Vedant Shah and Ismini Lourentzou},
  journal = {arXiv preprint arXiv:2609.04190},
  year    = {2026},
}