Analysis of Relevance and Appeal of Visual Presentation of Meat Products Generated Using Artificial Intelligence

Abstract

This article examines the application of generative artificial intelligence (GenAI) in visualizing meat products and evaluates its potential for use in the food industry. The study compares AI-generated images with conventional photographs in terms of professional accuracy and visual appeal. In a cited preliminary study, images of ten selected meat dishes were generated and evaluated by food technology professionals through a survey focused on realism and technical adequacy. Following this, comparable photographs were taken, and a second survey gathered feedback from the public on the appeal of both image types. Results revealed that while AI-generated images often lacked accuracy in texture, color, and structure, particularly for complex meat products, they were generally rated as more visually appealing by the public. This indicates that although current GenAI tools are not yet suitable for precise professional representation of meat products, they show strong potential for use in marketing and promotional content, where aesthetic appeal may outweigh technical accuracy. The findings suggest that with further development, AI-generated visuals could become more viable for professional applications in the food industry. In such cases, using accurate photographic references remains essential to ensure credibility and realism in food-related visual communication.

1. Introduction

Meat and meat products are vital components of a balanced daily diet. In current times, consumers frequently choose these items based on their visual appeal. As essential elements of human nutrition, meat, meat products, and meat-containing dishes pose a challenge in terms of presentation, since sensory perception significantly influences an individual’s food purchase decisions. Consequently, improved sales of meat and meat products result in greater financial benefits for all stakeholders involved in their production and marketing. In this article, we will evaluate the effectiveness and attractiveness of presenting meat dishes or products developed through artificial intelligence.

This article investigates the capabilities of GenAI to depict meat and meat products, comparing AI-generated images with traditional food photography in terms of professional accuracy and consumer appeal. In a preliminary study, we researched “How accurately can current GenAI render various meat products’ color, texture, and characteristic features when evaluated by food technology professionals?”, with the following hypothesis: “AI-generated images will fall short of professional photography in an accurate depiction of meat products.” The main results are cited in this research, which mainly addresses a new research question “To what extent do consumers prefer the look of AI-generated images to conventional photographs?” with the following hypothesis: “Consumers will find AI-generated images more attractive than real photographs despite accuracy limits.”

This article highlights the crucial role of visual presentation in food marketing, particularly for meat products. It examines how advancements in GenAI (Adobe Firefly 3) can equip producers and advertisers with innovative imaging tools. The overall purpose is to evaluate whether GenAI creates images that are at least as appealing to consumers as camera-taken photographs.

This study contributes to the literature on GenAI in food communication in the following ways:

-

Systematically comparing AI-generated and photographic images of meat products using both professional and consumer audiences;

-

Introducing a structured two-phase evaluation involving Adobe Firefly 3;

-

Highlighting the impact of reference images on output quality and public preference;

-

Providing a concrete hybrid workflow for applying AI in food marketing contexts;

-

Offering novel insight into demographic differences and trust in AI-generated food visuals.

This dual focus on expert judgment and consumer preference for a sensitive product category, such as meat, offers a unique perspective on the practical and perceptual challenges of using GenAI in food presentation.

2. Literature Review

Artificial intelligence (AI) is increasingly integrated into various aspects of our daily lives. One of the fastest-growing domains of AI is GenAI, which enables users to quickly create images, videos, code, and more through text. Correspondingly, the number of studies focusing primarily on differences in graphic perception between classical photography and artificially created images is rising [1]. Sensory perception of food products encompasses their visual presentation on plates, during preparation, and when displayed on shelves for customers. Consumers often rely heavily on a product’s appearance during purchase, even in the absence of prior information, making appearance extremely important. With raw products, the appeal of color is debatable because additives are generally undesirable, yet they often enhance attractiveness by stabilizing or improving color [2].

2.1. Artificial Intelligence

In 1955, John McCarthy provided one of the earliest definitions of AI, stating “The goal of artificial intelligence is to develop devices that behave as if they are intelligent.” Although this definition is not the most precise, it offers insight into AI’s long history of development. A more accurate definition by Elaine Rich states that “Artificial intelligence is the study of how to train devices to do tasks in which humans currently excel” [3]. Early intelligent systems depended on clear rules to aid decision-making, including knowledge bases and expert systems. AI encompasses a wide range of computer algorithms designed to perform tasks that typically require human intelligence, such as pattern recognition, language understanding, decision-making, and experiential learning [4].

