Stability A/B report: NERVE vs SPAN_S

Setup. Both networks were trained from scratch with identical configs - only network_g differs (NERVE vs SPAN_S). Same data, same degradation, same losses, same schedule. Everything is in configs/.

Validation. Urban100 x4 bicubic (standard, no self-made degradation). PSNR/SSIM at each validation step.\n\n## Graphs

losses.png - training losses (charbonnier, MS-SSIM, perceptual) for both networks.
validation.png - validation PSNR (left) and SSIM (right).
colour_stability.png - the average per-channel colour difference between the model output and the ground truth, in 0-255 units, at each checkpoint. One panel per colour channel plus a combined "largest bias" panel. Flat lines near 0 mean the model is not developing a colour cast while it trains; big swings or a rising trend mean it is.

Raw numbers: colour_bias.csv.