ConSynth-X was developed for evaluating the robustness of construction computer-vision and vision-language models under adverse environmental and visual conditions.

The resource defines 34,199 synthetic image records derived from 3,109 retained source scenes across 11 condition-specific subsets spanning precipitation, fog, nighttime illumination, adverse weather at night, and small-object/long-distance views.

Public Sample

A representative sample of ConSynth-X is publicly available through the project gallery:

View ConSynth-X Sample Gallery →

ConSynth-X is derived from third-party construction-image datasets with source-specific licensing and access terms. The release includes the generated synthetic images, source identifiers, source-derived annotations, provenance information, generation metadata, and quality-control results. Original source-image pixels are not redistributed and should be obtained from the respective upstream dataset providers.

Dataset preview

Figure 1. Representative paired examples from ConSynth-X. Five source construction scenes are shown with their corresponding synthetic variants across the 11 condition-specific subsets: light, moderate, and heavy fog; light and heavy rain; light and heavy snow; nighttime; rain at night; snow at night; and small-object/long-distance view. Bounding boxes are shown for examples with object-detection annotations.

What is shared

The public ConSynth-X release provides the information needed to identify, analyze, and reconstruct the dataset, including:

retained source-scene identifiers and source-dataset provenance;
mapping between the 3,109 retained source scenes and the 11 ConSynth-X conditions;
condition and condition-group labels;
generation methods and configuration metadata;
quality-control and validation measurements, including DINOv3-based source--synthetic similarity where available;
reconstruction/provenance manifests and integrity information; and
links to the generation and validation software.

Upstream image pixels and restricted upstream annotations are not relicensed by ConSynth-X. Users should obtain those materials from the original data providers under the applicable terms.

Dataset composition

Source dataset	Retained source scenes	Synthetic image records (× 11 conditions)
ConstructionSite 10k	1,586	17,446
SODA	608	6,688
SODA-KTSH	915	10,065
Total	3,109	34,199

Each retained source scene is represented across all 11 ConSynth-X conditions, enabling paired condition-specific robustness analysis after local reconstruction.

Synthetic conditions

Condition group	Conditions	Generation approach
Precipitation	`rain_light`, `rain_heavy`, `snow_light`, `snow_heavy`	InstructPix2Pix-based weather editing with physics-based precipitation rendering
Fog	`fog_light`, `fog_medium`, `fog_heavy`	Depth-guided Koschmieder atmospheric scattering
Low illumination	`night`	CycleGAN-Turbo day-to-night translation
Adverse weather at night	`rain_night`, `snow_night`	Weather editing → day-to-night translation → precipitation particle rendering
Small-object / long-distance view	`small`	FLUX.1-Fill-dev outpainting / canvas expansion

Quality control

A joint DINOv3 cosine-similarity threshold of 0.85 was used as the source-scene retention criterion for eight quality-controlled subsets:

fog_light
fog_medium
fog_heavy
rain_light
snow_light
snow_heavy
night
small

A source scene was retained only when all eight corresponding synthetic variants satisfied the threshold. The three additional conditions — rain_heavy, rain_night, and snow_night — were retained as part of the predefined condition set and evaluated separately rather than being included in the joint admission gate.

The released metadata include per-image quality information to support condition-specific filtering and sensitivity analysis.

Upstream datasets and access

ConSynth-X is a secondary resource built from the following third-party datasets. Users must obtain the upstream data directly from the corresponding providers.

ConSynth-X subset	Upstream dataset	Access and licensing
`cs10k`	ConstructionSite 10k	Obtain directly from the upstream provider; currently identified by the provider as CC BY-NC 4.0 and subject to the provider's access conditions
`soda`	SODA	Obtain directly from the original SODA provider; subject to the original SODA terms and permissions
`soda_ktsh`	SODA-KTSH	Obtain directly from the original provider; subject to the applicable upstream terms and permissions

ConSynth-X does not grant additional rights to the upstream datasets and does not replace their access procedures.

Reconstruction workflow

The ConSynth-X generation and validation code is available at:

https://github.com/ruoxinx/ConSynth-X

A typical reconstruction workflow is:

Obtain the required upstream dataset(s) from their original providers.
Match the upstream records to the ConSynth-X retained-source manifest using the released source identifiers/provenance information.
Install the generation environment and required third-party model weights following the repository documentation.
Run the corresponding condition-generation pipelines using the released generation configuration and metadata.
Use the released quality-control metadata and manifests to verify the reconstructed records and source--synthetic correspondence.

Users who have access to only one upstream dataset may reconstruct only the corresponding ConSynth-X subset.

Record organization

The ConSynth-X metadata are organized by source dataset and condition. Core fields include, where applicable:

source_id — identifier linking the record to its upstream source scene;
source_dataset — upstream dataset provenance;
condition — one of the 11 ConSynth-X condition labels;
condition_labels — condition-group / multi-label descriptors;
pipeline — generation-method metadata;
quality_scores — source--synthetic quality and validation measurements; and
reconstruction/provenance information used to associate locally generated images with the corresponding source scene.

Upstream image pixels and restricted source annotations are obtained from the original providers and are not relicensed by this repository.

License and provenance

ConSynth-X contains original metadata and software developed by the authors together with references to third-party source datasets. Licensing is therefore component-specific.

ConSynth-X metadata and documentation

Original materials created specifically for ConSynth-X, including condition labels, generation metadata, quality-control measurements, provenance mappings, reconstruction manifests, and dataset documentation, are released under the Creative Commons Attribution–NonCommercial 4.0 International License (CC BY-NC 4.0).

Software

Generation, validation, and reconstruction software is released separately under the Apache License 2.0 in the public code repository:

https://github.com/ruoxinx/ConSynth-X-Dataset

Upstream datasets

ConstructionSite 10k, SODA, and SODA-KTSH are third-party resources. Their images, annotations, and other source materials remain governed by the licenses, permissions, and access conditions specified by their respective providers.

Reconstructed synthetic images

Synthetic images generated locally from third-party source datasets may remain subject to applicable upstream dataset terms and third-party model licenses. The ConSynth-X CC BY-NC 4.0 license does not override or replace those upstream terms.

Intended use

ConSynth-X is intended to support research on robustness and generalization of construction computer-vision and vision-language systems under varying environmental and observational conditions. Potential applications include:

condition-specific robustness evaluation;
object-detection benchmarking;
image-captioning and visual-grounding evaluation;
visual question answering;
controlled source--synthetic comparison; and
analysis of model sensitivity to weather, illumination, visibility, and apparent object scale.

Synthetic conditions do not reproduce the full variability of field-captured adverse environments. ConSynth-X is intended to complement, rather than replace, evaluation using real adverse-condition imagery.

Citation

If you use ConSynth-X, please cite the ConSynth-X dataset record and the corresponding upstream dataset(s) used in your work.

@dataset{duong2026consynthx,
  title     = {ConSynth-X: A paired synthetic construction-site image dataset for robust computer vision under adverse conditions},
  author    = {Duong, Viet Huy and Xiong, Ruoxin and Al Forhad, Md Abdullah and Shi, Weishi},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/9597},
  url       = {https://huggingface.co/datasets/openconstruction/ConSynth-X}
}

Acknowledgements

The authors acknowledge the creators of ConstructionSite 10k, SODA, and SODA-KTSH for making the upstream resources available, and the Ohio Supercomputer Center for computational resources used in dataset generation and validation.

Downloads last month: 59

Total file size:

12.6 GB