Persistence of the EDM timing shift after class-proportion correction

Determine whether the approximately 11% timing shift between the unconditional EDM model and class-mass-matched data persists after correcting the model's class proportions within the sampler.

Background

Mass-matching the data to the unconditional EDM model's generated class proportions makes the model and data agree closely in lopsidedness and substantially reduces pairwise timing discrepancies, but the model still splits approximately 11% earlier than the mass-matched data.

After applying the class-proportion correction in the sampler, the estimated timing ratio is compatible both with exact agreement and with the uncorrected value. The available confidence interval therefore cannot establish whether the residual timing shift is a genuine model error that survives correction.

References

After the proportions are corrected inside the sampler (Section~\ref{sec:correct}), the shift is $1.02$ against the raw data and $1.07$ ($[0.90,1.30]$) against mass-matched data. That interval contains both 1 and the uncorrected value, so we cannot tell whether the shift survives the correction.

— When Does a Diffusion Model Decide What to Draw ?  (2610.05645 - Khilar, 5 Oct 2026) in Section 6, subsection "The decisive test: give the data the model's class proportions"

We could not determine on CIFAR-10 at which noise levels the correction must act.

— When Does a Diffusion Model Decide What to Draw ?  (2610.05645 - Khilar, 5 Oct 2026) in Section "Limitations", item "Correction"; related discussion in Section 6, subsection "Must the correction run during the whole of generation?"