VCReward-Bench contains 3,506 expert-annotated preference pairs for evaluating visual consistency in image editing. The dataset was created by GEditBench-v2 and was last updated in March 2026.
Use Cases
- Benchmarking reward models using expert preference pairs for image editing tasks.
- Training assessment models to predict human preferences on visual consistency.
- Evaluating the alignment of generative model outputs with human judgments on image edits.
Strengths
- 3,506 expert-annotated preference pairs provide a foundation for evaluation.
- Dataset is specifically designed for the focused task of visual consistency assessment in image editing.
Limitations
- Unknown row count and sample size for the underlying image data.
- Potential bias from the specific annotation process and expert pool used.
Provenance
- Source
- GEditBench-v2 via Hugging Face.
- Collection Method
- Expert-annotated preference pairs.
- Time Range
- null
- Freshness
- Last updated March 2026.
- Geography
- null