Research questionHow can synthetic image augmentation reliably improve computer vision when labeled real data are scarce?Generated images can be abundant but may contain quality defects and fail to match the real data distribution, so adding them does not consistently improve a model. The challenge is making synthetic examples useful without assuming synthetic-only training can match training on real data. Latest papersRecent research connected to this question, newest first.A Real-Calibrated Synthetic-First Data EngineEvidence centers on human pose estimation, with supplementary segmentation diagnostics. The source describes a modular CLI pipeline for controllable diffusion generation, filtering, selection, and validation, with optional uncertainty-driven selection and human verification; results support augmentation alongside real data, while synthetic-only performance remains below real-only performance.research paper · Sep 3, 2026Structured-Prior-Guided Diffusion Inpainting with Physical Consistency for Traffic Sign AugmentationThe source describes structured semantic, appearance, and geometric conditioning for diffusion inpainting, together with colour and edge consistency losses. Evidence comes from self-supervised reconstruction on in-house imagery, zero-shot evaluation on TT100K-2021, ablations, fidelity and controllability metrics, OCR accuracy, inference time, and downstream rare-class detection performance.research paper · Sep 2, 2026