2.2. Generative Artificial Intelligence

The main distinction of GenAI compared to earlier AI systems is its ability to collect data and produce new content—text, images, videos, audio, or code. This functionality resembles human creativity but accomplishes tasks faster and more economically [5]. AI systems often rely on machine learning, a subset of algorithms that can learn from data. Initially, these systems are exposed to training data, which enables them to automatically build analytical models for specific problem-solving tasks without explicit programming [5].

GenAI models use deep learning, a subset of machine learning, that utilizes artificial neural networks that are modeled after the structure of the human brain [6]. These interconnected neural networks effectively represent correlations and patterns across large datasets, making them suitable for various data types, including text, images, videos, or audio [4]. GenAI models, such as image-generating AI, require extensive datasets, like a variety of photographs, to train models capable of producing new images based on appropriate text prompts [7].

Since OpenAI introduced ChatGPT3 in 2022, the use of GenAI has surged, fundamentally changing digital platforms and creating new opportunities while also raising concerns about potential job replacements [7]. One of the newest and advanced freely available AI tools is DeepSeek, which differs from other AI tools in that it also understands the context of search inputs and does not just search by keywords [8].

2.2.1. GenAI Tools

Unlike earlier AI systems focused on pattern detection and prediction, GenAI utilizes vast datasets to craft new content (text, images, video, audio, and code) with human-like creativity at far greater speed and lower cost. Since OpenAI’s public release of ChatGPT-3 in 2022, these models have experienced significant growth, reshaping digital interactions and sparking new development avenues, even as they raise concerns over job displacement [7]. Tools like DALL-E, Stable Diffusion XL, OpenArt.ai SDXL, and Adobe Firefly let anyone visualize ideas without design expertise [4], yet they still falter on complex, multi-object compositions driven solely by text prompts [9].

In one study, N. Angelova found that Adobe Firefly 2 outperformed ChatGPT, DALL-E 3, Stable Diffusion, and OpenArt.ai SDXL in rendering realistic interpretations of descriptive prompts—and its successor, Firefly 3, further enhances control over size, composition, color, lighting, and perspective, drawing exclusively on licensed Adobe Stock and public-domain sources [6,10]. Building on this, Owen and Roberts (2024) introduced VisAlchemy, a structured five-step framework powered by SDXL Turbo that employs over 300 visualization-and-art-theory terms to guide designers from “Defining the Task” through “Generating Images,” enabling rapid, data-driven ideation without manual mock-ups [11].

2.2.2. GenAI in the Food Industry

With advances in GenAI, the food industry now uses AI tools for recipe development, product reformulation, and customer-preference analysis—yet consumer expectations for polished food imagery remain high, often set by polished advertising that can mislead consumers [12]. To address this, Güleç and Özkaya (2024) [13] asked 31 professional stylists and photographers to rate Firefly 3 and DALL-E 3 images against real photos of eight Turkish dishes in terms of lighting, color harmony, composition, presentation, background suitability, and “mouth-watering” appeal. They found the images generated by Firefly 3 to be indistinguishable from real photos across all dishes and superior to DALL-E 3 in five out of eight cases. Their subsequent SWOT analysis highlighted that both tools offer low-cost, rapid, and flexible production workflows, yet they sometimes yield surreal or overly “perfect” compositions, raising ethical and copyright considerations [13].

Diel et al. explored the concept of an “uncanny valley” for food images by comparing real photos, cartoonish AI renderings, and hyper-realistic yet imperfect AI outputs in 2025. Participants rated pleasantness and uncanniness, with higher food neophobia and disgust sensitivity amplifying eerie reactions, especially for images that are nearly, but not entirely, realistic. This suggests that novelty aversion underlies these responses [14]. Finally, Slerp’s experiment with food photographer David Robson had DALL-E recreate ten dishes from Eataly London. Of 100 public viewers, over 60% could not reliably distinguish AI from a photo; 73% misidentified the pizza, and two-thirds failed on the croissant and cacio e pepe, demonstrating AI’s ability to produce near-photographic, mouth-watering menu imagery quickly and cost-effectively, allowing professionals to focus on styling over logistics [15].

2.2.3. Credibility of AI-Generated Images

Most current research in GenAI focuses on the appeal rather than the accuracy of AI-generated images. Studies also examine how the awareness that an image is AI-generated affects product appeal. Additionally, extensive research has investigated whether people can distinguish between AI-generated images and traditional photographs. Findings from two studies by G. Califano and C. Spence indicate that artificially created images can be recognized, particularly when depicting processed food products. However, individuals initially preferred AI-generated images over conventional photographs. After revealing which photographs were real, preferences shifted significantly toward authentic images, demonstrating consumer distrust of GenAI [16].

