The honest answer is that neither the paper nor this card has a direct number for that. The 32.08 / 33.06 / 37.41 / 32.16 PSNR figures (Outdoor-Rain, RainDrop, Snow100K-S, Snow100K-L) are all for net_g_best, the checkpoint trained on the synthetic composite, evaluated against those same synthetic test sets. The paper's real-world comparison (RealSnow, 1,329 actual photos from Snow100K's real-world split) is qualitative only, Figure 6 in the paper shows side-by-side crops, there's no PSNR/SSIM table for it. That's not an oversight on the authors' part either, RealSnow has no paired clean ground truth, so a full-reference metric like PSNR isn't computable for it the same way it is for the synthetic sets. That absence of a number is itself informative: it means "where does it diverge" can't currently be answered with a metric, only by looking.

If I reason about it qualitatively rather than measuring it: the usual gap for this kind of model shows up where the synthetic renderer's assumptions break down. Snow100K's synthetic particles are fairly uniform in size and motion; real snow varies a lot more in particle size, translucency, and motion blur within a single frame, so I'd expect net_g_real to hold up better there than net_g_best would on the same photo. Outdoor-Rain's fog is a synthetic atmospheric-scattering model applied on top of rain streaks; real rain-plus-fog scenes usually also mix in uneven lighting and haze density that the synthetic pipeline doesn't fully capture, so that's probably where the bigger real-world gap sits too.

That's domain reasoning, not a measured result, and I'd rather say that plainly than imply a number I don't have. If it's useful, I do have both checkpoints and the real vs synthetic test sets locally already (that's what the demo's checkpoint selector is built from), so I could run an actual informal side-by-side and post real examples if there's interest. Happy to do that as a follow-up if it'd help more than the qualitative answer above.