Written by Ethan Smith

Table of Contents

Intro and review of failure cases

While SDXL has greatly improved adherence to prompts, there are still several notable places where things fall short. Some things that I think are somewhat inherent to using clip as the text condition.

Here’s a couple of failure cases I’ll roughly formalize:

Misattribution

• an adjective/color/characteristic specified for one subject bleeds into the remainder of the scene or other subjects.

You’ve probably seen it happen. You ask for “A cat with green eyes resting upon a pair of shoes” and the green infiltrates into other parts of the scene.

Neglect (Competition/Dominance):

• a specified concept does not appear in result, or maybe you can get it to in some cases, but upon adding more to the prompt, it vanishes favoring the other specified subject.

“Photo of a red sphere on top of a blue cube. Behind them is a green triangle, on the right is a dog, on the left is a cat”