Distribution distance to real-world weather datasets. Night augmentation achieves strongest improvement.
SigLIP2-based classifier accuracy on IP2P diffusion outputs. Rain reaches 63.5%.
GLCM + LBP + DCT + Haralick via Dempster-Shafer belief fusion. IP2P diffusion rain and snow preserve micro-texture.
Two-judge cross-tier VLM jury (InternVL2.5-8B + Claude Sonnet 4.6) accepts ConSynth-X fog at 0.98–1.00 — on par with the ACDC real-fog calibration baseline (0.95–1.00).
ΔFID against real ACDC / WeatherBench / WeatherNet-05 references: night −34%, snow −19%, fog −10% — synthesis shifts images toward genuine adverse-condition distributions.
IP2P rain/snow variants outperform the style-transfer ablation under both VLM judges (e.g. rain light 0.98–1.00 vs ST rain 0.58–1.00), validating diffusion as the primary weather pipeline.
Appearance-only conditions (rain, snow, fog, night) inherit source bounding boxes unchanged; small-object scale uses a deterministic affine remap — annotations stay exact across all 34,199 records.