| { |
| "@context": { |
| "@language": "en", |
| "@vocab": "https://schema.org/", |
| "arrayShape": "cr:arrayShape", |
| "citeAs": "cr:citeAs", |
| "column": "cr:column", |
| "conformsTo": "dct:conformsTo", |
| "containedIn": "cr:containedIn", |
| "cr": "http://mlcommons.org/croissant/", |
| "rai": "http://mlcommons.org/croissant/RAI/", |
| "data": { |
| "@id": "cr:data", |
| "@type": "@json" |
| }, |
| "dataType": { |
| "@id": "cr:dataType", |
| "@type": "@vocab" |
| }, |
| "dct": "http://purl.org/dc/terms/", |
| "extract": "cr:extract", |
| "field": "cr:field", |
| "fileProperty": "cr:fileProperty", |
| "fileObject": "cr:fileObject", |
| "fileSet": "cr:fileSet", |
| "format": "cr:format", |
| "includes": "cr:includes", |
| "isArray": "cr:isArray", |
| "isLiveDataset": "cr:isLiveDataset", |
| "jsonPath": "cr:jsonPath", |
| "key": "cr:key", |
| "md5": "cr:md5", |
| "parentField": "cr:parentField", |
| "path": "cr:path", |
| "recordSet": "cr:recordSet", |
| "references": "cr:references", |
| "regex": "cr:regex", |
| "repeated": "cr:repeated", |
| "replace": "cr:replace", |
| "sc": "https://schema.org/", |
| "separator": "cr:separator", |
| "source": "cr:source", |
| "subField": "cr:subField", |
| "transform": "cr:transform" |
| }, |
| "@type": "sc:Dataset", |
| "name": "VTBench", |
| "alternateName": [ |
| "vtbenchmark/VTBench" |
| ], |
| "description": "VTBench v1.0 is a test-only benchmark for evaluating virtual try-on models under four challenges. It contains 2,896 formal test pairs: 600 Font/Texture Fidelity (FTF) pairs, 433 Complex Background Consistency (CBC) pairs, 1,443 Hand-Occlusion Consistency (HOC) pairs, and 420 Cross-Category Plausibility (CCP) pairs. The 50,000 images referenced in the collection process form the shared pre-filtering candidate pool for FTF, CBC, and HOC and are not the released benchmark size. The repository also provides a derived qualitative-results release for 100 CBC input pairs across 19 virtual try-on methods, comprising 1,900 generated images; these outputs are evaluation artifacts and do not increase the 2,896-pair benchmark count.", |
| "version": "1.0", |
| "conformsTo": [ |
| "http://mlcommons.org/croissant/1.1", |
| "http://mlcommons.org/croissant/RAI/1.0" |
| ], |
| "creator": { |
| "@type": "Organization", |
| "name": "VTBench", |
| "url": "https://huggingface.co/vtbenchmark" |
| }, |
| "keywords": [ |
| "virtual try-on", |
| "fashion", |
| "benchmark", |
| "image-to-image", |
| "responsible AI" |
| ], |
| "license": "https://choosealicense.com/licenses/other/", |
| "url": "https://huggingface.co/datasets/vtbenchmark/VTBench", |
| "rai:dataCollection": "FTF, CBC, and HOC share an initial pool of 50,000 candidate images discovered through search engines on publicly accessible web pages. Three reviewers manually screened the candidates to construct the three challenge subsets. CCP was constructed by selecting images from an existing public dataset and creating new cross-category test pairs. VTBench releases a test benchmark only and does not provide a training split.", |
| "rai:dataCollectionType": [ |
| "Images collected from publicly accessible sources", |
| "Re-pairing of images from an existing public dataset" |
| ], |
| "rai:dataCollectionRawData": "Released images are renamed and processed with face blurring. Images with confirmed redistribution permission are released in processed form. When copyright status is unresolved, only publicly accessible source-page links are provided. For copyright-restricted images, authorization is requested from the rights holder and any release follows the granted terms. Images for which authorization cannot be obtained are not distributed and are replaced with evaluation-equivalent images that have clear usage and redistribution permission.", |
| "rai:dataCollectionTimeframe": "March 2026 to May 2026", |
| "rai:dataCollectionMissingData": "There are no missing files in the formal benchmark pairs: every referenced person image and target-garment image is present. The current CBC source directory additionally contains 41 candidate garment files that are not referenced by the 433 formal CBC evaluation pairs; these files are not counted as missing data or formal test pairs.", |
| "rai:dataPreprocessingProtocol": "Data are organized into four subsets: benchmark_texture (FTF), benchmark_complex_bg (CBC), benchmark_hand (HOC), and benchmark_crossvton (CCP). FTF, CBC, and HOC store person images, garment images, and pair lists in image/, cloth/, and test_pairs.txt. CCP uses category directories and four pair files for dress-to-lower, dress-to-upper, lower-to-dress, and upper-to-dress evaluation. Stable release filenames are used, with filename mappings retained when renaming is required. For the formal benchmark person images, RetinaFace is run on every image and detected face regions are blurred; all 1,528 benchmark images for which RetinaFace reported no face were manually reviewed, and any missed faces were manually blurred. The derived CBC qualitative-results release contains 1,900 generated images for 100 aligned input pairs and 19 methods. RetinaFace processing detected at least one face in 1,547 generated images and blurred 1,661 detected face regions using a Gaussian blur kernel size of 75 with a 5% bounding-box margin. No face was detected in 353 generated images, which were left unchanged by the automated pipeline; no processing errors were reported.", |
