VEFX-Reward-32B

VEFX-Reward-32B is the 32B-parameter variant of the VEFX-Reward family — a video editing quality reward model based on Qwen3-VL-32B-Instruct. It scores text-driven video edits on three dimensions on a 1–4 scale:

Dimension	What it measures
IF — Instructional Following	Does the edit accurately reflect the editing instruction?
RQ — Render Quality	Visual clarity, temporal consistency, physical plausibility
EE — Edit Exclusivity	Were only the intended regions modified, without side-effects?

• 📄 Paper: VEFX-Bench
• 💻 Code: Visko-Platform/VEFX-Bench
• 🤗 Dataset: xiangbog/VEFX-Bench
• 🤗 Smaller model: xiangbog/VEFX-Reward-4B
• 🏆 Leaderboard: vefx-leaderboard.com

Quick Start

git clone https://github.com/Visko-Platform/VEFX-Bench.git
cd VEFX-Bench
pip install -e .

from vefx_reward import VEFXReward

model = VEFXReward("viskoplatform/VEFX-Reward-32B", device="cuda")

scores = model.score(
    original_video="examples/sample_videos/object_removal_original.mp4",
    edited_video="examples/sample_videos/object_removal_edited.mp4",
    instruction="Remove the woman with the grey backpack walking on the right side of the frame.",
)
print(scores)
# {'IF': 3.69, 'RQ': 3.70, 'EE': 3.26, 'Overall': 10.65}

Hardware: ~65 GB VRAM (bfloat16). Tested on a single NVIDIA H100 80 GB.

Citation

@article{gao2026vefx,
  title={VEFX-Bench: A Holistic Benchmark for Generic Video Editing and Visual Effects},
  author={Gao, Xiangbo and Jiang, Sicong and Liu, Bangya and Chen, Xinghao and Yang, Minglai and Yang, Siyuan and Wu, Mingyang and Yu, Jiongze and Zheng, Qi and Wang, Haozhi and others},
  journal={arXiv preprint arXiv:2604.16272},
  year={2026}
}

License

Apache 2.0.