Another study in the food industry found that people’s trust in AI partially correlates with age. This research evaluated trust among different age groups and concluded that the so-called “baby boomer generation” (born between 1946 and 1964) was exceptionally skeptical. Such studies provide insights into consumer psychology in food marketing [17].

2.2.4. Ethical Aspects of GenAI Usage

Users of GenAI must also consider ethical aspects, especially since AI is replacing specific jobs by performing tasks more quickly and cheaply than humans. It is essential to examine the data sources used for AI training, addressing moral concerns about the balance between technological advancement and preserving human values, privacy, and fairness. These issues remain underexplored, necessitating further research to ensure AI models align with ethical standards [18]. To ensure more controlled and effective AI development, various methods of human oversight are employed, including crowdsourcing, human-in-the-loop, expert-in-the-loop approaches, or hybrids of these methods [18].

The use of GenAI for creating deepfakes in the graphic fields raises significant ethical concerns. Such usage violates fundamental human rights by falsifying individuals’ appearances in images or videos, potentially for intimidation, ridicule, or misinformation purposes, thus instilling fear and mistrust toward AI [18]. Organizations in the public and private sectors, such as UNESCO, IEEE, the European Union, and Google, have developed and published ethical guidelines to prevent such abuses of AI [18].

2.3. Meat as a Sales Object

Meat plays a crucial role in human nutrition, and consumers rely significantly on visual indicators, such as color, shape, contrast, and symmetry, to evaluate quality before making a purchase [19,20]. In early childhood, people learn to connect specific colors with food. Consequently, color becomes the strongest signal for how fresh and safe something is to eat [21]. When foods exhibit unnatural colors, like the blue steak and green fries causing nausea, as observed in studies in the 1970s, it highlights our instinctual distaste, emphasizing the importance of appearance in appetite and trust [22]. To align with these expectations, processors use approved additives to enhance the appearance of meat products while preserving their microbiological properties. For instance, nitrite salts stabilize the characteristic color of meat, contribute to aroma development, and slow down oxidative rancidity, ensuring products showcase the familiar colors that denote quality for consumers [19].

In professional food photography, achieving the perfect hues and textures requires more than just basic camera skills. Photographers must thoughtfully compose images, control lighting to enhance marbling and surface gloss, and manage color in-camera (via white balance and exposure) as well as in post-processing (through adjustments in contrast and saturation) to create appetizing and realistic visuals [23]. Particularly for meat, it is crucial to accurately represent characteristic recognition and meet quality standards, such as cross-section appearance, juiciness, and color, to ensure that consumer expectations align with actual products. This visual accuracy helps influence buying choices and narrows the divide between the polished image in advertisements and the reality on the plate [24].

3. Materials and Methods

This research comprised two linked studies. In a preliminary study, we used Adobe Firefly 3 to generate images of ten cooked and cured meat products. We obtained ratings from professional food-technology experts on the visual adequacy of their products [25]. Building on those results, the main study expanded the image set, produced matching real-product photographs, and then evaluated the public’s aesthetic preferences in a controlled comparison between AI-generated images and photographs. The goal was to assess the appeal of AI-generated images to the general public. Since different GenAI tools employed varying designs and technical approaches and were trained with other models, the findings of this study were applicable solely to Adobe Firefly 3 and did not universally apply to all GenAI programs.

3.1. Creating Images with Adobe Firefly 3

In the preliminary study, we utilized Adobe Firefly 3 to generate images of meat products from textual prompts. Adobe Firefly 3 is a commercial generative image model based on a latent diffusion architecture, likely derived from foundational models such as Stable Diffusion. While Adobe does not publicly disclose specific architectural details—such as the number of transformer layers, parameter count, or exact finetuning strategy—the platform is known to use a text-to-image pipeline trained exclusively on licensed Adobe Stock and public domain content. It supports image conditioning and offers controls for composition, lighting, and object placement, which make it particularly suitable for food imagery. However, due to the closed-source nature of the tool, we are unable to verify whether the model has been explicitly finetuned on culinary datasets or whether its performance stems from general domain pretraining.

The choice of Adobe Firefly 3 for image generation was based not only on image quality but also on practical factors such as licensing terms, accessibility, and workflow reproducibility. Unlike many open-source models, Firefly 3 is integrated into a commercial platform with clear usage rights, which simplifies legal compliance in a commercial or academic setting. Its browser-based accessibility also allows for consistent outputs without the need for local hardware configuration, ensuring greater reproducibility across users and systems.

