Occasionally failing to follow reference

[frozen sonnet]

I am working with the image-edit endpoint and trying to generate an image of a specific character using reference images and a detailed prompt.

Here’s what I’m doing:

I upload two reference images:

• The first is labeled as "Primary style" (the desired visual style for the output).
• The second is labeled as "@Alex" (the specific character I want the output to resemble).

My prompt specifies:

• Generate an image of @Alex doing some action. Make him look exactly like the attached @Alex

However, instead of following Alex's look, it takes a character from the primary style image.
The strangest part is that it only happens sometimes, and not always!

Why is that happening?

[raven roost]

Well, every model does that, I feel like. It's probably just something to do with image generation models not really having a good understanding of your prompt sometimes.

[amber root]

Occasionally? It's EVERY prompt. I just put in this description twice:

"A stylized portrait of a beautiful young East Asian woman with shoulder-length black hair and a warm, gentle expression. She has soft facial features. The art style is detailed and elegant, with warm lighting and a softly blurred background. Her eyes are expressive, her posture relaxed, and the overall tone is peaceful and confident. No accessories, no props — just a natural and flattering artistic depiction."

I got a no smoking sign and some dude.

Image generation has been broken for several days now.

[raven roost]

Hmm…

I could help.

Here you go.