In “Introducing Design Crit: We taught AI to judge design like a designer,” ten professional designers ranked four frontier image models across nine dimensions of real design work. The models were Nano Banana 2, FLUX.2 [max], GPT Image 1.5, and Seedream 5.0 Lite.
That post detailed CRIT and the models it tested. This article talks about how each model actually performed, and what designers preferred, criterion by criterion. This article covers the Aesthetics track, including color harmony and visual hierarchy, while the next covers Descriptions, which measures brief fidelity like color and text accuracy. On aesthetics, the short answer is that nobody dominates. Three of the four models are within a few points of each other, and the "best" model changes depending on which dimension of visual quality you care about. To see why, it helps to know how the ratings were collected.
How we measured taste
The launch post covers the full setup. This short version is that these results come from CRIT's Aesthetics cohort. Five designers, separate from the five who rated brief fidelity, judging five of the nine criteria: color harmony, mood and tone match, typography aesthetics, visual hierarchy, and an overall UI + aesthetics preference, an axis that just asks which design a viewer liked best rather than scoring one specific quality. Each criterion had its own 80 prompts drawn from real client deliverable categories, with all four models' outputs shown blind and compared pairwise giving 1,600 ratings per criterion, for a total of 8,000 ratings.
The race is tight
Pool the five aesthetic criteria together and the field compresses:
FLUX.2 (54%), Nano Banana 2 (52%), and GPT Image 1.5 (50%) finish at about the same win rate, ~52%, while Seedream 5.0 Lite trails behind at 44%.
That tightness is the real shape of the result. As the launch post showed, designers agree on good design about as much as people agree on a favorite movie, less than crowds agree on which photo is sharper, but with real shared consensus underneath. They share a rough sense of "good" with some personal variation at the top, and no group's taste runs the other way. That's exactly the kind of pattern a model can learn.
Pooled together, though, the aesthetic scores flatten out the differences. Split them by criterion and each model shows a distinct personality.
Each model has a signature
FLUX.2 [max] performs strongest on the structural criteria. It leads typography aesthetics (57%) and visual hierarchy (55%), the two that are really about how a layout is built, and both are clear separations rather than near-ties. If layout craft is what makes or breaks your output, FLUX.2 has the strongest case.
GPT Image 1.5 performs strongest on feel. It tops mood and tone match (56%) and edges the overall UI + aesthetics preference (54%, a hair above Nano's 53%), the criterion that just asks which design people liked best. Where it clearly slips is structure: typography aesthetics (45%) and visual hierarchy (48%) both land below parity.
Nano Banana 2's strength is consistency. It leads color harmony (55%) and stays close on mood and tone, typography, and overall preference, landing in the top group on four of the five criteria. Its one soft spot is visual hierarchy, where it sits at the bottom of a tight cluster (at 47%, with GPT and Seedream at 48% and 50% respectively) making it the most even performing model in the field.
Seedream 5.0 Lite's one bright spot is structure. It trails clearly on four of the five criteria, bottoming out at 39% on mood and tone, but climbs into the pack on visual hierarchy at 50%. Give it a layout problem and it holds its own; ask for polish and it falls behind.
What that means in practice
Because the aesthetic race is tight, the practical question becomes which model wins for your kind of work. Sliced by deliverable type, with every criterion pooled into one win rate per category, the winners move around:
Nano Banana 2 takes the most categories, winning logos, flyers, and Instagram posts. Its logo result (76%) is the widest margin anywhere in the study, while GPT Image 1.5's 15% in that same category is the weakest.
Seedream 5.0 Lite, last on most aesthetic criteria, wins art & design outright at 67%. Give it exploratory work rather than refined deliverables and it goes from last to first.
FLUX.2's wins in posters and social media line up with its typography and hierarchy strengths. Those are the formats where layout craft is the product.
This demonstrates the decision layer the launch post described: pick the model that's strongest where your deliverable lives. Typography for a poster, color and consistency for a logo instead of trusting one aggregate score.
Limitations
These results carry the same limitations as the launch post. Five raters per prompt, separate 80-prompt sets per criterion (so no design was ever rated on two axes at once), English-only prompts, and five aesthetic criteria that don't yet cover accessibility, brand consistency, motion, or audience fit.
What's next
The launch post tells the bigger story: how much designers actually agree, why nothing on the market can read their taste yet, and how a small model trained on CRIT already closes about half the gap to a human designer's eye. What we still don't know is whether signals like these can make the generators themselves better designers, not just easier to grade.
How we ran this study → Methodology