UniEditBench_models / README.md
nielsr's picture
nielsr HF Staff
Add model card and metadata for UniEditBench distilled evaluator
8b2af43 verified
|
Raw
History Blame
2.51 kB
metadata
license: apache-2.0
library_name: peft
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3-VL-4B-Instruct
tags:
  - image-editing
  - video-editing
  - reward-model
  - evaluation

UniEditBench Distilled Evaluator (4B LoRA)

This repository contains the distilled 4B MLLM judge (LoRA adapter) introduced in the paper UniEditBench: A Unified and Cost-Effective Benchmark for Image and Video Editing via Distilled MLLMs.

UniEditBench is a unified benchmark for image and video editing that standardizes evaluation across different paradigms and modalities. To enable scalable and cost-effective evaluation, the authors distilled high-capacity MLLM judges (such as Qwen3-VL-235B) into lightweight 4B and 8B evaluators. These reward models provide multi-dimensional scoring across structural fidelity, text alignment, background consistency, naturalness, and (for videos) temporal-spatial consistency.

Usage

Deployment

The reward model can be deployed using the swift framework. Below is an example command to deploy the 4B image-based evaluator:

CUDA_VISIBLE_DEVICES=0 swift deploy \
    --model Qwen/Qwen3-VL-4B-Instruct \
    --adapters /path/to/sft_image_lora_4b \
    --device_map balanced \
    --vllm_tensor_parallel_size 1 \
    --attn_impl eager \
    --infer_backend vllm \
    --port 8003 \
    --vllm_max_lora_rank 32 \
    --max_new_tokens 2048 \
    --served_model_name Qwen3-VL-4B-SFT-Image

For video evaluation or 8B model scripts, please refer to the scripts/deploy_qwen3vl.sh file in the official GitHub repository.

Citation

If you find UniEditBench helpful, please cite the following work:

@misc{jiang2026unieditbenchunifiedcosteffectivebenchmark,
      title={UniEditBench: A Unified and Cost-Effective Benchmark for Image and Video Editing via Distilled MLLMs}, 
      author={Lifan Jiang and Tianrun Wu and Yuhang Pei and Chenyang Wang and Boxi Wu and Deng Cai},
      year={2026},
      eprint={2604.15871},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.15871}, 
}