| "rai:dataManipulationProtocol": "CBC candidates were selected for street, outdoor, or non-simple indoor backgrounds. An initial set of 500 images was reviewed, and 67 images that did not satisfy the complex-background criterion were removed, leaving 433. FTF contains 600 manually screened garment images with clear text, logos, or complex textures. CCP directly uses 420 test pairs; an early plan to expand 400 base samples to 1,000 pairs through combinatorial pairing was not executed. Each of the four CCP direction files contains 105 pairs. In the current file structure, upper and lower each contain 105 person images and 105 garment images, while dress contains 210 of each; the conceptual skirt category is grouped under the dress directory rather than stored in a separate skirt directory. HOC candidates were first processed with HaMeR for hand detection, then manually screened to retain 1,443 quality images in which a hand overlaps or occludes the clothing region. The earlier count of 1,433 was a typographical error. Release preparation also includes filename normalization and face blurring.", |
| "rai:dataImputationProtocol": "Not applicable. VTBench performs no statistical imputation of missing pixels, attributes, or labels. Unusable samples are excluded or explicitly identified rather than imputed.", |
| "rai:dataAnnotationProtocol": "VTBench does not release item-level classification labels, bounding boxes, masks, attributes, or natural-language annotations. The human process is sample screening and quality review: each candidate sample is screened and reviewed by three people according to the subset criteria for complex backgrounds, fine-grained textures, hand-clothing occlusion, or cross-category pairing. HaMeR and RetinaFace provide automated assistance for hand detection and privacy processing, respectively.", |
| "rai:dataAnnotationPlatform": "No dedicated third-party annotation platform was used. Human screening and review were performed with local scripts and manual inspection of image folders.", |
| "rai:dataAnnotationAnalysis": "Because the released data contain no item-level supervision annotations, annotation-agreement statistics were not computed. Candidate selection was jointly screened and reviewed by three people.", |
| "rai:annotationsPerItem": "Not applicable to released annotations because VTBench does not provide item-level supervision annotations. Each candidate sample was screened and reviewed by three people.", |
| "rai:annotatorDemographics": "Demographic information about the people who constructed and reviewed the dataset was not collected; therefore, the influence of reviewer composition on sample selection cannot be quantified.", |
| "rai:machineAnnotationTools": "HaMeR was used to detect hands when constructing HOC candidates. RetinaFace was used for face detection before release. Local Python scripts were used for file organization, filename normalization, pair-list generation, and face processing.", |
| "rai:personalSensitiveInformation": "The dataset contains visible people, clothing, poses, and scene backgrounds and may indirectly reveal apparent skin tone, body shape, age presentation, gender expression, tattoos, or background text. The formal benchmark contains no minors and provides no identity labels. For formal benchmark person images, detected faces are blurred and all automatic no-detection cases have been manually reviewed and corrected when necessary. For the 1,900 derived generated outputs, the automated pipeline blurred detected faces but reported no face in 353 images; no manual-review record is provided for those generated-output no-detection cases. Face blurring does not guarantee complete anonymity or remove every indirect re-identification cue. Neither benchmark images nor generated outputs may be used for identity recognition, person tracking, biometric analysis, or sensitive-attribute inference.", |
| "rai:dataBiases": "Search-engine discovery, source websites, and manual challenge-oriented screening introduce source, selection, and challenge-selection biases. Public web, fashion, and e-commerce imagery may overrepresent particular photographic styles and garment types. VTBench has no verified demographic labels or quantitative demographic-balance audit and must not be described as representative across age, skin tone, body shape, gender expression, or geography. Subset sizes, garment categories, and conversion directions are not balanced, and the challenge subsets do not represent natural deployment frequencies. RetinaFace and HaMeR may also have different detection performance across poses, occlusion levels, image quality, and population groups.", |
