Krea co-founder and CEO Victor Poes has argued that AI-generated images can feel repetitive because some image models learn a narrow definition of what looks good and then keep reproducing it.
Generative images · Developer perspective
Why AI images can look alike: a training problem, not just a prompting one
Krea CEO Victor Poes argues that some image models learn a narrow definition of what looks good, then keep reproducing it. The result, he says, is a shared visual signature even when prompts differ.
The proposed mechanism
From narrow training to uniform output
Step 1
Training absorbs a narrow sense of "good"
→
Step 2
Model favors a limited range of styles and compositions
→
Step 3
Outputs converge, even for very different prompts
→
Step 4
Users perceive a recognizable visual signature
Poes places part of the cause in the model's learned preferences, not only in users' prompting habits.
Schematic, not measured data
Varied prompts, narrower outputs
Prompt A
Prompt B
Prompt C
Similar look,
shared signature
What the argument offers
Locates part of the repetition in learned preferences, not only in how users prompt.
Suggests better variety may require changes to training and evaluation, not just new interfaces or prompt guidance.
Gives developers a concrete explanation for a complaint long voiced by creators.
What remains unproven
It is one executive's perspective, without independent testing or benchmark results.
No Krea product data shows how widely the effect occurs or how much it can be reduced.
Consistency can be a benefit for customers who want a reliable house style.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...