edm2_sCM

Checkpoints for Stable Continuous-Time Consistency Distillation: An Empirical Study with a Multistep Extension (Transactions on Machine Learning Research, 2026, OpenReview).

Vinay Saji Mathew, Soundar R. Kumara and Gretta D. Kellogg (The Pennsylvania State University); William KM Lai (State University of New York at Buffalo; part of this work was done at Cornell University). Correspondence: wklai2@buffalo.edu.

Code: github.com/EpiGenomicsCode/edm2_sCM. The README there has the exact FID and training command for every checkpoint.

hf download vinaymatt/edm2_sCM --local-dir checkpoints

Contents

cifar10/
  teacher/    TrigFlow teacher, DDPM++
  scd/        sCD student
  sct/        sCT student
imagenet64/
  teacher/    TrigFlow teacher, EDM2-S
  scd/        sCD student
  sct/        sCT student
  ms_scd_m2/  MS-sCD, M=2
  ms_scd_m4/  MS-sCD, M=4
  ms_scd_m8/  MS-sCD, M=8
  mscd_m2/    MSCD baseline, M=2
  mscd_m4/    MSCD baseline, M=4
  mscd_m8/    MSCD baseline, M=8
  mm_s8/      moment-matching student, 8 steps, EDM ADM backbone
fid_refs/
  cifar10-32x32.npz           CIFAR-10 Inception statistics
  edm2_img64_custom_ref.pkl   ImageNet-64 training-set Inception statistics
configs/
  models.json                 sampler settings and reported FID per checkpoint

Each checkpoint is a network pickle that holds only what inference needs (ema, plus encoder where the model has one), with weights stored on CPU. Load it with the code above: pickle.load(f)['ema'].

Reported FID

50k samples, no guidance. The paper reports the best FID over a sweep of generator seeds.

CIFAR-10, unconditional. Reference: fid_refs/cifar10-32x32.npz.

Model	NFE	FID
TrigFlow teacher	35	2.08
sCD	1 / 2	3.59 / 2.39
sCT	1 / 2	2.88 / 2.09

ImageNet-64, class-conditional, EDM2-S. Reference: fid_refs/edm2_img64_custom_ref.pkl; other ImageNet-64 references give different values.

Model	NFE	FID
TrigFlow teacher	63	1.83
sCD	1 / 2	3.59 / 2.66
sCT	1 / 2	4.34 / 3.98
MS-sCD, M=2	2	2.53
MS-sCD, M=4	4	2.26
MS-sCD, M=8	8	2.08
MSCD, M=2 / 4 / 8	2 / 4 / 8	3.44 / 2.32 / 1.79

ImageNet-64 moment matching, EDM ADM backbone. Reference: EDM's imagenet-64x64.npz.

Model	NFE	FID
Moment matching	8	1.4

License and citation

CC BY-NC-SA 4.0, following NVIDIA EDM and EDM2.

@article{mathew2026stable,
  title   = {Stable Continuous-Time Consistency Distillation: An Empirical Study with a Multistep Extension},
  author  = {Mathew, Vinay Saji and Kumara, Soundar R. and Kellogg, Gretta D. and Lai, William KM},
  journal = {Transactions on Machine Learning Research},
  year    = {2026},
  url     = {https://openreview.net/forum?id=di6ofoWEU8}
}