GPT Image 2.5 vs Nano Banana 2.1: generating and reviewing rust segmentation data

Overview

Same prompt. Different rust-mask errors. I compared GPT Image 2.5 Sunburst and Nano Banana 2.1 on eight identical prompts to explore synthetic training data for rust segmentation. Both generated promising samples and masks with missing rust, shifted boundaries or false positives. I then expanded the work to 502 image mask pairs, including the pilot. After visual screening, 400 remained and 102 were excluded: a 79.7% retention rate. That is an encouraging starting set for training, with room to e

SpinGraph analysis pending — check back after processing.