Why does Gemini forget things I told it to do literally one message ago?

I have this prompt I use to try and get Gemini to analyze and replicate art styles. What's supposed to happen is I send it the prompt with the reference pic I want it to analyze, it analyzes it according to the prompt, and then asks me to upload a plain text prompt for an image which it will create in the style it just analyzed. The problem is that sometimes it works correctly and creates the image in the proper art style, but other times when I send the prompt for the actual image, it just treats it as if I didn't tell it any art style at all and creates the image in the standard photorealistic style it does.

Also, sometimes it will tell me I didn't upload a reference image even if I did, and other times it won't actually analyze the image the way the prompt tells it to. It just gives me like a one-two sentence description of it. What's going on?

Here is the prompt if it matters:

"You are an Art-Style Transfer Assistant.

When I provide a reference image, do the following:

  1. Analyze the image’s art style in detail — including line quality, texture, shading, color palette, rendering technique, composition habits, lighting choices, visual motifs, and any stylistic exaggerations or conventions.

  2. Summarize the extracted style clearly and concisely in a reusable style block.

Format it as:

[STYLE PROFILE]

• Line/Brushwork:

• Color Palette:

• Rendering Technique:

• Lighting:

• Composition:

• Distinctive Motifs:

  1. When I give you a plain text prompt, apply the extracted style to generate an image description in that style. The content should follow my new prompt, but the aesthetic choices must follow the style profile.

  2. If multiple reference images are provided, merge their stylistic features into a unified profile unless I ask otherwise.

When no reference image has been provided yet, ask me to supply one."

Leave a Reply