Will AI images replace stock photos? We put Firefly 3 integrated into Photoshop to the test

At the MAX London event Adobe presented Firefly Image 3, the third version of its AI to generate images also integrated into Photoshop, for now only in the beta version. Now Firefly 3 gives Photoshop the ability to also generate “images from scratch” within the program, whereas until now it was only possible to do so within a selection as a “generative fill”, or as an extension of the crop of an image.

We therefore wanted to put Firefly 3 to the test in Photoshop Beta to understand how well it can compete with photos available on a stock image platform.

The meaning of the test is to understand how much the personalized creation of an image, also in reference to the dimensions chosen by the user based on his needs, can be better than stock images which instead have to be “browsed” among a thousand others to find the one right.

It should be remembered that Firefly is trained on licensed content, such as Adobe Stock. And the company says it is designed to generate content for commercial use that does not infringe copyright and other intellectual property rights, such as trademarks and logos. Adobe also uses a multi-layered continuous review and moderation approach to block and remove content that violates Adobe policies. Even in our test, the model refused to generate “legitimate” images out of an excess of caution, which was however understandable.

For information purposes, we list what’s new in Firefly 3 in Photoshop:

• Generate image: creates an image from scratch, thanks to the Firelfy 3 generative model, based on its textual description

• Reference image: the user can use an image to guide the style of what Firefly 3 will create.

• Generate background: After removing the background from an image, Firefly 3’s AI can create a new one based on the user’s text description.

• Generate similar: Firefly basically generates 3 versions of an image. For each of these the user has the possibility to create a new one which will be similar to the one chosen.

• Enhance details: Only for images created within a selection, Firefly 3 can increase their details.

Photoshop versus stock images, the test

We started from the most classic example. You need a stock photo for a certain need, so we created it in Photoshop (PS) with a simple but fairly limited prompt and then tried to see if we could find a similar example in a stock image platform, in this case, Depositphotos.

The prompt is “A young woman in a swimsuit sitting by the pool”: the program also accepts prompts in Italian. PS created three variations and we chose this one. You can see slight inaccuracies on the fingers and on the wooden strips, but the photo has its own coherence.

On Depositphotos, the search didn’t take long, and we chose a profile photo that could be compared with the one on PS.

In the previous test PS fared quite well. So, having seen an interesting photo on Depositphotos, this time we tried to find out how well PS could replicate it by writing a prompt that described it. The stock photo is this.

We then created an image in PS with this prompt: “Photo of an elderly man with white hair and a pigtail, beard and moustache, smoking a cigar on a black background”. The result was this.

Beyond the usual problems with the fingers of generative AI, the photo only partially respected the prompt; therefore we took the opportunity to test the generation with a reference image, creating a new one inspired by the stock one. The result is slightly better, but still a little distant.

Having become familiar with the reference image we tried to create one that was similar to this stock image of a man on a construction site.

And then we tested a couple of cocktails generated based on the stock image.

We then moved on to something more complex, this time using the same resolution as the stock image as a “canvas” and using as a prompt the exact English text that describes the photo on Depositphotos: “Woman kiter in sunny weather on the sea” . The result has many inconsistencies.

Much better instead starting from something simpler, linear and without human beings: “A beach at sun with beach chair and umbrella, sea at horizon”. However, in this case we did not use the reference image.

For all examples we used generations and stock images in the order of approximately 2500 x 1500 pixels. We then wanted to try to push the resolution by imagining a 24 MP sensor with a resolution of 6000 x 4000 pixels for 300 ppi. The generation prompt is: “A steam train hurtling across an icy plain, 24mm focal length”.

The impact of the photo is excellent, but by zooming in you notice inconsistencies in many details and above all you understand that the generation was spread over the 24 million pixels, graining the original image created with a lower resolution.

Then we used generative fill to add, on the left, a train station with a wooden platform, and a UFO in the sky. The result is so-so.

The test did not go badly, if we consider that these diffusion models for image generation always have a certain degree of hallucination and inconsistency.

The best results were undoubtedly the woman in the swimming pool, the man in the construction site and the deckchairs by the sea, which can be considered a good starting point, perhaps for images to combine with a post on social media.

However, the replacement of stock photos with AI still appears to be a long way off, even if the step forward has been notable, especially when referring to the photo editing program par excellence which has given itself an important ability and which will only improve.