Most AI image generators are diffusion models — they start from random noise and refine it step-by-step into a coherent image, guided by a text encoder that converts your prompt into a vector the model can follow. Newer transformer-based generators (the GPT-Image family) produce images token by token similar to how language models produce text.
Rankings use TrueSkill (conservative rating: μ − 3σ) from blind human comparisons in the Image Arena. Each prompt generates 4 images from randomly sampled models. Users see the images side by side and pick the best and worst — without seeing model names, providers, or any identifying information. This eliminates brand bias and ensures rankings reflect actual image quality.
The leaderboard covers two distinct arenas. The text-to-image arena evaluates how well models generate images from text descriptions — including photorealism, illustration styles, concept art, and typography rendering. The image editing arena evaluates how well models modify existing images based on instructions, testing understanding of spatial relationships, style transfer, and selective editing.
Pricing data is pulled from provider APIs and shown per image. Costs range widely — from under $0.01 per image for open-weight or lightweight AI image generators to $0.10+ for frontier models. The scatter view in the rankings tab lets you compare quality vs. cost directly, which is useful if you need the best AI image generator for scale: product mockups, marketing assets, or creative workflows.