Collection of GPT-image-generator 2.0 issues, bugs, and work-around tips (check first post)

Very well, then bugs only :blush:
I’ll just describe it from my perspective (I’ll try to keep it simple).

To me, this pattern looks like:

  • ‘Undirected’ interference: interference arises from the diffraction (NOT refraction) of light rays.
  • Or superposition: frequencies overlap, and in some places the wave crests and troughs add together. In other places, they cancel each other out and are zero.

I recently conducted an experiment with a red kite. I wanted to depict the natural, physical flight dynamics.

This is one of the images:

Prompt:

“Minimal composition: a red kite flying through dense clouds with a visible rainbow.
The wings must visibly disturb the cloud medium, creating realistic turbulent airflow, vortex trails, and density variations in the mist.
Cloud particles should react with delay and form irregular swirling structures behind the wings.
The rainbow must follow physically accurate geometry, dependent on viewing angle and light direction, not a perfect decorative arc.
Focus on microstructure consistency: fine feather detail, cloud particle variation, and light interaction must remain sharp and plausible even when zoomed in.
No cinematic perfection, no symmetry, no idealization — only physically consistent interaction between air, light, and matter.”

Despite the clear prompt “red kite”, whose defining feature is a forked tail feather, the image generator began to smooth out precisely that feature. The swirls in the clouds also show the pattern you mentioned.

Perhaps this pattern is related to this observation.
Or perhaps this observation can help the developers deliberately provoke these patterns with a few straightforward prompts.

Here is the link to my test.

Note - LLMs are currently facing a similar challenge:

They tend to prioritise only highly relevant tokens in the prompt. As a result, nuances are overlooked leading to vague or distorted summaries of the context.

Some of the LLMs’ outputs are reminiscent of the ‘noise patterns’ produced by image generators.

I wonder if this might stem from a semantic and physical overlap in the definitions of words. Physics defines things precisely (e.g., refraction and diffraction of light), whereas poetry or philosophy can be somewhat too vague in their definitions.