Style consistency is improving, but text is still a liar.
Consistency up. Text still messy. Plan accordingly.
What happened
Midjourney, DALL·E, and Stable Diffusion all show measurable improvements in maintaining character consistency across generations and adhering to style specifications. Text rendering within images remains unreliable across all major platforms.
Why it matters
Image models are getting better at understanding and maintaining visual coherence through improved training on character consistency datasets. Text rendering fails because these models generate pixels, not fonts—they're predicting what text looks like, not actually typesetting it.
What changed
• More consistent characters across generations (better 'same person' continuity)
• Better style adherence when prompts are short and specific
• Text-in-image remains unreliable for anything important
Who it affects
• Creators
• Brand pages
• Design workflows
• Anyone making posters/logos via AI
What to do now
• Use reference images when possible
• Keep style descriptions short (avoid prompt novels)
• Don't put important text inside images (add text later in a real editor)
• Create 2–3 approved style presets and reuse them