Same Prompts, Different Lenses - Comparing Image Generation in GPT-4o and GPT-5
- Noemi Kaminski
- Aug 13
- 2 min read
Over the past few days as GPT-5 rolled out, I’ve been running a personal AI benchmarking experiment to test how different models handle creative reframing through prompt engineering.
The approach:
Start with one core prompt.
Reframe it for three different creative directions - cinematic, vector design, and editorial.
Compare how the model adapts tone, format, and detail to match the intended medium.
📍 The Process
For GPT-4o, the base prompt was "A crowded street market at night". For GPT-5, the base prompt was "An abandoned train station overgrown with nature".
Both were asked to reinterpret the concept in:
Cinematic Lens → immersive, atmospheric, sensory detail.
Vector Design Lens → simplified, scalable, graphic clarity.
Editorial Lens → realistic, usable in print/layouts, emotionally resonant.
🔍 Key Observations
GPT-4o
Good at shifting style between mediums.
Cinematic lens delivered strong mood and sensory cues.
Vector lens worked well, though not as deeply considered in design composition. Cluttered.
Editorial lens was relatable and human-focused, with richer storytelling emerging even without much guidance.
GPT-5
Sharper compositional awareness across all lenses - stronger use of symmetry, negative space, and framing.
Vector lens showed stronger design fundamentals with clean geometry and high-contrast palettes.
Editorial lens balanced realism with deliberate planning for headlines and copy placement.
📊 Takeaway
Both models adapted the base concept effectively across lenses. GPT-4o tended to produce richer storytelling without much prompting, while GPT-5 excelled at deliberate, compositional precision.
This “Same Prompt, Different Lens” method has been a valuable way to see how AI can intentionally shift creative execution for different media, a skill that’s essential if we want AI to act as a real creative partner rather than just a novelty tool.
Comments