[Report] Thorough Comparison of Gemini and GPT in Image Generation and Character Consistency

Survey Overview Combeez Inc. conducted a comparative study on image generation and character consistency using major generative AI models: Google's "Gemini 3.5 Flash" and OpenAI's "GPT-5.5 Instant."

In this study, the company utilized images of its original mascot, "Combee-chan," and analyzed the output results and specific strengths and weaknesses of each AI based on two patterns: a "simple prompt" and a "complex prompt."

Survey 1: Comparison via Broad Instructions For the prompt: "Generate an illustration of the attached character flying through the city." Both AIs successfully grasped the intention from broad instructions and produced illustrations of the character flying through the city without major issues. Both possess high practical utility for simple image generation tasks where the AI is left with some creative freedom.

Survey 2: Comparison via Complex Instructions We conducted a survey using highly challenging prompts that specified situations, touch, and composition in detail. Distinct differences in features and capabilities became apparent with detailed instructions.

Regarding overall quality and character consistency, GPT demonstrated superior performance. Gemini showed a slight tendency to struggle with reflecting all detailed instructions in the illustration (e.g., character having limbs when it shouldn't).

Additionally, distinct individualities emerged in drawing styles and texture. While Gemini produced warm illustrations with clear line art similar to a picture book, GPT excelled in expressing realistic textures, faithfully reproducing the characteristics of watercolor paintings.

Conclusion As of now, GPT demonstrates superior ability in detailed character consistency and complex composition specifications. However, AI is evolving daily, and there are high expectations for further improvements in functionality.

Survey Summary - Target: Major Generative AI (Gemini 3.5 Flash / GPT-5.5 Instant) - Period: April 17, 2026 – May 16, 2026 - Purpose: Investigation of output features based on prompt complexity.