How to use new mask feature?

[MoreColors123]

I tried but failed, how do I use this exactly? Could someone give me a hint?

[MoreColors123]

Did anyone already use this feature? I'm still not getting it.

[Acly]

Key things:

• Use Selection Masks to create your mask. You can edit them with soft brush, or gradient fill, etc.
• Use 90% Refine, or consider 100% Custom inpaint and disable seamless if you want more "drastic" changes in the image

Differential.mp4

[SyamsQ]

I use retangular selection, but red area doesn't appears. How to do it.?

[Muskelmagier]

while it probably is a nice feature but what it comes down to is that I have to manually soften the edges of every inpaint I do otherwise I have visible edges in my pictures.

[Acly]

Please post pictures/workflows that can ideally reproduced. I have a decently sized benchmark but not seen any issues. Mostly transitions improved slightly because it's now blending denoising strength, not alpha (which is more of a hack).

[Muskelmagier]

Raw

after inpaint strength: 40%

raw,

after inpaint, strength: 40%

a few examples I created in 3 minutes. it's not always there because it heavily depends on what kind of region you inpaint in and how many different colors are present, but when you have a big complex piece it gets increasingly more visible due to the slight color shift rims everywhere.

[Acly]

Thanks. I'm guessing it's an issue in those situations where for some reason the inpainting cannot match the colors of the context at all. Even a low denoising strength is enough to allow subtle color shift, which become more obvious when they are over a large region.

If you open comfy_workflow.py and find the line

def denoise_to_compositing_mask(self, mask: Output, offset=0.15, threshold=0.25):

you can tweak the values which determine the ratio between "soft inpainting" (differential diffusion) vs. alpha blending.

If you set offset=0 and threshold=1 you should get pure alpha blending (original behavior). But soft inpainting would no longer work.

It would be interesting to get values that still provide good results in all cases. Try something like offset=0.1 and threshold=0.35 etc.

(And yes we can make a slider, but that's always more work, and good defaults are still important regardless)