Skip to main content
T
turner111
Inspiring
May 6, 2026
解決済み

How to use generative AI on part of an image without a distorted result?

I’m trying to test out a few roofing materials, and have been using a rectangular marquee around the roof.

It generally works, with the caveat that it’s distorting the final result, even when given explicit instructions not to.

So - what’s the best way to avoid this?

Bonus question: Which models work best for this kind of stuff? That is, preserving the original image while changing a few details and retouching?

thanks

    解決に役立った回答 Trevor.Dennis

    What model did you use?

    If you selected a small part of the layer then the Firefly models work best.  Nano Banana tends to reposition the generated pixels so they don’t align with the original.  You need to select the entire layer (Ctrl A) with Nano Banana.

    Unfortunately this only generates at 1024 x 1024, so will stretch the generated area downgrading IQ.  What you can do is align the new and original layers as best as possible.  Temporarily setting the upper layer to Difference will aid this step  — get the layer to be as black as possible, and set the mode back to Normal.

    Now add a Hide All mask to the upper layer, and carefully paint in the area you want with a soft white brush.  That area will have the degraded IQ, but the rest will be back to full resolution.

    Alternatively, if you have a good upsize plugin like Topaz Photo or Gigapixel, reduce the image to 1024 on the long side.  Use Gen Fill and then upres.  Topaz Labs reduced the price of its subscriptions a couple of weeks ago, and you can now get the full Studio subscription for US$163.

    返信数 2

    Trevor.Dennis
    Community Expert
    Trevor.DennisCommunity Expert解決！
    Community Expert
    May 7, 2026

    What model did you use?

    If you selected a small part of the layer then the Firefly models work best.  Nano Banana tends to reposition the generated pixels so they don’t align with the original.  You need to select the entire layer (Ctrl A) with Nano Banana.

    Unfortunately this only generates at 1024 x 1024, so will stretch the generated area downgrading IQ.  What you can do is align the new and original layers as best as possible.  Temporarily setting the upper layer to Difference will aid this step  — get the layer to be as black as possible, and set the mode back to Normal.

    Now add a Hide All mask to the upper layer, and carefully paint in the area you want with a soft white brush.  That area will have the degraded IQ, but the rest will be back to full resolution.

    Alternatively, if you have a good upsize plugin like Topaz Photo or Gigapixel, reduce the image to 1024 on the long side.  Use Gen Fill and then upres.  Topaz Labs reduced the price of its subscriptions a couple of weeks ago, and you can now get the full Studio subscription for US$163.

      Trevor.Dennis
      Community Expert
      Trevor.DennisCommunity Expert
      Community Expert
      May 7, 2026

      BTW the workflow I detailed above came from Jesús Rameriz ‘Photoshop Ai Mastery’ course.  I thoroughly recommend it if you want to understand the quirks of the Ai models.  I think it might be a bit cheaper now, compared to what I paid for it.

        T
        turner111作成者
        Inspiring
        May 7, 2026

        Thanks - I used Nano Banana, and just selected the area containing the roof.
        Based on your notes, I think I might try selecting the area, then making a new document into which I just paste the part I’d like updated, generating the fill, then copy-pasting it back into the original.

        Thanks for the heads-up about resolution - did not know that. Kind of disappointing, as I thought it was Nano Banana Pro. I didn’t check, but just really started trying it out.

        Sameer K
        Community Manager
        Sameer KCommunity Manager
        Community Manager
        May 6, 2026

        Hey, @turner111. Welcome to the Photoshop Community. I'll help you figure this out, to some extent.


        Without examples, images, and results, I'll only have to guess based on the description. My guess is that the roof in question is slanted, so I'll recommend using the Polygonal lasso tool to select the desired area at an angle, with a slight margin of 5-8 px.


        Some Firefly models, and even partner models, may not support negative prompting, i.e., prompting them to do what not to do. So, any preventive prompts may not help, but confuse the generation. You can try focusing on what you want to generate. Firefly does a good job in contextualizing your prompt to the selection.


        There would be about 10 more ways to do this better. I'd recommend that you wait for our Experts & seasoned professionals to let us know.


        Thanks!
        Sameer K
        (Type '@' and type my name to mention me when you reply)

          契約条件Accessibility statement
          Using the community Terms of use Privacy Cookie preferences Do not sell or share my personal information ad choicesAdChoices