DiffPhyCon Jellyfish Checkpoints
Checkpoint archive for the 2D Jellyfish DiffPhyCon reproduction and Flow Matching experiments.
Contents
official/joint_full/model-100.pt: released joint DDPM checkpoint.official/w_full/model-50.pt: released opening-angle prior checkpoint.official/force_surrogate_model/force_model_epoch_9.pth: released force surrogate.official/boundary_updater/boundary_updater_epoch_9.pth: released boundary updater.flow_matching/joint_7ch/joint_step10000.ptthroughjoint_step160000.pt: the complete saved 7-channel Jellyfish Flow Matching joint-model training series, at 10,000-step intervals.flow_matching/joint_7ch/joint_latest.pt: the latest saved joint checkpoint (corresponding to the 160,000-step training state).flow_matching/joint_7ch/SHA256SUMS: integrity checksums for the Flow Matching joint checkpoints.flow_matching/prior_7ch/prior_step10000.ptthroughprior_step70000.pt: the complete saved 7-channel Jellyfish Flow Matching control-prior training series, at 10,000-step intervals.flow_matching/prior_7ch/prior_latest.pt: the latest saved prior checkpoint (corresponding to the 70,000-step training state).flow_matching/prior_7ch/SHA256SUMS: integrity checksums for the Flow Matching prior checkpoints.
The Flow Matching files retain the model, EMA model, optimizer, scheduler, configuration, and training-step state needed for inference and training resumption. Intermediate checkpoints are intentionally retained for exact experiment recovery and checkpoint-selection studies.
Do not confuse the official DDPM control prior at
official/w_full/model-50.pt with the separately trained Flow Matching prior
under flow_matching/prior_7ch/.
Provenance
The files under official/ were released by the authors of
DiffPhyCon. The files
under flow_matching/ are checkpoints produced by the Flow Matching
implementation in this project.
This repository is published for reproducible research with explicit separation between the original DiffPhyCon releases and the Flow Matching checkpoints produced by this project.