Book cover test

I’m not feeling too well today, but I mustered enough energy to test Ideogram’s new model on one of my favorite use cases, generating book covers.

• Being able to precisely place elements and describe them individually is a huge advantage. The JSON spec is almost like a text interface for a graphics program.
• The describe feature lets you study the composition and layout of existing images.
• Book covers benefit from having low feature counts (fewer things) and this system let’s you practice disciplined layout.

• On the downside, the system loses the image-based inputs from the previous version. This makes it harder to develop a style.
• The image outputs are, so far, low contrast and a bit middle-of-the-road.

I can especially see the issues when I shrink the images to thumbnail size (where most covers compete). Using the previous version and a style reference I get an image that works well at both scales, with contrasts and highlights that create an outline around the main figure. The new model’s output doesn’t translate as well in the lower scale image.

I’m hoping image styles get added into this new model. Until then I’ll have to keep trying to tweak my prompts to get better results.