Summary:
In a creative project involving a fictional penguin character and climate-themed comics, I’ve observed persistent residual artifacts in DALL·E image generations. Despite shifting context and specifying exclusions, the system reintroduces elements from earlier prompts—specifically a Gollum-like figure, a campfire, and a frying pan.
Details:
- The earlier prompt involved a satirical scene with a Gollum-like character reacting to food preparation.
- Subsequent prompts explicitly removed all references to that scene, asking instead for a penguin wedding disrupted by an airplane’s contrail (style: KC Green meets New Yorker).
- Despite this, the generated image included the previous campfire and character again—unrelated to the new request.
- The drift appears to have strengthened with time, not weakened.
- This happened after multiple successful prompt shifts and style adaptations.
Why it matters:
Creative users working with complex visual storytelling need model consistency and stylistic isolation between prompts—especially when evolving characters and narrative arcs. Residual generation undermines differentiation, originality, and user trust in output reliability.
Ask:
Is this a known behavior?
Are there plans for stronger prompt boundary enforcement or memory flushing across visual prompts?
Should we expect image model caching/contamination when working in stylistic or narrative series?
