When AI “mirrors” a detail… and breaks a character's consistency
I conducted a simple, but very revealing test on the consistency of characters in image generation.
Case: an asymmetrical character (e.g., a bag carried on the character's left shoulder). We start with a reference image (left profile), then we ask for the opposite profile… and the bag “moves” to the other shoulder, as if the model were mirroring rather than truly rotating.
➡️ This is not a “prompt wave” problem. I have tried multiple methods (constraints “left of character, not left of camera”, “no mirror”, “strictly the same person”, etc.) on several AI image generators… without obtaining robust consistency in this specific case.
Thanks to Baptiste Rognon (The AI guy) https://www.linkedin.com/in/baptisterognon/ ) who reproduced tests on his side and confirms: we are reaching a current limit of the models (for now).
My takeaway from this:
1. Some lateralized (left/right) details remain unstable as soon as the angle is changed.
2. The model may “prefer” an implicit symmetry rather than an anatomical continuity.
3. The right reflex is not to “cry bug”, but to design a workflow (editing/inpainting, double referencing, targeted retouching) when consistency is critical.
I've added an image below that clearly illustrates the mirror effect.
If you have found a reliable workaround (locked seed, local editing, multi-reference method…), I'm all ears: I'll compile the best leads and give feedback.
##IA ##GenerativeAI ##PromptEngineering ##CreationVisuelle hashtag#ImageGeneration hashtag#Workflow
