Open eeyrw opened 4 months ago
Hi, In my own experience, it does help improve visualization quality when trying a smaller image size (< 512). I have not measured with other quantitive metrics.
I see. Just another question: why not use VAE implement of diffusers? It seems less dependency such as pytorchlighting.
actually they are the same, just after training, you can replace vae's weight of diffusion with new weights directly. torchlightning is used simply for stable and easier running on multiple gpus.
How about new VAE? Does finetune improve the performance significantly?