AI4Science-WestlakeU/feature-information-dynamics-rae

Feature Information Dynamics: RAE checkpoint chain

Selected EMA weights from the historical pretrained RAE conditional fine-tuning experiment. These are inference checkpoints, not official initialization weights. The optimizer and private resource paths are omitted. The backbone is DiTDH-XL with DINOv2-B register features (768 x 16 x 16); RAE stage-1 encoder and normalization resources are downloaded separately.

File	Measurements
warmup.pt	Unconditional and class-conditioned
mask.pt	Class + mask
canny.pt	Class + mask + Canny

The adjacent JSON files record SHA256, byte counts, training steps, selected epoch, and validation loss. The warmup has no learned spatial gates; the public evaluator initializes these missing gates to zero. The other stages match all parameter keys.

Code and usage: https://github.com/AI4Science-WestlakeU/feature-information-dynamics

Prepared conditions: https://huggingface.co/datasets/AI4Science-WestlakeU/feature-information-dynamics

These finite-model estimates do not establish Bayes MMSE convergence. The runnable remeasurement uses paired images and noise, and reports signed densities; it is not a claim of exact numerical reproduction of the paper figure's larger evaluation set.