Trained checkpoints and evaluation archive for "A learned reversible cellular automaton for generative denoising on adiabatic hardware"

This deposit holds the trained checkpoints and the full evaluation archive behind the Article "A learned reversible cellular automaton for generative denoising on adiabatic hardware". It is the companion deposit to the code repository at https://github.com/coradian/icrc2026, which contains the code, the small recorded artefacts, and the per-checkpoint FID evaluations that the figures read; every figure except the sample grid of Fig. 3d reproduces from that repository alone. This deposit is needed only to regenerate the sample grid or to re-run the evaluations from the weights. Contents. Version 2 holds two archives, stored without compression, with no data file in both: icrc2026-large-data-v1.zip (245 files, 12.47 GB), unchanged from version 1: outputs/, 64 trained checkpoints in six runs, cifar10_192gen_hd64_actnorm (ep020–200), cifar10_deep_192gen_hd64 (ep020–200), cifar10_deep_exp21 (ep020–200), cifar10_fresh600_runpod2 (ep050–600), unet_cifar10_200ep (ep020–200) and unet_cifar10_600ep_ema (ep050–600); eval/, the full evaluation archive; eval_outputs/; and MANIFEST.sha256. icrc2026-large-data-v2-additions.zip, what the revised Article adds: outputs/cifar10_nonrev_600ep/ (the control of Table 1, with the bijectivity constraint removed, ep050–600), outputs/cifar10_64gen_hd64_ema_600ep/ (the 10.65M automaton, ep150–600 and model_final.pt), outputs/unet_cifar10_600ep_ema/model_final.pt, the added evaluation records under eval/, README-v2-additions.txt, and MANIFEST-v2-additions.sha256. Its README supersedes the references to release v1, to a version DOI and to Fig. 3c inside the v1 archive. Unpack both at the root of a checkout of the code repository (release v2), so that outputs/ sits beside model/ and figures/. The scripts read outputs/ from the repository root, or from the path in $RCA_OUTPUTS. Verify with sha256sum -c on each manifest. Licensing. This deposit is mixed-licence, and Zenodo records carry a single licence field. The trained checkpoints under outputs/ are released under the Apache License 2.0; the recorded evaluation artefacts under eval/ and eval_outputs/ are released under CC BY 4.0, which is the licence shown on this record. The full texts of both travel inside each archive under LICENSES/, alongside NOTICE and a README that states which files each licence governs; those files are authoritative. The methods described in the Article are the subject of pending patent applications. The Apache-2.0 patent grant covers the copyright holders' contributions and does not extend to other implementations.