Two SegFormer-MiT-B2 checkpoints from a step-matched comparison of real-only vs. diffusion-augmented training for skin-lesion segmentation, on the official ISIC 2018 Task 1 test set (1000 images).

Both runs use identical optimiser steps (32,400 = 2594⁄8 × 100), so the 1.38-Dice gap is not confounded by the augmented run simply seeing more gradient updates. Real-only training wins — consistent with two prior negative results on this exact question earlier in the same project (U-Net and a custom decoder both also lost Dice from the same kind of synthetic augmentation). Seed 1337, single seed per configuration.

Each folder contains best.pt (checkpoint at the epoch with the best validation Dice, used for the numbers above) and last.pt (final-epoch checkpoint).

Check config.json / summary.json in each folder for the exact training configuration and final metrics; loading code matches src/models/registry.py and src/engine/trainer.py in the code repo.

Encoder pretrained via segmentation_models_pytorch's mit_b2 (SegFormer/MiT) ImageNet weights — check that library's upstream licensing before commercial use. Trained and fine-tuned on ISIC 2018 Task 1 (CC0 masks, publicly released dermoscopy images).

Research artifacts from a benchmark measurement study. Do not use for diagnosis or any clinical decision-making.