I am truly blown away by the pace of generative AI advances. It is happening everywhere, but I’ve focused much of my play in this area on a single platform, Midjourney.
For reference, Midjourney first entered open beta on July 12, 2022, just four months ago. The company has been improving its algorithms, with new versions being released every few months. Version 2 of their algorithm was released before public beta in April 2022 and version 3 in July. On November 10th, 2022, the alpha iteration of version 4 was released to users.
This morning the company released the following updates to v4; Aspect ratios 2:3 and 3:2 are now enabled, slightly improved sharpness on upscale, reduced instances of over-cropping, slightly improved aesthetics, and –chaos now affects image prompts, and is slightly stronger overall. They also highlighted a few “known issues” in the release, including high-aspect images sometimes being over-cropped and sometimes showing borders, and upscaling could still be sharper / higher quality. Essentially saying that even before the release of v5, we’re going to see higher quality still.
I wrote about the future of stock photography on August 13, 2022, and the examples in that post now seem quaint. Take a look at the examples in that post to see for yourself. I have been blown away by how quickly these iterations have been released.
For a side-by-side comparison of the quality progression, I’ve used the same prompt in v2, v3, and v4. The prompt was, “10mm photo, West Virginia creek surrounded by lush forest, autumn, fall colors, thick woods, tall trees, low-angle perspective, morning fog, cinematic atmosphere, sony a7r III.” Each image was rendered at the highest quality possible in each version. They have been edited only to improve their display speed here.
Generative AI – Midjourney Version 2
Generative AI – Midjourney Version 3
Generative AI – Midjourney Version 4
Generative AI – Midjourney Version 5?
At their current pace, we should see the release of v5 sometime in March or April. Of course, we can only speculate on the level of improvement it will bring, but the results from v4 are already enough to know that any level of improvement only further endangers traditional stock photography markets. I will probably never use a stock photography site again. The two examples immediately below were built using the description of similar photos on Shutterstock.
The sample images below were created using the same method. I found an image on Shutterstock, copied the description, and changed only the model’s description. The results are stellar.
Generative AI – Limited Only By Your Imagination
I submit the following samples without further comment. The image prompts are listed in the captions.