Summary for Ghibli style
Replicating the iconic Studio Ghibli aesthetic—characterized by lush watercolor backgrounds, flat cel-shaded characters, and whimsical environmental storytelling—proves to be a towering challenge for AI image generators. The dataset reveals a stark divide between models that genuinely grasp this 2D-meets-painterly balance and those that default to glossy 3D renders or generic anime.
Key Findings:
- 🏆 Top Overall Performer: Muse Image dominates the category with an impressive 1502 Elo and zero refusals, consistently delivering the warmth, breathtaking flora, and true-to-style animation aesthetics required by the prompts.
- 🥈 Strong Contenders: Takumi 1 (1475 Elo) and FLUX 3 Image (1473 Elo) follow closely, offering exceptional mechanical and natural detailing. However, FLUX 3 Image suffered from a 50% refusal rate.
- ⚠️ The Refusal Bottleneck: The highest-rated model conceptually, MAI-Image-2.6 (1490 Elo), experienced a 90% refusal rate due to safety/copyright filters, making it highly unreliable for this specific creative task.
- 📉 Common Pitfalls: Weaker models, such as Grok 2 Image (477 Elo), consistently failed by applying heavy volumetric lighting, photorealistic finishes, or aggressive creature designs that shattered the gentle, hand-drawn illusion.