Google has unveiled a new version of its Imagen 3 artificial intelligence image generation model that promises improved realism, better prompt adherence and a wider range of custom styles from photorealism and impressionism to abstract and anime.
While you may not be familiar with Imagen 3 itself, if you’ve ever used Gemini to create an image, or even adapted images on an Android phone, chances are you’ve used the model from the Google DeepMind AI lab. The best place to use it is in the ImageFX labs experiment.
With the new update, Imagen 3 has not only gained an improvement in how it renders images but also in how it understands prompts. For example, it now understands the language of photography better than previous models including lens types and lighting. So Imagen 3 has the potential to be one of the best AI image generators.
The best way to put it to the test is to use the completely free ImageFX tool, which is part of Google Labs. This has one particularly unique feature that allows you to quickly adapt the prompt after the first version is generated, for example by switching out lens types.
Putting Imagen 3 to the test
To find out just how well Imagen 3 works I’ve come up with a series of photography-style prompts. Each of these prompts includes a different lens or camera type. Some also have different techniques such as sports photography or photojournalism.
The idea is to see how well the model generates the image, and, more importantly, captures the emotion and feeling of the moment outlined in the prompt.
1. A rainy day in London
(Image credit: Google Imagen 3/AI image)
One thing most models struggle with when asked to generate a street scene is placing the people. They can’t tell the road from the sidewalk but Imagen 3 seems to have got it right, having someone cross the street while others are on the side.
The prompt: “Street-level photograph of a bustling London street on a rainy day, people holding umbrellas as reflections shimmer on wet pavement, shot with a 35mm lens, shallow depth of field focusing on a red double-decker bus in the background, natural light, candid moment.”
2. A moment of reflection
(Image credit: Google Imagen 3/AI image)
This prompt could very easily have failed. Largely because of the fingers. Yes, almost all models have cracked the finger problem but when holding a cup or close up they still sometimes struggle. Add in complexities of depicting age and you easily get an uncanny valley — not so much here.
The prompt: “Golden hour portrait of an elderly woman with weathered hands holding a steaming cup of tea, soft sunlight highlighting her wrinkles and smile, taken with an 85mm f/1.4 lens for a creamy bokeh background, warm and intimate mood, natural outdoor setting.”
3. Feeding the nation
(Image credit: Google Imagen 3/AI image)
Here we had the model depict a specific type of lighting, the complexities of netting and correct shadows for the time of day. It also had to consider the requirement — a democracy-style image.
The prompt: “Photojournalistic image of a fisherman pulling a net from the ocean at sunrise, water droplets glistening in the light, shot on a Canon EOS R5 with a 24-70mm f/2.8 lens, high contrast with sharp detail in the man’s hands and the waves, capturing human resilience.”
4. The Barista’s art
(Image credit: Google Imagen 3/AI image)
Weirdly, latte art is something AI image models can struggle with. Imagen 3 not only got it right but also placed fingers correctly.
The prompt: “Natural light photograph of a barista pouring steamed milk into a cappuccino in a rustic European café, soft focus on the coffee cup while the background remains blurred, shot with a 50mm f/1.8 lens, capturing the steam rising and the texture of the foam.”
5. Caught in the moment
(Image credit: Google Imagen 3/AI image)
I had to do a few tweaks to this image. Originally I wanted to depict sweat droplets but it looked like rain, so I went for the rain motif. Looks good.
The prompt: “Dynamic long exposure shot of a sprinter mid-stride during a track and field race, muscles tensed and rain drops visible in the air, shot with a 70-200mm f/2.8 telephoto lens, fast shutter speed for pin-sharp focus, motion blur in the background.”
6. Full of potential
(Image credit: Google Imagen 3/AI image)
Here I wanted to see if Imagen 3 could capture emotion in an image. Or, at the very least, depict an artistic, model-style photograph and it achieved the goal. Properly capturing the right shadows and harsh light for a black and white image.
The prompt: “High-contrast black and white portrait of a young man standing under a bridge, sharp shadows and highlights emphasizing his angular jawline and intense gaze, taken with a Leica M10 and a 50mm lens, classic film grain effect for a timeless look.”
7. A candid moment
(Image credit: Google Imagen 3/AI image)
This was another image that required some tweaking to get it right. I wanted a casual, but impressive (taken with a good camera) shot of a farmer. It needed to position the farmer in such a way that he looks uncomfortable at having his photograph taken but also proud of his farm.
The prompt: “An environmental portrait of an elderly farmer standing proudly in the middle of a corn field at sunset, blue moonlight casting long shadows, shot with a Nikon Z9 and a 35mm f/1.4 lens, bokeh on the farmer ’s face and hands while the background shows rows of wheat softly blurred, capturing the grit and indifference of rural life.”
One more thing: Terrible photography
(Image credit: Google Imagen 3/Future AI)
I wanted to see how well Imagen 3 could handle bad photography. It is great that models are able to create stunning works of art, realistic brilliant photographs and abstract pieces that lead you to question whether it was human-made or not — but what about bad pictures?
I gave Imagen 3 this prompt to see how it handled the type of terrible photography commonly found in cameras in the 80s and 90s. I wasn’t disappointed.
The prompt: “A poorly lit indoor snapshot taken with a film camera using a harsh flash, saturating the faces of two people sitting at a dinner table, creating red-eye and deep, unflattering shadows on the wall behind them, taken at close range with slightly off-center framing.”
