In order to assess the relative ability of general practitioners (GPs) to detect diabetic retinopathy (DR), especially sight-threatening diabetic retinopathy (STDR) by direct ophthalmoscopy or by examining, on a separate occasion, retinal images as 35 mm colour transparencies, a South and Mid Wales primary care-based study was performed in four general practices (six GPs). The participating GPs were provided with standardized training and equipment. Both methods were compared to the 'reference' grade of DR provided by the Diabetic Retinopathy Reading Centre (London), based on the same retinal images. Ophthalmoscopy and retinal photography (Canon CR4 45NM) with mydriasis were all practice based. The clinical assessments were based on a protocol developed for screening for DR in Europe. A total of 996 people with diabetes were identified, representing a prevalence of known diabetes of 2.1%. After exclusions on medical grounds, 897 patients were available for screening, of whom 605 (68%) were photographed. Based on the retinal images, the reference centre identified DR in 43% and STDR in 14.4%. In total, 597 valid comparisons between GPs and the reference centre were obtained; of these, 462 (77%) were high quality photographs which were used in subsequent analysis. The sensitivity for detecting any DR increased from 62.6% (95% CI 55.9-69.4) with ophthalmoscopy to 79.2% (95% CI 73.6-84.9) using retinal photographs, specificity remaining essentially unchanged at 75.0 (95% CI 69.5-80.5) and 73.5% (95% CI 68.0-79.1) with the positive predictive value (PPV) increasing from 67.2 (95% CI 60.4-74.0) to 71.0% (95% CI 65.0-77.0), respectively. The detection of STDR sensitivity increased from 65.7 (95% CI 54.4-77.1) with ophthalmoscopy alone to 87.3% (95% CI 79.4-95.2) based on retinal photographs with specificity falling from 93.8 (95% CI 91.4-96.3) to 84.8% (95% CI 81.2-88.5) and PPV from 65.7 (95% CI 54.4-77.1) to 51.2% (95% CI 42.1-60.3), respectively. We conclude that the use of standardized 35 mm colour transparency retinal photographs for screening by trained GPs in a primary care setting achieves an acceptable detection rate (>87%) for STDR, contrasting with ophthalmoscopy alone (66%), which was below the proposed UK standard of 80%.