Summary for Text in Images
The Text in Images category reveals a significant divide between models optimized for typography and those focused purely on aesthetics. The standout performer in this analysis is Nano Banana Pro, which achieved near-perfect scores across diverse challenges, demonstrating an exceptional ability to handle complex layouts like magazine covers and movie posters without the common "gibberish" artifacts that plague other models.
Key Findings
- Top Tier Performance: Nano Banana Pro and ChatGPT 4o consistently delivered accurate spelling, even for complex strings and specific time displays.
- Aesthetics vs. Accuracy: While models like Midjourney v7 produced visually stunning textures and lighting, they frequently failed basic text accuracy tests (e.g., misspelled headlines, incorrect clock times), resulting in lower overall scores for this specific category.
- Secondary Text Evolution: A major trend is the improvement in "secondary text." Top models now populate background elements (like movie credits) with legible, logical words rather than the alien-like symbols seen in older generations.
- Material Understanding: High-performing models correctly rendered text materials (neon glass, icing, fabric), whereas some models hallucinated incorrect textures (e.g., DALL-E 3 rendering a stop sign as leather).
Comparative Strengths and Weaknesses
1. The Precision Leaders
Nano Banana Pro and ChatGPT 4o have set a new standard for text adherence. In complex prompts like The Tech Magazine, where layout and hierarchy are crucial, these models produced professional-grade results. They distinguish themselves by adhering to specific font style requests (serif vs. script) more reliably than competitors.
2. The "Gibberish" Barrier
A recurring failure mode for models like Flux 1.1 Pro Ultra and Midjourney V6.1 is the generation of nonsensical text in peripheral areas. For example, in the Times Square Billboard prompt, while the main billboard might be correct, the surrounding city signage often devolves into incoherent symbols, breaking the immersion of the scene.
3. Stylistic Integration vs. Overlay
Top performers integrate text into the physical world. For instance, in the Neon Sign prompt, models like Midjourney V6.1 (despite its text struggles elsewhere) and Nano Banana Pro rendered convincing glass tubing and light diffusion. Lower-scoring models often made the text look like a flat digital overlay floating on top of the image, lacking proper perspective or texture interaction.
4. Hard Failures on Specific Data
Prompts requiring specific numeric data, such as The Digital Clock requesting "10:45", revealed hard limitations in some models. DALL-E 3 and Midjourney v7 failed to reproduce the exact numbers, suggesting a disconnect between the prompt understanding and the visual generator for specific alphanumeric sequences.