| "rai:dataLimitations": "VTBench contains static test images only and does not cover video try-on. It is limited in scale and is not demographically representative. Automated face detection and blurring cannot guarantee complete anonymity, even with manual review. All four subsets are intentionally challenge-oriented and do not represent real-world usage frequencies. CCP covers four conversion directions and groups the conceptual skirt category under the dress directory, so users should rely on pair metadata rather than directory names alone. The released generated outputs cover only 100 of the 433 CBC pairs and do not provide corresponding 19-method outputs for the remaining CBC pairs or for FTF, HOC, and CCP. The generated outputs are qualitative artifacts rather than an additional benchmark split, and 353 generated images had no face detected by the automated privacy pipeline. Source-page availability and copyright status can change over time. Repeated use of this public test set may cause benchmark overfitting, and VTBench scores must not be interpreted as guarantees of product safety, demographic fairness, or user satisfaction.", |
| "rai:dataUseCases": "Intended uses are non-commercial academic research, evaluation and comparison of virtual try-on models, analysis of texture, background, hand-occlusion, and cross-category failure modes, reproduction of benchmark experiments, and research on evaluation metrics. Out-of-scope uses include face or identity recognition, person tracking, biometric or sensitive-attribute inference, demographic profiling, commercial model training or product deployment, redistribution without verifying source permissions, and high-impact decisions about individuals.", |
| "rai:dataSocialImpact": "Potential benefits include more transparent comparison of virtual try-on systems, clearer identification of texture, background, hand, and cross-category failures, and development of more reliable evaluation methods. Potential harms include privacy or indirect re-identification risks, amplification of biases present in web and fashion imagery, use of generated outputs for non-consensual appearance editing, and misinterpretation of benchmark rankings as product-level safety or fairness guarantees.", |
| "rai:dataReleaseMaintenancePlan": "VTBench is released as version 1.0. The paper, dataset card, Croissant metadata, and pair manifest use 2,896 formal test pairs as the shared primary count: 600 FTF, 433 CBC, 1,443 HOC, and 420 CCP pairs. The separately distributed 1,900 generated CBC outputs are derived evaluation artifacts and do not alter that count. Future data corrections or replacements must update these artifacts together and increment the dataset version when the formal evaluation set changes; additions or corrections to derived model outputs must update their manifests, processing report, dataset card, and Croissant metadata together.", |
| "distribution": [ |
| { |
| "@type": "cr:FileObject", |
| "@id": "main-repo", |
| "name": "VTBench main repository", |
| "description": "The main Hugging Face dataset repository containing benchmark files and derived model outputs.", |
| "contentUrl": "https://huggingface.co/datasets/vtbenchmark/VTBench/tree/main", |
| "encodingFormat": "git+https" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "cbc-generated-results", |
| "name": "CBC generated results for 19 methods", |
| "description": "A total of 1,900 generated images for 100 aligned CBC input pairs across 19 virtual try-on methods.", |
| "containedIn": { |
| "@id": "main-repo" |
| }, |
| "encodingFormat": "image/png", |
| "includes": "VT_model_results_retinaface_100/results/*/*.png" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "cbc-results-pairs", |
| "name": "CBC result input pairs", |
| "description": "The 100 aligned CBC person-garment input pairs used for the released generated results.", |
| "contentUrl": "https://huggingface.co/datasets/vtbenchmark/VTBench/raw/main/VT_model_results_retinaface_100/pairs.tsv", |
| "encodingFormat": "text/tab-separated-values", |
| "separator": "\t" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "cbc-results-model-sources", |
| "name": "CBC result method manifest", |
| "description": "The 19 method names and their repository-relative result directories.", |
| "contentUrl": "https://huggingface.co/datasets/vtbenchmark/VTBench/raw/main/VT_model_results_retinaface_100/model_sources.tsv", |
| "encodingFormat": "text/tab-separated-values", |
| "separator": "\t" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "cbc-results-processing-report", |
| "name": "CBC result privacy-processing report", |
| "description": "Machine-readable processing summary for the 1,900 generated images.", |
| "contentUrl": "https://huggingface.co/datasets/vtbenchmark/VTBench/raw/main/VT_model_results_retinaface_100/processing_report.json", |
