Type | VAE |
Stats | 375 12,051 |
Reviews | (27) |
Published | Nov 9, 2024 |
Base Model | |
Training | Steps: 360,000 Epochs: 30 |
Hash | AutoV2 9723DEDCC4 |
v2(?Future?) - Resume from v1.5, increase resolution for testing. Release if viable.
v1.5 - Same dataset, loss rebalanced, longer training.
It seem to improve cluttered line fidelity a bit, some artefacts are reduced. Tends to lead to less issues at higher denoise hires fix from what i've seen.
Content reproduction on smal lscale seems to improve(?)
Is less vibrant tho. Would suggest a bit higher CFG, or contrast prompting, if required.
Other than that, it's not a big difference.
This is a VAE trained on 13k anime dataset at pretty standard conditions, to improve and fix existing issues with some models.
Particularly - line artifacting, especially at hires in abstract art, in Illustrious model.
This model also has potential to smooth and improve clarity of small details.
Colors were not changed significantly and remained balanced.
Reduces jpeg and overall noise.
As a side effect, makes image size smaller, lmao.
Why so few showcase screenshots?
Man, i can't be bothered anymore. You won't see it in full images anyway, download and check for yourself. Im tired.
How did i make it
Trained fully in BF16 precision at batch 24 with accumulation of 4 (Effective = 96)
I have not tested if accumulation is effective with VAEs.