Dataset Pipeline Evaluation Gallery Human Eval Download

Download & Licensing

Apache 2.0

Source Code

Generation pipelines, dataset configurations, and evaluation tools.

View on GitHub
CC BY-NC 4.0

Dataset

ConSynth-X contains 34,199 synthetic records: 3,109 source scenes × 11 synthetic conditions (fog light/medium/heavy, rain light/heavy, snow light/heavy, night, rain_night, snow_night, small), admission-gated at DINO ≥ 0.85 similarity to source. Records include condition labels, bounding-box annotations, generation metadata, and quality scores in 11 Parquet files (one per condition).

Bring-Your-Own-Source (BYOS): source images are not embedded in the release. Users reconstruct the clear↔condition pairing from their own copy of the upstream datasets (CS10K via LouisChen15/ConstructionSite, SODA-D, SODA-ktsh) using the provided rebuild_consynthx.py script.

View on HuggingFace

Acknowledgments

Source Datasets

Construction Site 10K & SODA

Original construction datasets are drawn from Chen & Zou's Construction Site 10K and Duan et al.'s SODA (incl. SODA-ktsh extension). We thank the authors for releasing their datasets.

Computing Resources

Ohio Supercomputer Center

This work was supported in part by an allocation of computing time from the Ohio Supercomputer Center.

osc.edu