A preprint introduces Metropolis-Adjusted Diffusion Models to correct discretization bias in diffusion model samplers by applying Metropolis adjustments to corrector steps

We propose two approaches: 🔵Exact Barker corrector via a two-coin algorithm: the first exactly adjusted corrector for score-based models 🟢Simpson's rule approximation: efficient with order 5/2 accuracy

Consistent FID gains on FFHQ, AFHQv2 & ImageNet-64 without retraining [2/3]

[📄preprint] Diffusion models 🤝 MCMC !

Diffusion model samplers are biased due to discretisation

💡The fix: Metropolis-type adjustment on corrector steps ❗️Challenge: no access to the density ratio, only the score 🔑Insight: the score (and some maths) is all you need... [1/3]

@tylerfarghly nice work

@tylerfarghly Very cool