GPT Image 2.5 vs Nano Banana 2.1: 502 synthetic rust masks reviewed, 79.7% retained
Gloomy_Recognition_4 · reddit · 2026-10-21
The author benchmarked GPT Image 2.5 Sunburst against Nano Banana 2.1 on identical prompts for generating synthetic rust-segmentation training data.
- A pilot with 8 prompts showed both models producing missing rust regions, shifted mask boundaries, and false positives
- Scaled to 502 image-mask pairs; after visual screening, 400 were kept and 102 excluded — a 79.7% retention rate
- Two engineering notes document the comparison and dataset review with interactive mask overlays
- Next step: training RF-DETR and evaluating on real images
More from Multimodal
- Fibo2 old-photo restoration and colorization tested on a historic fire scene — Striking-Long-2960 · 2026-10-21
- Full Seedance 2.5 prompt shared for generating ultra-realistic handheld smartphone video — techhalla · 2026-10-21
- AI-generated 'Doner Kebab Taxi' video goes viral, poster holds the music rights — average1804 · 2026-10-21
- Fibo 2.0 ComfyUI workflows: Edit Turbo is fast and solid at image editing — nomadoor · 2026-10-21
- MOSS t2i anti-blur tip: write 'photography', not 'photograph' — Bulky_Chemistry_196 · 2026-10-21
- PVBench: New Benchmark Scores Video Generation Across 46 Physical Plausibility Dimensions — Kaiwei Chen · 2026-10-21