In the company's press release, they say that 900 million images have been generated with Generative Fill -- and the feature is still in beta. I'll bet that nearly every user has tried expanding the canvas using Generative Fill. So what's so special about Generative Expand?
To answer that, let's make sure we're all on the same page by first discussing two powerful Photoshop fill commands.
Content-aware fill is a feature that's been around since about 2010. While it uses some intelligence and has definitely improved year over year in terms of quality, it's not using a generative AI engine to do its work.
Where it excels, and where I've used it extensively, is getting extensions of a textured surface to look pretty good. Take a look at this first image.
Filling the concrete and asphalt, as well as the green muck line in-between, into the white area is a lot of work using other Photoshop tools, like the rubber stamp tool. But by selecting the white area and choosing Content-Aware Fill, you get this:
You can see where the two images join, it's not quite perfect. But it gets the job done enough that a bit of retouching is all it takes to make the fill look good. Five minutes instead of an hour.
Content-Aware Fill, ironically enough, isn't really aware of its content. Oh, it knows that there are bit patterns it wants to reflect, but it has no idea that you're looking at concrete and asphalt.
Take this next example.
Here, we have a nice Oregon barn with a fence and a tree branch. Using Content-Aware Fill, we get this:
It doesn't know that the ground is the ground, the fence is the fence, or the sky is the sky. What it sees are green pixel textures, some white pixel textures, and some bluish pixel textures. As such, you wind up with the grass making it all the way up into the sky.
Content-Aware Fill lets you erase areas of the image where you don't want it to consider pixels, so for the above, I erased the barn. If you just select the white rectangle and don't erase the barn, Content-Aware Fill shows its complete lack of awareness for the content of the image:
Generative Fill, which is still in beta for Photoshop, is in a completely different league. Here's how Generative Fill handled filling in the white area:
Notice how the ground is on the ground. The fence extends to the right of the tree and then behind the tree. The tree casts a shadow, including on the fence. There's even another fenced-in area behind the tree that looks as if it could well be part of the scene.
Yes, the tree is a little anemic, but that can be fixed with more generative fills near the branches. And yes, the main trunk of the tree has a little green in it, but that's a pretty simple retouch.
Perfect, it's not. But if I hadn't told you that the right side of the image was generated by an AI, you probably wouldn't have known.
Expanding on Generative Fill
Almost everyone who has experimented with Generative Fill has done this. Take a base image, increase the canvas size, and then select the white space. Like this:
The red dashes are there so that you can more easily see the area selected.
Then, once there's a selection, all you need to do is click the Generative Fill button. What was once white is filled in with the rest of the scene:
The results are amazing. You can see how it added trees, extended shadows, and even put a building in the background where it makes sense to have a building. All the additions are aware of the direction of the sun, and the shadow generation is spot on.
The newly announced Generative Expand does exactly what I just showed you with Generative Fill, but does it using a different tool and a slightly streamlined workflow.
My method took three steps, three tools, a bunch of clicks, and some typing.
Select Canvas Size from the Image menu and type in the new dimensions, choosing to center the original image. This creates the white block.
Choose the Magic Wand tool and click on the white space. This selects the white space area.
Choose Modify→Expand from the Select menu and type in a pixel value. This creates some overlap with the original image (which makes Generative Fill work more reliably).
Generative Expand uses the Crop tool. Here's how Adobe describes its new functionality:
First, select the Crop tool and drag beyond the image's original canvas to your desired aspect ratio.
Generated content can be added with or without a text prompt. Without a prompt, click "Generate" in the Contextual Task Bar, and Photoshop will fill in the new white space with generated content that seamlessly blends with the existing image.
When using a prompt, the image will be expanded and will include the content you entered into the prompt. Select your favorite variation, and the expanded image will be added non-destructively in a new Generative Layer.
Essentially, we've all been doing Generative Expand, but now there's a version of the feature that works within the Crop tool and may save some steps.
Availability and languages
Adobe announced that the Generative Expand functionality will be available in the Photoshop beta being updated today.
The company also announced that you're now able to talk to the AI in 100 languages. They include Afrikaans, Albanian, Amharic, Arabic, Armenian, Assamese, Azerbaijani (Latin), Bangla, Bashkir, Basque, Bosnian (Latin), Bulgarian, Cantonese (Traditional), Catalan, Chinese (Literary), Chinese Simplified, Chinese Traditional, Croatian, Czech, Danish, Dari, Divehi, Dutch, English, Estonian, Faroese, Filipino, Finnish, French, French (Canada), Galician, Georgian, German, Greek, Gujarati, Haitian Creole, Hebrew, Hindi, Hungarian, Icelandic, Indonesian, Inuktitut, Irish, Italian, Japanese, Kannada, Kazakh, Khmer, Konkani, Korean, Kurdish (Central), Kurdish (Northern), Kyrgyz (Cyrillic), Lao, Latvian, Lithuanian, Macedonian, Maithili, Malagasy, Malay (Latin), Malayalam, Maltese, Marathi, Mongolian (Cyrillic), Myanmar (Burmese), Nepali, Norwegian, Odia, Pashto, Persian, Polish, Portuguese (Brazil), Portuguese (Portugal), Punjabi, Romanian, Russian, Serbian (Cyrillic), Serbian (Latin), Sindhi, Sinhala, Slovak, Slovenian, Somali, Spanish, Swahili, Swedish, Tamil, Tatar, Telugu, Thai, Tibetan, Tigrinya, Turkish, Ukrainian, Upper Sorbian, Urdu, Uyghur, Uzbek (Latin), Vietnamese, Welsh.
That's how you know it's an AI. The only two people I know who speak that many languages are Hoshi Sato and Saru, and neither has been born yet (and, technically, never will be because they're fictional Star Trek characters).
More to come
Before I close, I want to remind you of two particularly valuable differences between Adobe's Generative Fill/Expand and tools like Midjourney.
First, Adobe lets you select an area of the image and tell it what you want to be filled. This article shows you exactly how and why that's so powerful. With Midjourney, you have to create the entire image from text.
Second, and probably the most important aspect of all of Adobe's generative AI offerings: the images generated are suitable for use without copyright concerns. Most AI image generators trained their AIs on all sorts of random and unspecified images on the Internet.
As such, you may well be using part of a copyrighted image in your output if you use one of those tools. But Adobe only used images in its own Adobe Stock collection, all of which it owns and licenses as part of your Creative Cloud license. So you're very safe if you use a Photoshop-generated image and very exposed if you use something from one of the other AIs.
Adobe says it will be introducing more generative AI goodness as Photoshop's new version gets closer to formal release.