Please implement color correction for img2img results to match original colors

[andrewwashington732]

Feature Idea

SwarmUI is an amazing tool and there is only one missing feature that keeps me from using it fulltime:

In the old Automatic1111 WebUI and in Forge there is an option to "Apply color correction to img2img results to match original colors."

This applies color correction after generation so that the colors are close to those of the original image. It is possible to do in ComfyUI using color matching nodes, so I think it could be possible to implement in Swarm as well.

There is something like this already implemented, but I think it only applies to inpainting.

This feature would be very useful for those like me that are using AI to enhance their own digital art.

[mcmonkey4eva]

That was a hack for SDv1 models, because there was a pink-shift effect each time the VAE re-encoded.
SDXL and later models don't have that issue.
So I'd recommend just... not using SDv1 models in the first place. There's far more quality issues than just that one. Newer models are a lot better.

Flux has its own vae color issue, but it's more complex than the SDv1 one -- you can find a color correction parameter for it under advanced sampling:

(This param is not perfect for the flux issue, but can help with inpaint color shifts a bit)

[andrewwashington732]

Thank you so much for the (incredibly fast) reply.

I am not using SDv1 models. I mostly use XL checkpoints and sometimes Flux.

Yes, that option was implemented for the pink-shift problem, but I use it all the time in XL. Most XL-based checkpoints have a tendency to change the colors when doing img2img at higher denoising values. Even more so when using certain loras.

Here is a quick example I set up now so you can see why that option is useful to me. Take a look at the woman's dress:

This is using an Illustrious checkpoint.

Here is a screenshot of what I'm using in ComfyUI for color correction:

In WebUI the processing is done in the apply_color_correction function in the processing.py file

Without this option I'd have to manually adjust the colors in Photoshop when integrating the generated image into my artwork.
I know you closed this issue, but I hope you could reconsider it because it really is very useful in my process.

That was a hack for SDv1 models, because there was a pink-shift effect each time the VAE re-encoded. SDXL and later models don't have that issue. So I'd recommend just... not using SDv1 models in the first place. There's far more quality issues than just that one. Newer models are a lot better.

Flux has its own vae color issue, but it's more complex than the SDv1 one -- you can find a color correction parameter for it under advanced sampling:

(This param is not perfect for the flux issue, but can help with inpaint color shifts a bit)