sihc-research-checkpoints

SiHC research checkpoints

Weights and resume checkpoints for the residual-stream burden experiments. Use the catalog below to distinguish architectures, training objectives and weight formats.

Choose a checkpoint

• group1/: weights-only pixel-space controls and SiHC variants.
• resume/group1/: corresponding full training checkpoints.
• resume/group6/: clean/velocity models trained in whitened patch coordinates.
• resume/external_group1/: controlled B-size DeCo, DiP and PixelDiT checkpoints.
• flagship/ and scaling_law/: historical blockwise runs. The paper sublayer XL model is linked separately below.
• raev2/: additional research runs.

The withdrawn write-only control is excluded from the paper result index. Its original archive files remain in place for provenance.

Weights-only derivatives omit the optimizer. Resume archives carry additional training state. Use raw model weights for representation measurements and the documented EMA state for generation.

Files and configuration

Machine-readable checkpoint catalog records revisions, file sizes, configuration and architecture classification.

Code and reproduction guide · Research archive · Paper flagship XL

The DINO teacher is needed only for REPA training. Follow the model-specific loader; blockwise and sublayerwise weights are different architectures.

Paper flagship

The selected SiHC-XL + REPA checkpoint is hosted by our coauthor: step_00650520.pt. Its recorded primary-EMA Heun-50 evaluation is FID 1.71491, IS 299.5535, CFG 2.4, interval [0.1, 0.9], 50,000 samples. Historical checkpoints above retain their own identities and metrics.