| "encodingFormat": "application/json" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "cbc-results-errors", |
| "name": "CBC result processing errors", |
| "description": "Processing error manifest. The current release contains a header and no error rows.", |
| "contentUrl": "https://huggingface.co/datasets/vtbenchmark/VTBench/raw/main/VT_model_results_retinaface_100/errors.tsv", |
| "encodingFormat": "text/tab-separated-values", |
| "separator": "\t" |
| }, |
| { |
| "@type": "cr:FileObject", |
| "@id": "parquet-repo", |
| "name": "Hugging Face converted Parquet repository", |
| "description": "The Hugging Face automatically converted Parquet branch.", |
| "contentUrl": "https://huggingface.co/datasets/vtbenchmark/VTBench/tree/refs%2Fconvert%2Fparquet", |
| "encodingFormat": "git+https" |
| }, |
| { |
| "@type": "cr:FileSet", |
| "@id": "parquet-files-for-config-default", |
| "containedIn": { |
| "@id": "parquet-repo" |
| }, |
| "encodingFormat": "application/x-parquet", |
| "includes": "default/*/*.parquet" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "dataType": "cr:Split", |
| "key": { |
| "@id": "default_splits/split_name" |
| }, |
| "@id": "default_splits", |
| "name": "default_splits", |
| "description": "Splits for the default config.", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "default_splits/split_name", |
| "dataType": "sc:Text" |
| } |
| ], |
| "data": [ |
| { |
| "default_splits/split_name": "test" |
| } |
| ] |
| }, |
| { |
| "@type": "cr:RecordSet", |
| "@id": "default", |
| "description": "vtbenchmark/VTBench - default subset", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "default/split", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "parquet-files-for-config-default" |
| }, |
| "extract": { |
| "fileProperty": "fullpath" |
| }, |
| "transform": { |
| "regex": "default/(?:partial-)?(test)/.+parquet$" |
| } |
| }, |
| "references": { |
| "field": { |
| "@id": "default_splits/split_name" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "default/text", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "parquet-files-for-config-default" |
| }, |
| "extract": { |
| "column": "text" |
| } |
| } |
| } |
| ] |
| }, |
| { |
| "@type": "cr:RecordSet", |
| "@id": "cbc_result_images", |
| "name": "CBC generated result images", |
| "description": "One record per generated image: 100 CBC input pairs evaluated by each of 19 methods, for 1,900 records.", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "cbc_result_images/model", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "cbc-generated-results" |
| }, |
| "extract": { |
| "fileProperty": "fullpath" |
| }, |
| "transform": { |
| "regex": "VT_model_results_retinaface_100/results/([^/]+)/[0-9]{4}\\.png$" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "cbc_result_images/sample_id", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "cbc-generated-results" |
| }, |
| "extract": { |
| "fileProperty": "fullpath" |
| }, |
| "transform": { |
| "regex": "VT_model_results_retinaface_100/results/[^/]+/([0-9]{4})\\.png$" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "cbc_result_images/result_image", |
| "dataType": "sc:ImageObject", |
| "source": { |
| "fileSet": { |
| "@id": "cbc-generated-results" |
| }, |
| "extract": { |
| "fileProperty": "fullpath" |
| } |
| } |
| } |
| ] |
| }, |
| { |
| "@type": "cr:RecordSet", |
| "@id": "cbc_result_pairs", |
| "name": "CBC result input pairs", |
| "description": "The 100 CBC person-garment pairs shared by all 19 released methods.", |
| "key": { |
| "@id": "cbc_result_pairs/sample_id" |
| }, |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "cbc_result_pairs/sample_id", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "cbc-results-pairs" |
| }, |
| "extract": { |
| "column": "sample_id" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "cbc_result_pairs/person_image", |
| "dataType": "sc:ImageObject", |
| "source": { |
| "fileObject": { |
| "@id": "cbc-results-pairs" |
| }, |
| "extract": { |
| "column": "person_image" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "cbc_result_pairs/cloth_image", |
| "dataType": "sc:ImageObject", |
| "source": { |
| "fileObject": { |
| "@id": "cbc-results-pairs" |
| }, |
| "extract": { |
| "column": "cloth_image" |
| } |
| } |
| } |
| ] |
| }, |
| { |
| "@type": "cr:RecordSet", |
| "@id": "cbc_result_models", |
| "name": "CBC result methods", |
| "description": "The 19 virtual try-on methods included in the generated-results release.", |
| "key": { |
| "@id": "cbc_result_models/model" |
| }, |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "cbc_result_models/model", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "cbc-results-model-sources" |
| }, |
| "extract": { |
| "column": "model" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "cbc_result_models/result_directory", |
| "dataType": "sc:Text", |
| "source": { |
| "fileObject": { |
| "@id": "cbc-results-model-sources" |
| }, |
| "extract": { |
| "column": "result_directory" |
| } |
| } |
| } |
| ] |
| } |
| ] |
| } |
|
|
