Somnai/pytti-vqgan-checkpoints

pytti VQGAN checkpoints (preservation mirror)

First-stage VQGAN checkpoints from CompVis/taming-transformers and community retrainings, as used by pytti and the wider VQGAN+CLIP ecosystem (2020-2021 era).

Why this mirror exists

The original hosts for these files have been decaying for years. eaidata.bmk.sh (the canonical WikiArt host) is already dead, batbot.ai and dl.nmkd.de have been intermittent, the FacesHQ koofr share rejects plain HTTP clients, and the remaining heibox (Heidelberg University) shares are single points of failure. This repo is a verbatim, hash-verified mirror so the pytti toolchain has one stable place to download from.

Every file was verified before upload: size sanity, torch/zip container integrity (torch.load(weights_only=True) where the checkpoint format allows it, zip central directory check otherwise), and sha256 recorded below and in manifest.json.

Layout

<model_key>/<model_key>.ckpt
<model_key>/<model_key>.yaml
manifest.json

from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("Somnai/pytti-vqgan-checkpoints", "wikiart/wikiart.ckpt")
yaml = hf_hub_download("Somnai/pytti-vqgan-checkpoints", "wikiart/wikiart.yaml")

Files

imagenet

VQGAN f16, 16384-token codebook, trained on ImageNet (taming.models.vqgan.VQModel)

• imagenet/imagenet.ckpt — 980,092,370 bytes, sha256 845a68805098cb666420d5db93df53f3a3b6dd443e6dd85c05759c5b998cd663
• imagenet/imagenet.yaml — 692 bytes, sha256 00e2c6189926f1d89ecfef73e9598db77981c1982f0555fbade963ffd16143c7
• ckpt provenance: https://heibox.uni-heidelberg.de/f/867b05fc8c4841768640/?dl=1
• yaml provenance: https://heibox.uni-heidelberg.de/f/274fb24ed38341bfa753/?dl=1
• note: downloaded 2026-08-06 from the original heibox (Heidelberg University) links

coco

VQGAN f16, 8192-token codebook, trained on COCO (taming.models.vqgan.VQModel, first-stage export)

• coco/coco.ckpt — 296,030,214 bytes, sha256 106cc20fde571df14afc4349d62218d8213cf47177c682ecaa78a8c28b12f9de
• coco/coco.yaml — 458 bytes, sha256 17d0c2d9fda59eccade2a6070d833804b759dc817da42c2268be8cf1eb047676
• ckpt provenance: http://batbot.ai/models/VQGAN/coco_first_stage.ckpt
• yaml provenance: http://batbot.ai/models/VQGAN/coco_first_stage.yaml
• note: uploaded from a verified local pytti cache copy; earlier canonical host was https://dl.nmkd.de/ai/clip/coco/

wikiart

VQGAN f16, 16384-token codebook, trained on WikiArt/Artemis (taming.models.vqgan.VQModel, wikiart_f16_16384_8145600 lineage)

• wikiart/wikiart.ckpt — 1,005,327,541 bytes, sha256 bd08bb46301f98be1712bb2be9f8868cea30b53137cd985d4d6e8da8b3e02c36
• wikiart/wikiart.yaml — 920 bytes, sha256 6e78241d2828ff35b8381839ce249c24a6c468ed429760d33a25e98e7f24499a
• ckpt provenance: https://github.com/pixray/pixray/releases/download/v1.7.1/vqgan_wikiart_16384.ckpt
• yaml provenance: https://github.com/pixray/pixray/releases/download/v1.7.1/vqgan_wikiart_16384.yaml
• note: canonical host http://eaidata.bmk.sh/data/Wikiart_16384/ is dead; pulled from the pixray project's GitHub release mirror

sflickr

S-FLCKR Net2NetTransformer checkpoint; pytti extracts the first_stage_model.* VQGAN weights

• sflickr/sflickr.ckpt — 4,263,525,412 bytes, sha256 8a8adea3da8dab412675772831370dd948e0aa97bb11a9488a2f328b58dbc929
• sflickr/sflickr.yaml — 1,603 bytes, sha256 c27f012996f3f1f02580f063be47fd60bdd8528efb04d7318b2cbe8a86505b1b
• ckpt provenance: https://heibox.uni-heidelberg.de/d/73487ab6e5314cb5adba/files/?p=%2Fcheckpoints%2Flast.ckpt&dl=1
• yaml provenance: https://heibox.uni-heidelberg.de/d/73487ab6e5314cb5adba/files/?p=%2Fconfigs%2F2020-11-09T13-31-51-project.yaml&dl=1
• note: downloaded 2026-08-06 from the original heibox (Heidelberg University) links

openimages

Gumbel-VQGAN f8, 8192-token codebook, trained on OpenImages (taming.models.vqgan.GumbelVQ)

• openimages/openimages.ckpt — 376,581,823 bytes, sha256 5cd6c74810ab97e00e942c25403f73afc081e8b19987b31ec0d9ff5b68e7ab14
• openimages/openimages.yaml — 745 bytes, sha256 91110bab325067b37f85d1e75895a837d59c292bae4f3097f80c518714d5caff
• ckpt provenance: https://heibox.uni-heidelberg.de/d/2e5662443a6b4307b470/files/?p=%2Fckpts%2Flast.ckpt&dl=1
• yaml provenance: https://heibox.uni-heidelberg.de/d/2e5662443a6b4307b470/files/?p=%2Fconfigs%2Fmodel.yaml&dl=1
• note: downloaded 2026-08-06 from the original heibox (Heidelberg University) links

faceshq

FacesHQ Net2NetTransformer checkpoint; pytti extracts the first_stage_model.* VQGAN weights

• faceshq/faceshq.ckpt — 3,973,478,502 bytes, sha256 3274bba51bd565c90990bbcf7bb5dad0b9b95a6a77011980afc4c3cfc4d1cad1
• faceshq/faceshq.yaml — 1,451 bytes, sha256 726ca97fb266c3c744f26e5aba2e579a51e310eee711a15f22224031dfdd57e2
• ckpt provenance: https://app.koofr.net/content/links/a04deec9-0c59-4673-8b37-3d696fe63a5d/files/get/last.ckpt?path=%2F2020-11-13T21-41-45_faceshq_transformer%2Fcheckpoints%2Flast.ckpt
• yaml provenance: https://drive.google.com/uc?export=download&id=1fHwGx_hnBtC8nsq7hesJvs-Klv-P0gzT
• note: downloaded 2026-08-06; the koofr link requires a Referer header of https://app.koofr.net/links/a04deec9-0c59-4673-8b37-3d696fe63a5d (plain fetch returns 400 'hotlinking not allowed')

License / attribution

These are the standard community-mirrored research weights from the taming-transformers project and its community retrainings (WikiArt by EleutherAI-adjacent community training, COCO first-stage export by @nmkd). No license claim is made here beyond what the upstream project states; see the taming-transformers repository for the original release terms.