Artificial Analysis just dropped a leaderboard that actually means something for image generation. Their Text to Image Arena has collected over 45,000 human preference votes, pitting the biggest names in image models against each other in direct head-to-head matchups. Midjourney’s latest, OpenAI’s DALL·E, Stable Diffusion, Playground, and a mix of open-source contenders are all in the ring. The scoring uses an ELO system, the same approach that made Chatbot Arena the gold standard for comparing language models.
The mechanics are straightforward. You see a prompt and two images. You pick the one that better matches what was asked. No metrics, no FID scores, no academic benchmarks — just people deciding what actually looks good. Artificial Analysis generates more than 700 images per model across a deliberately broad prompt set: human portraits, groups, animals, nature, art, and beyond. That breadth matters, because image preference is wildly subjective in ways text evaluation isn’t. One person’s photorealistic masterpiece is another’s uncanny-valley nightmare.
What’s notable here is the scale of the effort. Running human preference studies for image models has historically been expensive and slow, which pushed the industry toward objective metrics that never quite captured what people actually like. A crowdsourced arena sidesteps that bottleneck. And if you’re willing to put in the votes — 30 of them — you get a personalized leaderboard ranking models according to your own aesthetic biases. That’s a genuinely useful feature for anyone trying to figure out which tool fits their workflow.
The timing feels right, too. Text-to-image quality has improved so dramatically over the past 18 months that the gap between top proprietary models and strong open-source releases has narrowed considerably. A leaderboard that includes both camps gives you something most rankings don’t: a realistic picture of whether you actually need to pay for Midjourney or DALL·E, or whether a self-hosted Stable Diffusion variant gets you close enough. Artificial Analysis already tracks speed and pricing for image API endpoints on their main site, so pairing that cost data with these preference rankings makes the comparison genuinely actionable rather than just academic.