While generating images with Adobe Firefly 3, particular attention was paid to using correct terminology in both languages to achieve accurate representations. The prompts were assessed and refined by a qualified expert in the field of food technology to achieve optimal results. Selected meat products included a burger, Wiener schnitzel, hot dog, Carniolan sausage, mortadella, chicken skewers, prosciutto, steak, dry salami, and steak tartare.

Example of Creating Images with Adobe Firefly 3

During image generation in January 2025, we found that short, direct prompts yielded the best results, without unnecessary descriptive elements. The initial English prompt was “Prepared meat product on a white plate on a white background”, while the Slovenian equivalent read “Mesni izdelek, pripravljen za serviranje, postavljen na okrogel bel krožnik na beli podlagi”. These were adjusted as needed based on the outcome for each product. English proved ineffective for generating an accurate image of Carniolan sausage (as shown in Figure 1), and the initial Slovenian prompt was similarly unsuccessful. After several failed attempts, we refined the prompt by including specific visual attributes, assuming the model lacked sufficient training data for this traditional item. The prompt that ultimately yielded the closest approximation of the desired product is displayed in Figure 2: “Carniolan sausage (longer, rounded, tapered) on a plate with sauerkraut, white background, white plate”. A similar issue with generation arose with mortadella, which, despite being more widely known, was initially rendered too similarly to ham and lacked the characteristic white fat inclusions. After multiple iterations, a version presented on bread more closely resembled the intended product, though it still lacked full professional adequacy.

3.2. Survey Evaluating the Accuracy of AI-Generated Images

In the preliminary research, we presented ten AI-generated images of meat products to 22 food-technology professionals via an online survey. These experts, all with normal or corrected vision, evaluated the visual adequacy of each image based on specific criteria: correct product color, characteristic appearance, realistic texture of the cross-section, appropriate placement, and overall presentation on a plate. Each image was rated as either adequate or inadequate. When an image was deemed inadequate, participants selected one or more predefined categories such as inadequate product color, incorrect shape or geometry, unrecognizable or incorrect cross-section composition, and inappropriate presentation or composition of the image. They also had an option to provide open-text comments. This binary evaluation system, combined with qualitative feedback, offered both quantitative data on adequacy proportions and expert insights into the limitations of AI-generated food images. Demographic information collected included participants’ gender, age, and visual status. Those with uncorrected visual impairments were disqualified to ensure reliable visual assessments. The findings identified which sensory characteristics were most critical to professional evaluations of visual adequacy [25].

3.3. Meat Product Photography

For comparative analysis, we photographed each of the meat products generated in the preliminary study. These photographs were taken in a controlled indoor environment, using a white surface and background to minimize distracting visual elements and to direct the viewer’s focus exclusively to the primary subject—the meat product. Illumination was provided by two lights positioned on either side of the subject, resulting in an even distribution of light and reduced shadows. Lighting conditions were adjusted for each shot to approximate the appearance of the AI-generated images. A Nikon Z50 mirrorless camera equipped with a 16–50 mm lens was used to capture all photographs. Each meat product or dish was presented on a plain white plate, without additional decorative garnishes that could divert attention. This method prioritized clear composition and allowed for easy comparison with the AI-generated images. Additionally, the subjects were always centered in the frame, and the camera angle was kept consistent to align with the perspective of the generated images.

While we aimed to make the AI-generated image and photograph as comparable as possible, it is important to acknowledge a key limitation: each meat product was represented by only a single photographic execution. This does not fully capture the variability found in professional food photography, where different lighting styles, angles, and artistic choices can influence perception. Including multiple photographic interpretations per item would provide a more robust basis for comparison but was beyond the scope of this study. The primary goal was not to evaluate specific items like mortadella or beef tartare in isolation, but rather to examine the general viability of GenAI tools for representing meat products as a category. Therefore, our findings are based on the average performance across all products, and the overall appeal and accuracy trends are more meaningful than item-level conclusions. This broader perspective allows for a more realistic assessment of GenAI’s practical utility in the food industry, where batch production of promotional content is common, and individual visual perfection is not always feasible.

