From Professional Photography Workflow to AI Training Data: A Practical Framework for Structured Image Description, Quality Control, and License-Compliant Dataset Curation
The quality of training data remains a critical bottleneck for vision-language models(VLMs). Widely used datasets rely on generic, web-scraped captions that lack technicaldepth, information about the illumination, and explicit quality control. This paperproposes a reproducible framework that translates professional commercial photographyworkflows into structured annotations for AI training data. The framework has fourcomponents: (i) a four-layer annotation schema (factual description, technical analysis,contextual purpose, usage decision); (ii) an acquisition-metadata block that records thespectral quality of the light source (CRI/Ra, R9, TM-30, TLCI, CCT, Duv); (iii) a quantitative,binary quality-control protocol with proposed initial thresholds; and (iv) a licensing andprivacy model compatible with CC BY 4.0, GDPR and the EU AI Act. We hypothesise thatillumination sources with limited spectral fidelity may produce residual colour shifts afterwhite-balance correction, particularly in mixed-light and shadow regions; the direction,magnitude and practical significance of these shifts are to be tested empirically. Wetherefore propose recording illumination quality, colour-fidelity metrics and processingprovenance as explicit annotation variables, rather than allowing such conditions toremain undocumented in training data. We state a testable hypothesis and anexperimental protocol, and we discuss limitations. The paper is a framework and protocolproposal; empirical validation is described as future work. The proposed numerical thresholds are initial, context-dependent working criteria rather than validatedstandards; they must be calibrated on a pilot dataset before being used as acceptance orrejection rules.Keywords: vision-language models; image annotation; data quality control; colourrendering index; LED lighting; photography workflow; dataset curation; CreativeCommons; data provenance
Authors
- Dejan Nikolić (ORCID: https://orcid.org/0009-0007-8233-9647)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-12
- DOI
- https://doi.org/10.5281/zenodo.24161364
- Primary Topic
- Multimodal Machine Learning Applications
- Type
- preprint