Photoshop AI with Midjourney
I’m sure you know by now that Photoshop has a beta out that can generate images via AI. I’ll go through my experience working with it with the following sets.
Besides having very nice UX (floating panel), the AI engine is quite impressive. To illustrate this, I have prepared 3 sets of images. I will use this Yoga / Ballet set to for the main blog post. While the backgrounds and floors understandably aren’t complex, these are great demonstration of how well Photoshop’s AI engine understand lighting and optics.
Example 1
Yoga
Ballet
Yoga
Ballet
In all three sets, Photoshop accurately extended the walls and floors into a large studio setup. While Midjourney’s original is impressive, the images it generates are often overly tightly cropped — this is somewhat necessary, otherwise your subject will be too small and will lack detail. So… Photoshop AI to the rescue. I expanded each image to twice its original width and height, or 4x the original. Now the figure is properly framed and can be presented in a nice fine art print. For the last image, I also took the opportunity to fix the feet — though I admit hat it a lot of gens (~10) and they’re still not perfect.
Body Type
A lot of people assume that I only like muscular guys because of the content posted here. I actually like lean and and sculpted bodies of dancers quite a bit. I like all types — but the reason a lot of my images are muscular is because MJ tends to generate bodybuilders without wrapping them completely in clothes, thereby reducing time wasted just to get the correct figure. I hope that they will consider changing this practice. I understand why they won’t want to show the upper bodies of women, but when the prompt is clearly asking for a man, it’s often frustrating to not be able to show the musculature.
Example 2: Warriors at the Roman Baths
I thought that I would use an image that I have already created for this second example. The original is a painting, and I was not sure if the Photoshop AI engine can replicate the styles accurately. To my surprise, it picked up the forms and colors nicely. it even attempted to replicate the column designs of the architecture. Where it’s lacking are the intricate details present in the the original painting.
As an important note, I did generate these in small sections — my PSD has 10+ generative layers. one thing that many people didn’t know is that Photoshop AI generates image content at 1024x1024, so if you have made a selection larger than that, your generation will be scaled. Here I did in fact make selections smaller than 1024px side, yet while the generations closely match the existing painting, it is comparatively not as detailed as it could be. Still, it’s very impressive.
Originally, it generated a column on the left because the original had a column (2). I took the opportunity to remove it during the session (3), thus making the image feel less claustrophobic. The final image is the first one in the set (1), while the last image in this set (4) is the Midjourney original that I have originally posted earlier.
Example 3: Cupid
The third example is a painting of Cupid, also seen earlier on my IG feed. The original from Midjourney frames the figure nicely, and I was mostly wondering if it could handle that.
The prompt I used for the all of these outpaints was decidedly simple: “classical painting of greek mythology”.
Photoshop AI did a decent job. The oval frame was kept relatively consistently. I thought that it would just add a bunch of ornaments and call it a day. But instead, it added figures and tried to match the art style. The level of details is frankly lacking, but it’s still very impressive. Where it struggled a lot was the feet, but that’s to be expected. There are also some instances where it tried to keep the whole figure inside the frame, such as the man to the left. Overall, I won’t say that this is a masterpiece, but certainly if I needed a less cropped image to work with for design, it tackled that task nicely.
As the previous set, the first image is the one with the Photoshop generations. 2-3 are edited sections. The 4th and last image is what I have originally published, which was generated with Midjourney with minor tweaks.
Technical Info
Text prompts in Midjourney v5.1, with Photoshop Generative Layers in Photoshop Beta.