Custom GPT understands Knowledge but cannot use official templates and logos when generating images

GPT builders
1

Hello,

I have built a Custom GPT for my real estate company to generate branded marketing materials, including social media posts.

The GPT has been carefully configured and contains a comprehensive Knowledge base with:

  • Operating Manual
  • Brand Guidelines
  • Official PNG logos
  • Official social media templates
  • Design assets
  • Example outputs

The configuration is correct, the Knowledge files are uploaded correctly, and the GPT successfully reads and understands all of them.

For example, if I ask why a generated image is incorrect, the GPT correctly explains that it should:

  • use the official template;
  • use the official logo files;
  • never recreate logos;
  • preserve the official layout;
  • only replace the photo and text.

So the GPT clearly understands the instructions and recognises exactly what is expected.

However, when I ask it to generate the final image, it does not follow those instructions.

Instead of using the official template and logo files stored in Knowledge, it creates a new layout and recreates similar-looking logos.

In other words, the GPT understands the task correctly but is unable to generate the image accordingly.

Interestingly, when performing similar work in a standard ChatGPT conversation (outside the Custom GPT), image generation is much closer to the official template.

Could you please clarify whether this is expected behaviour?

Specifically:

  1. Can images (PNG templates and logos) stored in a Custom GPT’s Knowledge actually be used as graphical assets during image generation?
  2. Or are they only used as reference material for reasoning?
  3. Is this a current limitation of Custom GPTs?
  4. If so, what is the recommended workflow for organisations that need a Custom GPT to generate marketing materials that strictly follow an existing corporate identity?

I have attached:

  • the official template;
  • the image generated by the Custom GPT.

My objective is simply to understand whether this behaviour is an expected platform limitation or whether there is another recommended approach.

Thank you.

2 Likes