A design can be gorgeous and still wrong. The brief said navy but the logo came back teal. The headline that’s supposed to sit above the product sits behind it. For working designers, this is the difference between an asset they can and can't use.
So alongside our aesthetics evaluation of Nano Banana 2, FLUX.2 [max], GPT Image 1.5, and Seedream 5.0 Lite, we ran a dedicated brief-fidelity track. And unlike aesthetics, where three models were in a near-tie, fidelity has a clear winner.
How we measured it
These results come from the Descriptions cohort of CRIT, rated by five professional designers entirely separate from the five who rated aesthetics in Part 1. The cohort covered four criteria the designers themselves identified as verifiable against the prompt: color accuracy (did the palette match the brief?), spatial accuracy (did elements land where the brief placed them?), typography from descriptions (did the specified text appear, spelled correctly, as directed?), and an overall preference given the brief.
Each criterion used its own set of 80 prompts, drawn from real client deliverable categories on Contra's marketplace. All four models generated from each prompt, presented blind behind code-names. Each designer ran all six pairwise comparisons per prompt, aggregated into strict four-way rankings for a total of 1,600 ratings per criterion.
On the two overall-preference tracks, designers also flagged every image for hallucinations, rating both major and minor elements that contradicted the brief or shouldn't exist at all.
Nano Banana 2 sweeps all four criteria
Nano Banana 2 took first place on every fidelity criterion with win rates ranging from 62% to 67%, while no other model crossed 52% on anything. Its strongest result on typography from descriptions means that when a brief specified a type treatment, designers picked Nano Banana 2’s outputs 67% of times.
The rest of the field inverts the aesthetics standings. Seedream 5.0 Lite, last on most aesthetic criteria, is the consistent runner-up here, placing second on color accuracy, spatial accuracy, and overall preference. FLUX.2 [max], the typography-aesthetics champion of Part 1, falls below parity on every fidelity criterion, including a field-worst 40% on color accuracy. GPT Image 1.5 posts the single weakest number in the track: 36% on spatial accuracy.
The aesthetics–fidelity split
Plotting each model's aesthetics win rate against its fidelity win rate sorts the four into three groups.
Nano Banana 2 is the only model above parity on both axes, and that fidelity edge makes it the overall leader of the study, beating FLUX.2 and GPT Image 1.5 in 57% of all head-to-head matchups and Seedream in 59%.
FLUX.2 and GPT Image 1.5 sit in the lower-right, where designers like how their work looks more than how well it listens. Seedream 5.0 Lite runs the other way, scoring higher on brief fidelity than on aesthetics, though it remains below parity on both.
Hallucinations: a different leaderboard again
Fidelity leadership and reliability aren't the same thing:
Three models produce a clean image at the same 57% rate, but they differ in how badly the other 43% goes wrong. FLUX.2 [max], which ranked last in fidelity, also has the lowest major-hallucination rate at 7%. Nano Banana 2, the fidelity champion, majors at 12%, near GPT Image 1.5's field-worst 13%. GPT Image 1.5 is the outlier overall: it hallucinates on more than half of its generations.
Following the brief and not hallucinating are separate abilities with Nano Banana 2 being best at executing what you asked for, but when it misses, it's somewhat more likely to miss big. FLUX.2 drifts from the brief in small ways constantly but rarely fabricates something severe.
Limitations
The sample is small. Each prompt was rated by five designers, which is enough to measure agreement and rule out noise but not enough to be sure of any single comparison. Each criterion used its own set of 80 prompts, so no design was ever rated on two criteria at once. That kept each rating clean, but it means we can't see how one designer weighs fidelity against feel on the same design, because no one judged the same design on both. Every prompt was in English, so behavior across languages isn't captured here. The nine criteria cover a lot but not everything. Accessibility, brand consistency, motion, and audience fit are all natural axes to add as the work grows.
How we ran this study → Methodology