Halo on Frequency Separation for Cutout Images

[8BPixelLab]

Hey all,

Newbie here, but a fairly experienced journeyman in retouching. However, this problem is new to me, and has proved perplexing thus far. I was doing a frequency separation on a portrait, but unlike previous images, I had cut this image of the girl out of her background, so she resides on a layer cutout without background or fill/bg color. I noticed once I set up the frequency sep, the high layer gives a gnarly halo around the cutout image.

My question is, is this because the FS was done on a cutout image, and is there anyway to correct for this that doesn't involved adding fill color to the cutout?

Appreciate any insight on this.

[klev]

Re: Halo on Frequency Separation for Cutout Images

That is the expected behavior given your depiction. Ignoring the underlying goals of fs, it uses gaussian blur or some other blur on one layer. If you do that to any image with some fully transparent pixels, you will have some bleed. That's not unexpected.

You get around this somewhat by using a layer mask on the layer containing the girl. Make other adjustments as needed. Blur the outline by 1 pixel if it's too harsh. Check to make sure you don't compromise the transparency of the girl anywhere.

[Tulack]

Re: Halo on Frequency Separation for Cutout Images

Use surface blur instead.

[skoobey]

Re: Halo on Frequency Separation for Cutout Images

You should put a split into a masked group. Deleting pixels instead of using masks is not a good workflow.

[Profesh]

Re: Halo on Frequency Separation for Cutout Images

Hey all,

Newbie here, but a fairly experienced journeyman in retouching. However, this problem is new to me, and has proved perplexing thus far. I was doing a frequency separation on a portrait, but unlike previous images, I had cut this image of the girl out of her background, so she resides on a layer cutout without background or fill/bg color. I noticed once I set up the frequency sep, the high layer gives a gnarly halo around the cutout image.

My question is, is this because the FS was done on a cutout image, and is there anyway to correct for this that doesn't involved adding fill color to the cutout?

Appreciate any insight on this.

While I echo skoobey, above as regards workflow, in the short-term this is probably what you're looking for:

https://snag.gy/GzaHe8.jpg

[8BPixelLab]

Re: Halo on Frequency Separation for Cutout Images

That is the expected behavior given your depiction. Ignoring the underlying goals of fs, it uses gaussian blur or some other blur on one layer. If you do that to any image with some fully transparent pixels, you will have some bleed. That's not unexpected.

You get around this somewhat by using a layer mask on the layer containing the girl. Make other adjustments as needed. Blur the outline by 1 pixel if it's too harsh. Check to make sure you don't compromise the transparency of the girl anywhere.

Why I never 2+2d that the Gaussian blur and the fact that I didn't use a mask is just sloppy on my part.

Use surface blur instead.

I'll test drive that the next time I delete a background.

You should put a split into a masked group. Deleting pixels instead of using masks is not a good workflow.

Yeah, I generally do use masks, but this was a personal project with a redundancy of copies of the original. But you're right, it's a bad habit, and it haunted me here.

While I echo skoobey, above as regards workflow, in the short-term this is probably what you're looking for:

https://snag.gy/GzaHe8.jpg

Thanks, I try to heed the advice to mask out of principle. I usually know better, but occasionally I cave to college habits.

[klev]

Re: Halo on Frequency Separation for Cutout Images

Why I never 2+2d that the Gaussian blur and the fact that I didn't use a mask is just sloppy on my part.

I don't understand what you mean by 2+2d, but I do know some of the mechanics of gaussian blur. It samples neighboring pixels and doesn't lock transparency. I think if you use a mask and then select based on that mask, it might work. It probably requires some experimentation. Otherwise it will either include a contribution from pixels that are later removed from the image or include some level of transparency, as the alpha value is implicitly sampled.