A larger image count establishes a change in file count. Evidence of model improvement requires comparative results on independent real samples. Even when generated examples cover more lighting or occlusion combinations, check label correspondence, clusters of similar images and whether those conditions represent the target site.
How should the models be compared?
Keep the original baseline and record the data and training settings used after adding synthetic samples. Compare on the same real test set, annotation rules and metrics. Examine false detections, missed detections and error examples for key classes and operating conditions. State which conditions the test set covers. List any conditions with insufficient real test samples.
What can be delivered when real data is limited?
Initial delivery can include synthetic images, generation and screening records, applicable scope and a list of real samples to collect. If task annotations are required, check label fields and their correspondence to samples against the delivery specification. Evaluate any training benefits and field performance on independent real samples. Until then, mark both results as unverified. Record sources and versions for any added samples to preserve traceability.
Related services: Data, training and evaluation, project consultation.