Meta Muse is the most dependable style-transfer model we tested. In a blind tournament across 10 real-world style-transfer briefs, Muse finished last in just 14% of matchup rankings, tied for the fewest in the field, and it was the only model to finish top-two on both tournament wins (29.5%) and Elo (1,552.6). ChatGPT Images 2.0 took first place most often, but its results swing brief to brief.
We ran a blind, head-to-head evaluation of four current AI image models on the capability Meta put front and center in Muse's launch: style transfer. Take a real photo (a person, a pet, a landscape) and re-render it in a named style: a dreamy watercolor postcard, a classical oil painting, a flat illustration, a 16-bit game character. The four contenders were Meta Muse, ChatGPT Images 2.0, Nano Banana Pro, and Flux 2.
Muse is the steady performer
Across all 88 tournaments, ChatGPT Images 2.0 took first place most often at about 34% of time (30/88), with Meta Muse close behind at 29.5% (26/88), ahead of Nano Banana Pro (23.9%) and Flux 2 (19.3%).
But first-place counts hide the shape of each model's results. Muse's rankings cluster at the top: it finished first or second in most of its tournaments and came in last just 14% of the time, the lowest rate in the field alongside Nano Banana Pro. ChatGPT Images 2.0 tells the opposite story. When it wins, it wins outright, but when it doesn't take first, it usually falls to the bottom half of the ranking. Muse almost never bombs.
The Elo ratings sharpen the picture. Nano Banana Pro leads at 1,562.3 with Muse just behind with a gap of about 10 points, while ChatGPT Images 2.0, despite winning the most tournaments, sits roughly 80 points back at 1,473.6, weighed down by its bottom-half finishes.
Where Muse won: graphic, stylized briefs
Muse's wins concentrated in briefs where the target style was bold and graphic. It won a brief asking it to transform an image into a plushie with a 62.5% win rate and another making it into a watercolor postcard at 55.6%.
When reviewers picked Muse, the two most common reasons were aesthetic and style execution and composition, followed by visual quality and fidelity. On a prompt asking for a flat illustration, which Flux 2 ultimately took, reviewers who chose Muse described exactly the kind of confident stylization the prompt asked for.
Model G [Muse] best captures the requested flat illustration style through bold geometric shapes, simplified forms, and a clean, cohesive composition. The color palette is vibrant yet balanced, and the soft painterly transitions along the edges closely match the prompt.
I preferred G [Muse] because the playfulness really came through in the use of the black 'handdrawn' strokes for windows/doors as well as the 'handdrawn' red loops for the roofing texture. The colors and shapes were also bold.
Where Muse fell short: material realism, by a narrow margin
Even where Muse lost, reviewers rarely dismissed it. On a brief asking to reimagine a pet as a classical pet oil painting, where ChatGPT Images 2.0 had the most decisive win at 75%, reviewers repeatedly credited Muse's fundamentals before choosing a rival on a single dimension: whether the image looked like actual paint.
"Model G [Muse] has the best lighting," wrote a reviewer who ultimately picked ChatGPT Images 2.0 because "many of the models did not capture the paint-strokes of an oil painting."
"Model G [Muse] also captures the classical style well but has softer lighting and less pronounced brush texture," wrote a reviewer who chose Nano Banana Pro for its "rich impasto brushwork, dramatic side lighting, and strong chiaroscuro contrast."
The critique data shows the same picture in aggregate. When reviewers picked another model over Muse, the top critique categories were composition (9 comments) and aesthetic/style (7), the same dimensions Muse gets praised on, meaning most losses were judgment calls between two strong outputs. The recurring differentiator was texture and materiality: when the target style is itself a physical medium (oil paint, engraved linework), reviewers chose the model that rendered the material most convincingly, and that was usually ChatGPT Images 2.0 or Nano Banana Pro.
The bottom line
All four models won at least one brief outright, and each showed a distinct shape. ChatGPT Images 2.0 swings hardest: it took first place more than any other model and spent most of its remaining tournaments in the bottom half of the ranking. Meta Muse is the steady one. It won nearly as often, ranked second on Elo, and finished last as rarely as any model in the field. Reviewers who picked Muse praised its aesthetics, composition, and cohesion. Reviewers who chose a rival usually decided on material realism: briefs that hinged on convincing physical texture went to ChatGPT or Nano Banana Pro, often with compliments for Muse's lighting and style along the way.
For teams choosing a style-transfer default: Muse is the pick that will almost never embarrass you and wins nearly as often as the leader, ChatGPT Images 2.0 is the pick most likely to produce the single image a reviewer loves, and the choice between them comes down to whether your brief is built on graphic style or physical material.
Methodology
Nine creative professionals sourced from Contra's top-earning talent evaluated ten style-transfer briefs, each pairing a source image with a target style: flat illustration, 16-bit game character, child's crayon portrait, Renaissance portrait, watercolor postcard, classical pet oil painting, 19th-century botanical engraving, 1980s action figure, anime, and plushie. Each brief specified the stylistic requirements in professional detail (medium, texture, proportions, lighting, and compositional elements), forcing each model to execute a specific style rather than a generic aesthetic. Every brief was run through all four models:
- Meta Muse
- ChatGPT Images 2.0
- Nano Banana Pro
- Flux 2
The four outputs per brief were shown to reviewers blind: model identities were masked behind letters and order was randomized, so brand reputation couldn't tilt the results. Reviewers completed a round-robin pairwise preference (six matchups per brief), which forces a clean first-to-fourth ranking we treat as one tournament, then wrote an open-ended explanation of why they preferred the winner, with instructions to be specific and reference models by letter. From the pairwise outcomes we computed Bradley-Terry Elo ratings, which weigh every matchup rather than just first-place finishes.
Limitations
Reviewers judged single-pass generations, so the results reflect first-attempt capability, not each model's ceiling under iteration or prompt refinement. The study covered one task category, style transfer from a source image, and says nothing about text rendering, detailed scene construction, or pure text-to-image generation. Preference judgments carry inherent subjectivity; the blind, randomized presentation removes brand bias but not taste. With 10 briefs and nine reviewers, the sample is small enough that we treat these findings as directional rather than definitive, and we place more weight on patterns across briefs than on any single result.
While ChatGPT Images 2.0, Nano Banana Pro, and Flux 2 were generated via API, Meta Muse outputs came from the Meta.ai web app and were uploaded after the fact, so differences in the generation harness may have influenced Muse's results in either direction.
With 10 briefs covering 10 distinct styles, each style was tested by only a single prompt, so a model's result on any one brief reflects its performance on that specific image and prompt, not its general capability in that style.
How we ran this study → Methodology