Why the upscaled image is noisy?
Your upscaler is nice! It's very fast.
My model, Cool Japan Diffusion, is the finetuned model of Stable Diffusion 2.1 .
I generated a image by my model.
However, the image upscaled by your model is noisy.
Why the upscaled image is noisy?
This is because the VAE of my model is difference from your model?
Or this is because the U-Net of my model is difference from your model?
Do I need to finetune the your model?
The image by my model (1024x1024) .
The upscaled image by your model (2048x2048) .
@alfredplpl
I love your space! Just had quite some fun playing with it:)
I've only tested it on 512 x 512, but It should work with your fine-tuned model!
you can swap out the VAE in upscaler, and use the VAE in your model instead to see if it improves:)
Thanks! I will try to swap the VAEs.
I have the same issue. Using the same VAE for the model and the upscaler. Using SD 1.5 stock. I've tried both using CFG 0 and using a prompt and setting CFG at 0.1, 0.5 and 7. Nothing seems to change, the output seems over-sharpened or something. What am I doing wrong?
1024x1024
Upscaled
Any resources or pointers (or ideally code) on fine tuning the VAE for a custom dataset?
the same issue when upscale images generated by finetuned Stable diffusion Unet. I think this latent upscaler model has a lot of limiitations.