Add PixelModel v5 to text-to-image leaderboard
#9
by TobiasLogic - opened
FID 34.06 / CLIP 0.2927 at cfg 5.0, n=5000 MS-COCO val2014, same protocol as v4.
Architecture identical to v4 (40,013,980 trainable params); only the training data changed (83k COCO images -> 2,986,571 CC12M images recaptioned in COCO style).
Model card: https://huggingface.co/bench-labs/pixelmodel-v5
TobiasLogic changed pull request status to merged
BenchLabs deleted the
refs/pr/9 ref