Pix2Pix to Img2Img?

[mykeehu]

I was wondering if it is possible to transfer the old Pix2Pix checkpoint colors to Img2Img in some way? I tried, but it keeps the colours. I tried using Denoising to correct the distortions, but unfortunately there was always a minimal difference.
Is there any way to switch from Pix2Pix? Because it has poor generation and takes a long time to train, while the SD quality is excellent and fast, so I would replace Pix2Pix.
Anyone have any experience?

[ClashSAN]

What is Pix2Pix ?
What was desirable in a custom pix2pix model?

It is a matter of finding the right prompts to mimic your model.
There are some nicer things that older models like "dalle-mini" seem to have. It is good for making certain characters.
Use Dreambooth: https://colab.research.google.com/github/TheLastBen/fast-stable-diffusion/blob/main/fast-DreamBooth.ipynb

Or Full finetune Training: https://colab.research.google.com/drive/1vrh_MUSaAMaC5tsLWDxkFILKJ790Z4Bl?usp=sharing

[mykeehu]

Here is more information, this is an older development for image replacement and generation:
https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix

[mykeehu]

The essence of Pix2Pix was to replace parts painted in given colours with objects based on a trained model.
Here the objects could be specified with text, but could be defined for the colours, e.g.:
car: #339966, house: #3366FF, lamp: #FF00FF and so on.
So in Img2Img you could specify which colour is what, so that you don't have to recolour previous images :)

[mykeehu]

This add-on would replace Pix2Pix because it would label the colors. I could imagine it as an extension.

[victorca25]

Pix2pix is trained using label images as inputs and the full image as target, each model is trained for a single translation from one domain to the other, it doesn't have the general purpose behavior of SD and won't work how you think it will, sadly.

The paint with text idea looks interesting, as you're basically labeling the instances of the pseudo label map for each image.

[mykeehu]

The reason I would replace it is because, as with Pix2Pix, the colours correspond to a single label. If I can label the colors in SD as well, I can get pretty much what I wanted. True, I get variations, whereas with Pix2Pix I get a constant end result, but then in SD I can end up with a more flexible and varied result.
I hope you get adjustable blur options, which would definitely be good for triggering sharp color transitions.