Generating impressive images is no longer the bottleneck for AI image models. The bottleneck is producing images a professional creative would actually approve for client work. Visual quality and commercial production readiness are different standards, and current models still struggle to meet the second one consistently.
GenPeach v1 Preview is a new image model built to close that gap for human-centered commercial and editorial photography, where skin, anatomy, styling and art direction all have to hold up under professional review.
Contra Labs ran a blind evaluation to test how GenPeach v1 Preview performs against that standard. 53 professional creatives compared GenPeach’s images head-to-head with outputs from eight widely used image models across 100 professional shoot briefs. The results show where GenPeach leads, where other models came close, and how far the gap between visual preference and production readiness still extends.
Methodology
53 working creatives sourced from Contra’s top talent, including brand designers, art directors, photographers and creative technologists, evaluated GenPeach against eight image models on 100 editorial and fashion photography prompts. Evaluators had a median of 6 years of professional experience, and 17 of them had 10 years or more.
The prompts were sourced by GenPeach from professional creatives working across advertising, fashion and entertainment. Prompts were finalized before any image generation and were not modified based on model outputs. They covered beauty close-ups, fashion editorials, lifestyle photos, passport-style headshots, black-and-white portraits, multi-person scenes and avant-garde concepts, with some prompts spanning multiple categories. Each was written as a professional shoot brief, specifying the subject, wardrobe, lighting, composition and framing.
The evaluation included GenPeach v1 Preview, GPT-Image-2.5 Sunburst, Krea 2 Medium, Ideogram 4.5, Gemini 3.1 Flash (Nano Banana 2), Muse Image, Flux 3, Luma Uni-1 and Recraft V4.1. The competing models were selected based on their rankings across public leaderboards and relevance to professional creative workflows based on GenPeach’s interviews with professionals working across advertising, fashion, media and entertainment, and photography.
All nine models were evaluated under the same generation protocol, with exactly one image per prompt and no retries. Contra Labs conducted the blind evaluation using outputs from a fixed GenPeach model checkpoint, provided by GenPeach, and outputs generated through the other eight models' public APIs.
GenPeach served as the anchor model. Each matchup paired GenPeach’s image with another model’s image for the same prompt, shown side by side with model names hidden and left/right placement randomized. Evaluators selected a winner on four dimensions before indicating their overall preference:
- Human realism: Does the person look real?
- Skin and material: Do the subject’s skin, hair and fabric look convincing?
- Aesthetic and art direction: Do the light, composition and styling work?
- Brief adherence: Does the image deliver what the prompt asked for?
Along with the overall preference, evaluators had to provide qualitative feedback for each image, highlighting its strengths and weaknesses.
Experts then answered two yes-or-no questions about every image, each followed by a written reason:
- Would you deliver this image to a client as-is?
- Does this image look AI-generated?
Each matchup was rated by five different experts. In total, the study produced 3,930 head-to-head votes across 786 matchups and 4,430 single-image ratings, and experts spent about 27 minutes per prompt.
Results
Across 3,930 blind head-to-head votes, GenPeach v1 Preview was preferred in 71% of comparisons against eight competing models, leading across all four evaluation dimensions. GenPeach won the majority vote on 88 of the 100 prompts, and its widest margins were against Krea 2 Medium, Ideogram 4.5 and Flux 3, at 76% each. Nano Banana 2 came closest, winning 37% of its votes against GenPeach, followed by Muse Image (35%), Recraft V4.1 (34%) and GPT-Image-2.5 Sunburst (31%). GenPeach’s strongest advantage was in aesthetics and art direction, while competition was closer in human realism and brief adherence.
The following sections examine each evaluation dimensions and explore what drives professional creatives' preferences.
Aesthetics & Art Direction: GenPeach out front
Aesthetic and art direction covered how the light falls, how the frame is composed and how the subject is styled. It also assessed whether the image looks like the work of an art director rather than a merely competent generation.
This was where GenPeach's lead was widest and most consistent. No model won more than 38% of its votes against it, and a majority of the five experts preferred GenPeach in 77% of matchups. Muse Image, Recraft V4.1 and GPT-Image-2.5 Sunburst came closest, at 36% to 38%. GenPeach's aesthetic advantage was not attributed to a single dominant feature. Evaluators cited composition, lighting, color and depth, without one factor consistently standing out.
Aesthetics was also the dimension most closely tied to the overall vote. When experts split their picks across dimensions, their overall choice matched their aesthetic pick slightly more often than any other.
Human Realism: A closer race at the top
Human realism asked whether the person looked real: believable proportions, anatomy, faces, poses and expressions. This was the closest dimension at the top. GenPeach showed the strongest results and won 65% of votes on this dimension. Nano Banana 2 nearly matched GenPeach, winning just under half of its realism votes, and the rest of the field won roughly a third of their votes against GenPeach, with Krea 2 Medium lowest at 23%. On convincing people, two models stood apart from the rest – GenPeach v1 Preview and Nano Banana 2.
