How this is ranked: these are
ELO scores from blind human preference — people are shown two outputs
from the same prompt, without knowing which model produced them, and pick the one they prefer. It is the only
honest way to rank media models: no automated benchmark can tell you whether an image is good.
A higher ELO means people preferred it more often. The
votes column shows how many comparisons back the score —
a model with few votes has a less reliable position.
Open-weight means the model can be downloaded and run on your own hardware. The ranking API does not say
whether a model is open, so we identify them by model family. Where we cannot confirm it either way, we mark the
model
unverified rather than guess — and the "best open-weight" figure above only counts models we can
actually confirm. We would rather understate open-source than mislead you.
Spotted a mistake?
Rankings by
Artificial Analysis.