Photographs were captured in RAW format and processed in Adobe Lightroom Classic (version 14.3.1). All image editing was performed in the sRGB color space with the explicit aim of bringing the photographs’ appearance closer to that of the AI-generated images, thereby ensuring equivalence of conditions for the subsequent survey comparison. Key objectives of minimal post-processing included standardizing the background, enhancing the whiteness surrounding the primary subject, and applying slight corrections to contrast and color saturation, particularly for the meat products themselves. We made these adjustments because we shot in RAW format to ensure high-quality photographs. These changes were necessary to prevent the actual photos from looking “flat” compared to the generated images. Our goal was to make sure they could stand out equally in terms of aesthetic appeal during the survey.

It is essential to emphasize that the survey’s purpose was not to distinguish between authentic and AI-generated images, but instead to evaluate their aesthetic appeal and professional suitability. For this reason, we sought to equalize conditions through minimal post-processing while preserving the authenticity of the photographs. Editing interventions were therefore limited to essential adjustments, such as white balance calibration, tonal contrast leveling, and targeted exposure corrections in specific image regions. To illustrate the effect of minimal post-processing, we present an example in Figure 3: the unedited original photograph is shown on the left (Figure 3a), and the final version after digital optimization appears on the right (Figure 3b). This example demonstrates how modest adjustments can enhance clarity, cleanliness, and visual appeal without compromising the realistic representation of the meat product.

3.4. Likability Comparison Between AI-Generated Images and Photographs

The preference comparison between AI-generated images and traditional photographs was conducted via the online platform fototeka.si. To accommodate a broader audience, including both domestic and international respondents, the questionnaire was administered in both Slovenian and English. Participants were randomly assigned to one of two groups (A or B) to determine whether prior knowledge of an image’s origin would influence their aesthetic judgments. Group A received introductory instructions that explicitly stated the survey would involve a comparison between images created by GenAI and real-life photographs. In contrast, Group B’s instructions omitted any reference to AI, allowing us to examine whether awareness of the technology alone affected participants’ choices. Aside from this initial distinction in wording, both groups were presented with the identical set of image pairs.

Upon accessing the survey, each respondent first completed a brief demographic questionnaire requesting information on gender, age, and nationality. These data were essential for subsequent subgroup analyses, enabling us to explore potential differences in preference patterns across demographic categories. In addition, to ensure that participants could objectively evaluate the visual content, they were required to confirm their professional expertise in food technology and to indicate whether they had normal or corrected vision. Respondents who reported uncorrected visual impairments were excluded from further participation, as accurate color discrimination and detail recognition were critical for the study’s objectives.

Instructions for both groups were deliberately concise yet comprehensive. Participants were informed “You will see 15 pairs of images displayed side by side on a white background. For each pair, please select the image you find more visually appealing.” This simple directive minimized cognitive load and focused respondents’ attention solely on visual preference, rather than other criteria such as perceived realism or technical execution.

The main body of the survey consisted of 15 sequential image pairs, each presenting a real photograph of a meat product alongside an AI-generated rendition of the same item. To prevent any positional bias, the position of the photograph and its corresponding AI image were randomized for each pair. Images were presented without captions, labels, or additional decorative elements, ensuring that participants’ selections were based purely on visual appeal. The collection of food items included a diverse range of meat products and dishes: burger, Wiener schnitzel, hot dog, young beef liver, Carniolan sausage, mortadella, roasted chicken, chicken medallions, chicken skewers, prosciutto, steak, dry salami, spaghetti bolognese, ham in dough, and beef tartare. Each participant made one binary choice per pair, resulting in 15 independent preference decisions (Figure 4).

Upon completing the fifteen selections, participants were directed to a brief closing page where they received a written acknowledgment and were provided with contact information for the research team. This follow-up option allowed respondents to submit any questions, comments, or requests for further information regarding the study’s findings. By implementing this dual-language, randomized, and professionally vetted survey design, we were able to collect rich quantitative data on image preference, while also establishing a channel for ongoing engagement with participants in future related research endeavors.

4. Results

4.1. Evaluation of the Suitability of AI-Generated Images

The preliminary study we conducted to determine the appropriateness of images, generated with Adobe Firefly 3, is shown in Figure 5. The selected products included burger, Wiener schnitzel, hot dog, Carniolan sausage, mortadella, chicken skewers, prosciutto, steak, dry salami, and steak tartare [25].

Twenty-six people participated in the professional evaluation; however, based on predefined criteria, we included responses from only twenty-two participants. Qualified participants were food industry professionals with good vision or who used corrective lenses or other visual aids. Four participants who did not meet these criteria were excluded. Demographic data revealed a predominance of female participation (21 women, 1 man), with an average age of approximately 54 years. Most participants (64%) were aged between 55 and 65, followed by 27% aged 40–54 and 9% aged 25–39. Regarding vision, 16 participants reported adequate vision without corrective aids, and 6 used corrective aids, ensuring reliable and visually objective evaluations [25].

