Question on intended bin definitions for L2D Loss in RFDiffusion

[halasadi]

I’m working with RFDiffusion and have been comparing the L2D loss described in the Supplementary Information against how it’s computed in RosettaFold2’s publicly available code (for example, in c6d_to_bins2). I noticed that, although both use 37 bins for distances and dihedrals, the supplement defines equal-width bins over 0–18.5 Å (or 2π/37 for angles), whereas the code uses DMIN = 2 Å, DMAX = 20 Å split into 36 uniform bins plus a catch-all bin. The RFDiffusion code‑base uses those same DMIN/DMAX and bin counts, so I wondered if the Supplementary Information might contain a typo. Could you please confirm which scheme is the intended one?

Thank you very much for your time and for sharing your work.