Evaluating generative models is shifting from a "one-size-fits-all" approach to a more nuanced, domain-specific strategy. The Arena Team has unveiled a significant update to the Text-to-Image Arena, moving beyond a single global ranking to introduce seven distinct category leaderboards. By analyzing over 4 million user prompts, the team identified that model performance fluctuates significantly depending on the intent, such as 3D Imaging & Modeling or precise Text Rendering.

This granular approach reveals fascinating insights into current foundation models. For instance, while high-profile models like GPT-image-1.5 lead overall, the Nano-banana-pro model demonstrates superior performance specifically in 3D construction. Meanwhile, Qwen-image-2512 punches above its weight in human portraits despite having a lower general ranking. These findings highlight the importance of choosing specific tools for specialized creative tasks rather than relying on a single general score.

To further refine the data, the Arena now employs a Large Language Model based filter to scrub "noise"—low-quality prompts like accidental resume pastes or video instructions that the system cannot fulfill. By removing approximately 15% of these outliers, the leaderboard achieves higher statistical reliability, ensuring that rankings reflect actual text-to-image capabilities. This update provides a more transparent and dependable framework for evaluating the rapidly evolving state of the art in AI imagery.

Evaluating generative models is shifting from a "one-size-fits-all" approach to a more nuanced, domain-specific strategy. The Arena Team has unveiled a significant update to the Text-to-Image Arena, moving beyond a single global ranking to introduce seven distinct category leaderboards. By analyzing over 4 million user prompts, the team identified that model performance fluctuates significantly depending on the intent, such as 3D Imaging & Modeling or precise Text Rendering.

This granular approach reveals fascinating insights into current foundation models. For instance, while high-profile models like GPT-image-1.5 lead overall, the Nano-banana-pro model demonstrates superior performance specifically in 3D construction. Meanwhile, Qwen-image-2512 punches above its weight in human portraits despite having a lower general ranking. These findings highlight the importance of choosing specific tools for specialized creative tasks rather than relying on a single general score.

To further refine the data, the Arena now employs a Large Language Model based filter to scrub "noise"—low-quality prompts like accidental resume pastes or video instructions that the system cannot fulfill. By removing approximately 15% of these outliers, the leaderboard achieves higher statistical reliability, ensuring that rankings reflect actual text-to-image capabilities. This update provides a more transparent and dependable framework for evaluating the rapidly evolving state of the art in AI imagery.