Participants assessed AI-generated images based on product color, texture, presentation, and overall visual appeal. If an image was deemed unsuitable, participants provided comments explaining their reasoning. Overall, experts rated GenAI as 43.7% successful in producing professionally suitable images.

Table 1. Professional accuracy of AI-generated meat product images (n = 22).

Meat Product	Professionally Accurate Image	Professionally Inaccurate Image
Burger	82%	18%
Prosciutto	67%	33%
Chicken skewers	59%	41%
Wiener schnitzel	57%	43%
Steak	50%	50%
Carniolan sausage	36%	64%
Dry salami	29%	71%
Hot dog	29%	71%
Steak tartare	19%	81%
Mortadella	9%	91%

shows detailed expert evaluations [25].

Experts highlighted GenAI’s current limitations in replicating complex meat structures, especially in color accuracy, texture realism, and composition. In Figure 5, expert approval varied: the burger (5a) was rated as adequate by 82%, while 18% noted color issues, and cited an issue with visibility of the meat inside it. Prosciutto (5b) was accepted by 67%, while 33% criticized its color and presentation. Chicken skewers (5c) received only 41% approval, with complaints regarding their artificial appearance. Additional comments mentioned that the meat appeared to be of lower quality. The Wiener schnitzel (5d) was rated adequate by 57%, with 43% citing issues with color and plating. Carniolan sausage (5e) was found to be inadequate by 64%, lacking its distinctive tapered tip. It was also noted that the appearance made it seem more like a grilled sausage. Both dry salami (5f) and hot dog (5g) were rejected by 71% for unrealistic proportions and color, with suggestions to use natural casing in the hot dog image. Beef tartare (5h) was deemed inadequate by 81% for its unnatural appearance, specifically color. Mortadella (5i) received the harshest critique, with 91% of experts rejecting it due to its inaccurate cross-section and color [25].

These findings from this experimental phase of the study underscore the current limitations of Adobe Firefly 3 in accurately depicting meat products for professional use. In 9 out of 10 cases, color was the most inappropriate factor, followed by the texture of the meat and the plate appearance in some cases. Although specific images, such as those of the burger and prosciutto, were judged relatively acceptable, more complex items like mortadella and beef tartare were consistently deemed inadequate due to unrealistic color, texture, and cross-section details. Color inaccuracies were the most frequently noted issue, as slight deviations in hue can imply differences in freshness, quality, or even edibility. Additionally, errors in cross-section composition and plate presentation indicate that Adobe Firefly 3 struggles to reproduce the intricate structural details of both processed and raw meats. Creating visually attractive images does not guarantee technical accuracy, which is crucial in the food industry, where consumer trust relies on accurate product representation [25].

It is important to highlight notable failure cases where Adobe Firefly 3 produced anatomically implausible textures in meat products. For instance, the AI-generated image of mortadella was deemed professionally inaccurate by 91% of food experts, primarily due to the absence of its signature white fat particles and an unconvincing cross-section structure. Similarly, the image of beef tartare was rejected by 81% of professionals for its overly smooth texture and unnatural color, resulting in an unrealistic, plastic-like appearance. These examples expose GenAI’s current struggles with replicating fine-grained internal structures and natural irregularities—such as consistent marbling or tissue variation—that are critical for conveying authenticity. Identifying and documenting such flaws helps define the technical hurdles that must be overcome for AI to generate images suitable for professional food communication.

These results highlight the indispensable role of professional photographers. Unlike AI, photographers apply their expertise in lighting, composition, and post-processing to present meat products in an aesthetically pleasing and realistic manner. They can also make minor but essential adjustments to meet client requirements and industry standards—capabilities that AI currently lacks. Thus, while GenAI offers a cost-effective solution with marginally lower quality, it remains unable to replace professional photography for accurately and professionally representing meat products [25].

4.2. Analysis of AI-Generated Image Attractiveness Compared to Photographs

In the second study, we compared the attractiveness of images generated with Adobe Firefly 3 with that of actual photographs. Fifteen photographs were taken under controlled lighting conditions; these were minimally processed in order to match the AI-generated images visually. An additional five images (young beef liver, chicken nuggets, roasted chicken, spaghetti bolognese, and ham in dough) expanded the survey to 15 image pairs, shown in Figure 6, Figure 7, Figure 8, Figure 9, Figure 10, Figure 11, Figure 12, Figure 13, Figure 14, Figure 15, Figure 16, Figure 17, Figure 18, Figure 19 and Figure 20.

