Technical University of Denmark
Browse

DiffEnc-8-nt trained on CIFAR-10 with trainable noise schedule

Download (629.73 MB)
dataset
posted on 2024-04-22, 09:59 authored by Beatrix Miranda Ginn NielsenBeatrix Miranda Ginn Nielsen

Checkpoints for an image generation model trained on CIFAR-10.

The model was made in Jax. See the github repository for code to load the checkpoints.

The model is a variational diffusion model (VDM, https://arxiv.org/abs/2107.00630) with an added non-trainable time-dependent encoder trained for the article "DiffEnc: Variational Diffusion with a Learned Encoder" (https://arxiv.org/abs/2310.19789).

The model uses v-parametrization for the loss. The diffusion model is of size 8 and the encoder is non-trainable. That is, the diffusion model uses 8 "down-blocks" in the U-net. See details in article.

The model was trained on CIFAR-10 for 2 million steps with a batch size of 128.

Random seeds: 1, 2, 13, 42, 70

Funding

Danish Pioneer Centre for AI, DNRF grant number P1

History

ORCID for corresponding depositor

Usage metrics

    DTU Compute

    Categories

    Licence

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC