ReasonEdit: Towards Interpretable Image Editing Evaluation via Reinforcement Learning

📦 Install

python -m pip install -r requirements.txt

📥 Model Weights

Download the model weights from the link and put them into:

weights/qwen35_9b
weights/qwen35_4b
weights/ReasonEdit
weights/RE-Reward

You can override the defaults at launch with --base_model_path and --lora_path.

The default registry points to:

base model: weights/qwen35_9b
LoRA weights: weights/ReasonEdit

⚡Inference

CLI Inference

Single sample, default CoT mode:

python -m cli.infer_cli \
  --source_image example/example_source.jpg \
  --edited_image example/example_edited.jpg \
  --instruction "$(cat example/example_instruction.txt)"

Score-only mode:

python -m cli.infer_cli \
  --mode score_only \
  --source_image example/example_source.jpg \
  --edited_image example/example_edited.jpg \
  --instruction "$(cat example/example_instruction.txt)"

Batch CSV:

python -m cli.infer_cli \
  --input_csv test_samples/official_noleak_test/test.csv \
  --output_csv outputs/reasonedit_predictions.csv

The batch CSV must contain source, edited, and instruction columns. Image paths in the CSV are resolved relative to this repository root when they are not absolute. The output CSV keeps only source, edited, instruction, v_score, e_score, p_score, and cot_text. When the input CSV also carries labels, the CLI prints three-dimensional PLCC, SRCC, and KRCC metrics in its JSON summary.

Multi-GPU Batch

python -m cli.infer_cli \
  --devices cuda:0,cuda:1,cuda:2,cuda:3 \
  --mode score_only \
  --input_csv test_samples/official_noleak_test/test.csv \
  --output_csv outputs/reasonedit_predictions.csv

Reward Model Package

The reward model release is split into code and weights:

RE-reward/
weights/RE-Reward/

Single-sample inference:

python RE-reward/reward_infer.py sample \
  --image_0 example/example_source.jpg \
  --image_1 example/example_edited.jpg \
  --instruction "$(cat example/example_instruction.txt)" \
  --critique "$(cat example/reward_example_critique.txt)"

CSV batch inference:

python RE-reward/reward_infer.py csv \
  --csv_path test_samples/rm_test_stratified_2000/test_2000_sft_like.csv \
  --output_csv outputs/rm_predictions.csv

The CSV can use either image_0/image_1/critique or source/edited/cot columns. If labels are present, the script reports three-dimensional PLCC, SRCC, and KRCC metrics. The default base model path is weights/qwen35_4b, and the default LoRA package path is weights/RE-Reward.

--pixel_budget defaults to 262144. calibration.json is optional; when it is absent, scores use identity calibration. The final reward is 0.3 * visual_quality + 0.4 * editing_alignment + 0.3 * content_preservation.