Demographics showed participants from 14 countries: Belgium, France, Greece, Hungary, Italy, the Netherlands, Norway, Portugal, Spain, Sweden, Ukraine, the United Kingdom, Luxembourg, and Slovenia. Participants were divided into two groups, designated Group A and Group B, based on the differing introductory instructions they received. Group A was explicitly informed in the instructions that the study concerned preference judgments between images generated by AI and actual photographs, whereas Group B received no mention of AI. Consequently, at the outset of the survey, 253 participants were aware that they were comparing AI-generated images with photographs of real products, whereas 117 respondents completed the comparison without this prior knowledge.

A comparative analysis of responses from both groups revealed no significant differences in the results of each test group; therefore, we combined the data from both groups for a unified analysis. Participant feedback indicated that the majority became aware of the AI-generated nature of some images even when this was not explicitly stated in the instructions. This suggests that mentioning AI in the introductory text did not substantially influence participants’ perceptions of image preference, thereby supporting the assumption that users across all age cohorts increasingly respond to AI-created visual content in a spontaneous and unbiased manner.

In examining gender differences, women exhibited a slightly greater tolerance toward AI-generated images, selecting them as preferable in 60.68% of cases. Men chose AI-generated images in 57.71% of instances, indicating a somewhat lower but still majority preference for AI-created visuals. Age-related differences were also notable. The highest rate of preference for AI-generated images (70.24%) was observed among respondents aged 25–44 years. This cohort is likely the most AI-aware and also the most accepting of AI’s presence in everyday contexts such as advertising, design, and visual communication. Respondents under 25 years of age selected AI-generated images in 67.46% of cases, which is consistent with the fact that this generation frequently engages with computer technologies and possesses higher levels of digital literacy. Among participants aged 45–64 years, the preference rate for AI-generated images was slightly lower yet still above half, at 55.74%. This finding suggests a gradual acceptance of AI usage among older users, who are typically more cautious and skeptical about new technologies. As anticipated, respondents over 65 years of age demonstrated the least inclination toward AI-generated images; however, even within this group, 43.33% preferred AI-generated visuals, illustrating a general openness to such technologies.

These demographic trends were supported by statistical analysis. Statistically significant gender differences were observed in the youngest age groups (under 25 and 25–44) for several products and both image types, with p-values below 0.05. However, the effect sizes (Cramér’s V ranging from 0.13 to 0.21) indicate that while the differences are statistically significant, they are modest in practical terms. No significant gender differences were found among participants aged 45–64 or 65+, suggesting that the visual preferences of older respondents were not influenced by gender.

Across all age groups, genders, and products, an average of 59.19% of participants selected AI-generated images as their preference. This result confirms that, from the standpoint of general preference, AI-generated images are visually comparable to photographs of real products. While the overall preference for AI-generated images was strong, particularly for items like burgers and roasted chicken, the standard deviation of ±19.75 highlights significant variation across products. This suggests that AI’s visual appeal is not uniform and may depend on product complexity or recognizability, which should be explored in future research. Nonetheless, it is essential to note that this does not imply professional adequacy, particularly in terms of accurate color rendition, realistic texture, and the recognizability of each product’s distinguishing features. In other words, visual appeal does not necessarily equate to fidelity or reality.

Figure 21 presents detailed preference results for AI-generated images, showing, for each meat product, the number of respondents who selected the AI-generated image as preferable versus those who chose the photograph.

AI was used to generate a highly varied range of meat product images in terms of preference percentages among non-expert viewers. The most preferred AI-generated items were chosen by 92.90% of respondents, while the least preferred were selected by 31.42%. Notably, even the least-selected image garnered a substantial proportion of preferences, suggesting a broad increase in the acceptance of AI-generated content, irrespective of specific demographic factors.

Adobe Firefly 3 was most successful in generating appealing images of a burger (Figure 15) and roasted chicken, both of which received a 92.90% preference rate. The burger image was preferred by 93.75% of women and 91.27% of men. Notably, these were the only items created with the help of sample photographs, suggesting that reference images enhance the realism and attractiveness of AI outputs. The Carniolan sausage image (Figure 20) also scored highly (90.16%), despite experts previously rating it as inadequate due to missing product-specific features. A similar pattern emerged with mortadella (Figure 19), which lacked defining traits but was preferred by 65.30% of respondents. These cases highlight a gap between professional standards and general public preference, underscoring the potential of combining AI tools with expert guidance and reference imagery to improve both visual appeal and accuracy.

