How to improve image fidelity?

[soon-yau]

(Left:generated. Right: from dataset)
I have trained on mannequin dataset and the results look quite good. However, the generated images are a bit blurry and fine details are lost. Therefore, I wonder what changes do I need to do make them look crispier.

I currently use VQGAN pretrained on imagenet. I have also tried to train a VAE from scratch (using default train_vae.py) but it is blurry. I tried increasing the number of layers, number of tokens etc but didn't see improvement and made it a bit more unstable. Any advise on what VAE parameters to change?

[rom1504]

You could try to use vqgan 16k or even the f=8 ( see the details on their page)

…

On Thu, Nov 25, 2021, 01:21 Soon-Yau Cheong ***@***.***> wrote: [image: brown] <https://user-images.githubusercontent.com/19167278/143328331-6fb1f1a7-1094-478d-a0c5-615192a9b3e9.png> I have trained on mannequin dataset and the results look quite good. However, the generated images are a bit blurry and fine details are loss. Therefore, I wonder what changes do I need to do make them look better. I currently use VQGAN pretrained on imagenet. I have also tried to train a VAE from scratch (using default train_vae.py) but it is blurry. I tried increasing the number of layers, number of tokens etc but didn't see improvement and made it a bit more unstable. Any advise on what VAE parameters to change? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#390>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAR437XBIS5MZ6L4P3HSRQLUNV6RRANCNFSM5IXI2X6Q> .

[soon-yau]

Managed to train a mannequin-only model with vqgan_gumbel_f8_8192 that produces quality matching OpenAI.

[huang-xx]

@soon-yau Hi, I am trying to train a model using vqgan_gumbel_f8_8192 as you do, but I get terrible results. What model parameters did you use to get this result, especially the number of layers and dimensions? Also, could you please provide a link to the "mannequin dataset"? Thanks