diffusion_spill-ckpts

Training checkpoints from a representation-alignment study on diffusion transformers, moved off a shared research filesystem to free space. This repository is an archive: the checkpoints are kept so the experiments stay reproducible, not because they are intended as ready-to-use models.

1 files, 1.96 GiB, 1 research line(s). The probed checkpoint is at step 1.

Training configuration

Read from the checkpoints themselves (args in the payload):

setting	value
model	SiT-B/2
dataset resolution	256
batch size	64
max train steps	1
encoder (reference only)	pe-vit-g
encoder depth	4
projection coeff	1.0
path type	linear
prediction	v
mixed precision	fp16
seed	0

Checkpoint payload keys: args, ema, model, opt, rng_state, scaler, steps

Runs included

• diffusion_spill/symlinked_20260627_155316/iREPA/ldm/exps/path_switch_bs64_probe/checkpoints

Paths inside this repository mirror the original paths relative to the project root, so a restore puts every file back where the project's provenance checks expect it.

External pretrained models

These runs were trained against: pe-vit-g.

Those third-party weights are not included here. The checkpoints name the encoder in their config and load it at runtime from a local pretrained-models directory; only parameters trained in this project are published. Verified by inspecting the parameter names in every uploaded checkpoint group: no external encoder tensors are present.

Restore

offload_manifest.tsv lists every file with its sha256, byte size, path in this repository and original absolute path. To restore:

from huggingface_hub import hf_hub_download
p = hf_hub_download("sjrksjek/diffusion_spill-ckpts", "<repo path from offload_manifest.tsv>")

Then verify the sha256 against the manifest before use.

Status and licence

Research artifacts, published as-is under CC BY-NC 4.0, with no guarantee of correctness, reproducibility or fitness for any purpose. The numbers derived from these checkpoints live with the project, not in this repository. The non-commercial licence follows the terms of the datasets and pretrained encoders these runs were trained against.