Skin, Material & Brief Adherence: Different strengths across the field
The remaining dimensions produced a slightly different competitive order. On skin and material rendering, GenPeach won 64% of votes. Nano Banana 2, Recraft V4.1 and Luma Uni-1 came closest, each winning 42% to 45% of their votes against GenPeach.
Brief adherence produced the closest comparisons. GenPeach led at 64%, with Muse Image (46%) and GPT-Image-2.5 Sunburst (44%) coming nearest to parity. Yet brief adherence was the dimension least closely associated with overall preference. Following the brief gets the requested details right. It doesn't necessarily make the result the preferred image. In one comparison against Recraft V4.1, an evaluator favored Recraft on realism, skin and brief adherence, but selected GenPeach on art direction and overall.
Overall Preference, Client Readiness & AI Look: The tougher test
Performance on individual dimensions, however, only tells part of the story. For professional creative work, two further questions matter: whether an image is ready to deliver to a client without further edits, and whether it looks convincingly photographic rather than AI-generated. The first measures immediate usability; the second captures perceived authenticity, an important consideration in professional visual production.
The dimension that best predicted the overall preference depended on whether the images looked real or AI-generated. In 52% of votes, experts judged both images to look AI-generated. In these votes, human realism and art direction predicted the overall pick at similar rates: the overall choice agreed with the realism choice in 84% of votes and with the aesthetic choice in 83%.
In votes where both images looked real, the pattern shifted. The overall choice agreed with the aesthetic choice in 86% of votes and with the human realism choice in 78%. Once realism no longer separated the two images, art direction became the strongest predictor of preference. The matchup with Nano Banana 2 illustrates this pattern. Nearly level with GenPeach on the human realism axis but behind on art direction, it lost the overall vote 37% to 63%.
Of the four dimensions, brief adherence was the least predictive of which image experts preferred overall. Even in matchups where experts preferred the other image on the brief, they chose GenPeach overall a third of the time. Following the brief worked as a baseline requirement rather than the deciding factor.
Winning a comparison and being ready for a client are different standards. Across the study, the preferred image in a matchup was one the expert would deliver as is only 53% of the time.
Delivery rates varied widely across models. GenPeach led the client-readiness measure at 49%, more than double the 23% average for the other eight models. Nano Banana 2 followed at 34%, Muse Image at 29%, and the remaining models ranged from 17% to 24%. When experts preferred GenPeach's image, they rated it deliverable 58% of the time. When they preferred the other model's image, they rated that image deliverable 43% of the time.
Perceived photographic authenticity was strongly associated with the delivery decision. Evaluators would deliver 70% of images they judged not to look AI-generated, compared with 22% of images identified as AI-generated.
But realistic image appearance was not sufficient. Even among images judged not to look AI-generated, 41% were considered unsuitable for delivery without changes. Photographic authenticity gets an image closer to production, but composition, anatomy and brief execution can still keep it from being usable.
GenPeach had the highest share of images judged not to look AI-generated, at 40%, followed by Nano Banana 2 (34%) and Luma Uni-1 (30%). Brief adherence and delivery didn’t always line up. Muse Image and GPT-Image-2.5 Sunburst stayed close to their briefs, but evaluators often flagged their images as looking generated.
Bottom Line
These findings suggest that the models face different challenges in achieving production readiness: some images fail because they appear visibly synthetic, while others look convincing overall but contain specific errors that prevent delivery.
Technical accuracy wasn't always the strongest indicator of overall preference. When both images looked photographic, art direction was more closely aligned with the final choice.
For human-centered editorial and fashion photography, GenPeach led on overall preference, art direction and client readiness. Nano Banana 2 was the closest alternative on human realism, while Muse Image and GPT-Image-2.5 Sunburst came nearest on brief adherence. The choice depends on whether the priority is photographic credibility, precise execution of the brief or an output ready for client delivery.
The gap between visual preference and production readiness remains substantial across the field. Even the highest-performing model in this evaluation produced images experts would deliver as-is only about half the time.
Limitations
Some models declined to generate images for certain prompts. Muse Image declined five, four of them involving smoking. GPT-Image-2.5 Sunburst and Krea 2 Medium declined three each, Flux 3 two and Nano Banana 2 one. Refusals were not counted as losses and did not negatively affect the models' reported scores.
Disclaimer
This study was independently conducted by Contra Labs and commissioned by GenPeach. Because every matchup paired GenPeach with one other model, the results describe GenPeach’s performance relative to each competitor and should not be read as a ranking of the full field.
How we ran this study → Methodology