Every model renders the same brief, and a vision judge picks one image without knowing which model routing chose.
Does routing make better images?
All 5 models rendered each of 10 briefs, and a judge blind to routing picked one image. It was the routed model's on 7 of 10, the default's on 2, against 2 by chance. Scored alone out of 100, the routed image came top on 4 of 10 and the default on 3. Every brief is below.
Lifestyle shot of a woman jogging at sunrise wearing our new running shoes, the shoes should be the focus
Photoreal · 95%
Typecast's pickz-imagebest of 5
Overall score
69
Brief
69
Style
90
Craft
49
Defaultqwen-2.1
Overall score
61
Brief
53
Style
78
Craft
51
Judge's pick: the routed model (z-image). Highest overall score: krea2.
All 5 models, same prompt and seed. Outlined: the judge's pick. Click any image to enlarge.
A small, early sample: 10 briefs that routing was tuned on, judged by a model (Jev reading the images in every order), not yet checked against people. The other jobs noted on some briefs aren't measured at all. The lineup is open-weight models run locally, standing in for a real product's roster.