How Do You Get CS5's Content-Aware Fill to Succeed?

[GB1]

I recently upgraded to Photoshop CS5, and one of its big draws its new 'content aware' fill that Adobe showed off last spring (see www.youtube.com/watch?v=NH0aEp1oDOI).

So far I have found that it doesn't quite meet expectations, though I am getting better results as of late. I've noticed that it sometimes leaves strange artifacts in the filled area, and that the filled area is almost always discernible to the discriminating eye. After checking Help, I found "Content-aware fills randomly synthesize similar image content. If you don’t like your original results, choose Edit > Undo, and apply another Content-aware fill."

Uh, yeah.. that's real helpful, Adobe It doesn't work, either.

After getting ready to write it off, someone told me that you have to use smaller fill areas to get decent results. I have seen some success doing this, though not sure how you would approach a big area you want to fix.

What is your approach to the content aware function? Anything to share?

[drg]

The graphic designer mavens I've seen using it are mostly using it as a super cloning brush type tool. Smaller areas that instead of getting edges that need blending, Content Aware has already done it for them.

The bigger 'fills' are very dependent upon what it is being replaced and what in the photograph is similar that can be 'copied'. If you have ever taken or viewed one of the tutorials about how to use the 'cloud generation' tools in Photoshop you probably already get the idea on what can and can't be done.

Pixel peeping is certainly going to show artifacts; after all this tool generates artifacts to get rid of artifacts! Then the Graphic Design magic tools like Gaussian blur and noise and all those other tricks need to be employed to distract the eye! This is a pattern generator and if there isn't a pattern or a gradient type of orientation in the area to be filled, it may look really bad.

Light to dark works well from the center of the image out.
Light to Light only so so.
Dark to light will do strange things most of the time.
Edges between patterned areas (even complex) and smooth areas can even change color!
The later is a problem when shooting something against a sky or a backdrop. CAF needs a demarcation that isn't to be changed to be the most successful from the work I've seen so far. Still a very handy tool for those who are fixing up work for publication.

Content aware is a speed tool. There are newer tutorials from Adobe than the pre release one you referenced, better may be not. But lots of 'types' of images that it will work with better than others.

It seems to be also very dependent upon how complex selections are made (order dependent) plus the use of the various settings for delete. Fixing perspective first seems to be required or it can distort the random content too!

[n8]

I think drg probably covered it all, but I think it really is a tool that shines aspects where there is a lot of repetition in the image for it to safely sample. It won't work miracles, but it will make some jobs a lot easier. I do find that using the smudge tool afterwards can help with some of the artifacts.

[drg]

Two more things with the initial release of the Content Aware Fill tool in the full PS-CS5 product:

1. It doesn't work on a background layer. It may appear to but it isn't doing the same thing. I've seen this demonstrated a number of times now, but will admit the technical reasons haven't been fully explained to me yet.

2. With Skin tones or those type of 'mid-tones' in lighter skinned people it is not the right tool to smooth out skin. CAF just blurs up the skin or creates strange freckles in many cases.
However:
It will work to remove tattoo's and large blemish's quicker than anything I've seen when used with the right combo of selection (manual adjustment of the selection or by using a selection brush). Auto/Magic selection will grab onto a tattoo very quickly it seems!

[GB1]

Thanks for replying fellows.

I played with it a little again last night. It seems to work a tad better now that I took your experiences into consideration, and if I kept the area small.

But I still have big areas that need 'fixing', and not sure how to hit those.

I think I'll post pieces of the problem images here, let you see what I have and what the CAF is doing.

Can you also post examples?

Thanks!

G

[GB1]

Messing w/ it more, I see that it really depends on what area you select. If you want to fix a telephone line then it seems to work OK if you select the line AND a lot of blank adjacent area. Still experimenting though.

Here's the results of one pic. But I also adjusted the shadow density of the fixed one, which sort of skews the evaluation... You can see that the part of the poles are still there in the ground ... context fill couldn't handle those, so they would need to be manually fixed or fixed via another effect. Note too that the fixed area still looked rough where you could tell that something was going on, so I had to guassian blur the background too.

I keep thinking of that demo they gave .. wow, that was a load of stuff, as it's nowhere near as easy as that ... they let the sales team run wild

G

[KmH]

Often, rather than using the Content Aware Fill tool, I just use the Magic Healing brush
and turn on the content aware option.

http://tv.adobe.com/watch/the-russel...e-fill-part-1/
http://tv.adobe.com/watch/the-russel...e-fill-part-2/

[GB1]

Keith - That seems to have its own can of worms. I'm not sorry the tool is available and it does work better than the old Healing brush on occasion, but ... it's just not as they advertised. (gee GB.. not as good as the sales team says .. a first time? )

[drg]

Sorry I didn't return to this post sooner. All that Spam for a few days put me off and I couldn't log on part of the time!!

The one part I did not mention it seems in my earlier comments was that the Content Aware tool looks for a pattern. For it to try and replace a solid object with another solid object (telephone line with solid sky) is not going to fit into pattern recognition algorithms usually.

Now if there were lots of clouds in the sky around the telephone line it might work much better. I've tried somethings like brick walls and architectural features that don't work reliably though they are a pattern. Single objects or big areas are iffy. I do like the added control though when using it a spot mode and it works along edges of photos where there's little else but color or background.