In the retail industry, acquiring high-quality image data is essential for accurate audits, stock analysis, and compliance verification. However, image collection is often hindered by inconsistencies, as sales professionals, driven by incentive-based targets, frequently capture low-quality photographs that fail to adhere to established guidelines. Such noisy inputs degrade downstream performance in automated systems. Traditional manual quality assurance methods are labour-intensive, inconsistent, and lack scalability, particularly given the vast diversity of India’s 12–13 million retail outlets. While machine learning-based image classifiers present a scalable solution, their performance is intrinsically tied to the availability of large-scale, meticulously labelled training datasets which are often difficult, time-consuming, and expensive to obtain. To address these challenges, we propose a novel framework for automated retail image quality assessment using Multimodal Large Language Models (MLLMs). Our approach begins with domain-adapted prompt engineering to generate supervision signals from multimodal models such as BLIP, LLaVA, and Qwen2-VL. We then employ an ensemble-based consensus mechanism to self-label images with improved reliability. The resulting high-confidence pseudo-labelled dataset with weak manual supervision is used to train and evaluate image classifiers, specifically ConvNeXt, ResNet-50, and Vision Transformers, using a carefully optimized training strategy. Experimental results demonstrate the effectiveness of this approach in evaluating multiple quality dimensions, including blurriness, composition, content clarity, and contextual relevance. By drastically reducing reliance on manual labelling, our pipeline enables scalable, accurate, and cost-effective retail image validation. This innovation ensures that only high-quality images contribute to business operations, leading to better decision-making and more efficient retail management.