PotentialEP checkpoints

Checkpoints for PotentialEP: Backprop-Free Training of CIFAR-Scale Generative Models via Equilibrium Propagation (Alex Gower, 2026). Code: github.com/alexgower/PotentialEP.

Unconditional CIFAR-10, 50k samples, EMA weights, after the contrastive phase.

folder	model	gradient	params	FID	τ_s
em_replication/	Energy Matching (replication)	backprop	50M	3.54	3.25
ift/	PotentialEP	implicit solve (IFT)	50M	3.64	3.25
ep/	PotentialEP	EP	50M	3.64	3.25
norm_attn_free/	Energy Matching, norm- and attention-free	backprop	78M	3.51	5.0

PotentialEP rows draw phase-2 contrastive negatives with a feedforward twin; see the paper.

Each phase1/ folder holds the checkpoint the contrastive phase started from (FID at τ_s = 1.0):

file	FID	note
em_replication/phase1/checkpoint_145000.pt	6.42	the checkpoint every correspondence diagnostic in the paper runs on
ift/phase1/checkpoint_145000.pt	6.70
ep/phase1/checkpoint_100000.pt	6.23	what ep/checkpoint.pt phase-2 started from
ep/phase1/checkpoint_145000.pt	6.87	the end of the EP phase-1 run

Usage

Every file is a training checkpoint dict (net_model, ema_model, sched, optim, step) saved by the code repository's training scripts, so it loads directly with --resume_ckpt. All paper numbers use ema_model (--use_ema). The exact FID command for each checkpoint, including the relaxation settings for the PCN arms (the EP model is sampled with --pcn_gamma=0.003), is in checkpoints_main_arms/EXPLANATION.md.

hf download alexgower/PotentialEP ep/checkpoint.pt --local-dir ./ckpts

SHA256

file	sha256
ep/checkpoint.pt	98a9884c81142bd21a5c8d2f1685da85b571616e52caa9e857a65e8dc61b6a4b
ift/checkpoint.pt	58b8ea2d4b0fa904a4afb4de2b4bced992a9323f580c251071453cfdd6f06b06
em_replication/checkpoint.pt	72e87f6d90e251a0990d61cd706248d27cb8c644e4bdb1b6631cd585661bf211
norm_attn_free/checkpoint.pt	5b3141e436804b7f5db0bb580f6f624da65aead8485f49a6e84fc8926fc6db97
ep/phase1/checkpoint_100000.pt	94ba3773f97567c9daef8a82f41f011c5e766054f4343efc495ccaf694d09325
ep/phase1/checkpoint_145000.pt	67cc46749eb3b3cb762d5a3aa45abedff547442bf38421554307626428eb71fe
ift/phase1/checkpoint_145000.pt	f67a5bc89416e357ae11eb3da394e6e5200b9a85011a0aa25c3086f7c83e9d46
em_replication/phase1/checkpoint_145000.pt	cf69079e2e813b2a5083952075778fb1db4d5433693bb7eb4c86d714aa122464

Citation

@article{gower2026potentialep,
  title   = {PotentialEP: Backprop-Free Training of CIFAR-Scale Generative Models via Equilibrium Propagation},
  author  = {Gower, Alex},
  journal = {arXiv preprint arXiv:2610.14409},
  year    = {2026}
}