Experiments with Imagen 3
I've been sleeping on Imagen 3, and I'm betting you have been too. Google's new (but not that new) AI image generation model is truly amazing. I have been testing it against Midjourney, FLUX and Leonardo Phoenix and in my opinion it won on all three of my tests; 1) realism, 2) recency of training data, 3) depth of training data.
In these tests I was specifically looking for: image quality (realism) and specificity (nailing the details of the car)
Test 1: Realism - which image model can best reproduce a car that should be very well represented in it's training data (but in an unlikely scene). Winner: line-ball between Imagen 3 and Midjourney for quality, but Imagen 3 has the most realistic car.
Test 2: Recency - which image model can best create an image of a reasonably new (2024) car that is not that well known (the Porsche 911 Dakar)? Winner: Imagen 3 by a mile
Test 3: Depth - which model can create the most realistic scene of a well known older Porsche, and nail all the fine details of the car? Winner: Imagen 3 again by a country mile
I have been using Imagen 3 in ImageFX which is available from Google Labs. All of these tests were from a text prompt.
