EditScore-code

Mirror of the EditScore source code: the editscore Python package, evaluation scripts and training examples for image-edit reward models.

What it is for

The repo copies the EditScore code repo. It holds the editscore package, which loads an EditScore reward model and scores an image edit against its instruction. It also holds evaluation scripts that benchmark reward models on EditReward-Bench, and examples for training EditScore reward models, best-of-N selection and reinforcement fine-tuning of image editing models. The model weights live in separate repos.

Provenance

• Upstream source: VectorSpaceLab/EditScore. Paper: arXiv 2509.23909. Project page: EditScore.
• Related upstream data: EditReward-Bench, EditScore-Reward-Data, EditScore-RL-Data. Models: the EditScore collection.

Use

hf download chibifire/EditScore-code --local-dir EditScore
cd EditScore
pip install -e .
bash evaluate.sh

Install PyTorch for your CUDA version first. vllm is optional and speeds up inference. See the upstream README for full setup, usage and the training and RL guides under examples/.

Licence

Apache 2.0, the licence of the EditScore code repo. Terms: LICENSE.