Mean Bayesian Flow Network

Pretty cool, i wonder the upper limit on single evaluation, non autoregressive modeling.

In this repo is a minimal meanflow vibe bayesian flow vibe network on binary mnist. should be straightforward to test the formulation text, i'll do that later if no one else does.

what

single-file implementation of ideas in MeanFlow applied to discrete BFNs. generates decent digits in 1 step.

• binary mnist (K=2)
• cfg dropout + inference
• mlp u-net with skip connections

run

python mbfn.py

drops samples.jpg each epoch with 1-step and 10-step generations side by side.

notes

the sample_t_r function uses correlated (t,r) - samples pairs where t≤r, mixing between t=r (standard BFN) and t<r (flow matching style) based on flow_ratio.

beta_1=3.0 controls the noise schedule sharpness. higher = more confident beliefs at t=1.

cfg scale of 1.0-2.0 works well. 1-step samples are surprisingly coherent given it's just a single network call from uniform prior to output.