In addition to the most highly rated cases, eight other images, generated with Adobe Firefly 3, were preferred by more than 50% of respondents. The image of spaghetti bolognese (Figure 16) received the highest preference at 88.25%, followed by chicken medallions (Figure 11) at 75.96%, and the hot dog (Figure 9) at 71.58%. Chicken skewers (Figure 10) were preferred by 70.22% of participants. Preference rates between 65% and 70% were recorded for ham in dough (Figure 18), young beef liver (Figure 7), and mortadella (Figure 19). The last AI image to receive more than 50% preference was Wiener schnitzel (Figure 8), at 63.11%. In contrast, the least preferred AI-generated images were those of beef tartare (Figure 17), dry salami (Figure 13), prosciutto (Figure 6), and steak (Figure 12). These same images had also been rated as less professionally adequate by food technology experts. The main reason for their low evaluations was the unnatural coloration in the AI-generated versions, which discouraged many respondents. This clearly indicates that color accuracy and realism are crucial for the perceived quality of AI-generated images, particularly for meat products, where hue and saturation significantly impact impressions of freshness and quality.

To assess whether the observed preferences for AI-generated images were statistically significant, we conducted chi-square tests for each product. The results confirm that for 14 out of 15 items, the preference distributions were statistically significant at the p < 0.05 level. For example, preferences for AI-generated images were particularly strong for burger and roasted chicken, as well as Carniolan sausage. The only product for which the difference was not statistically significant was steak (p = 0.12). This suggests that, while average preference leaned toward AI imagery, the strength of this preference varied by product.

One of the most notable findings of this study is that the use of reference photographs during the AI image generation process significantly enhanced the visual quality and appeal of the resulting images. Specifically, the AI-generated images of the burger and roasted chicken—both created with the help of photographic samples—achieved the highest preference ratings among all image pairs, each selected by 92.9% of respondents. This strongly suggests that combining photographic inputs with textual prompts enables GenAI systems to better replicate real-world visual features such as texture, structure, and color accuracy.

This ‘hybrid’ creation process offers a practical solution for food marketers and content creators, balancing the visual appeal of AI-generated images with the credibility and recognizability of real products. To facilitate implementation, we propose a structured five-step workflow:

• Capture or identify a reference photograph that illustrates the desired visual attributes;
• Describe key features and stylistic targets;
• Prompt the AI using both the image and detailed textual input;
• Validate the output against a visual checklist (e.g., color fidelity, shape realism, product recognizability);
• Refine results through human-in-the-loop feedback or light editing.

This reference-supported strategy offers a path forward for professionals seeking efficiency without sacrificing accuracy, and should be considered a central contribution to this study.

Overall, it can be concluded that Adobe Firefly 3 achieves highly variable outcomes. For meat products where details such as color and cross-section composition are crucial, the results are less satisfactory. By contrast, items that are more visually recognizable and allow greater aesthetic flexibility, such as burgers or spaghetti, achieve much higher preference rates.

5. Discussion Results

Based on the studies conducted, it is evident that the use of AI in generating visual content for the food industry holds significant potential, while simultaneously raising multifaceted questions regarding product perception, authenticity, and communication. Among survey participants, images generated with Adobe Firefly 3 were preferred over actual photographs, confirming that contemporary consumers of all ages are primarily visually oriented and often respond more positively to aesthetically enhanced imagery than to strictly realistic depictions.

The statistical analysis confirms that the preference for AI-generated images was not only visually apparent but also statistically robust. Chi-square tests showed that for 14 out of the 15 meat products evaluated, the preference for AI images over photographs was significant at the p < 0.05 level. The strongest preferences were observed for burger, roasted chicken, and Carniolan sausage, each exceeding 90% AI preference and yielding extremely high chi-square values (p < 10⁻²⁹). Products like spaghetti bolognese, chicken medallions, and hot dogs also showed strong and significant favorability toward AI images. In contrast, steak was the only item where the observed difference did not reach statistical significance (p = 0.12), suggesting that for certain products with more complex or recognizable textures, AI may still fall short in visual persuasion. These results confirm that the observed preference trends are not due to chance and support the broader conclusion that GenAI imagery, particularly when visually optimized, can surpass conventional photography in consumer appeal across a wide range of meat products.

The preference for AI-generated images, such as the 92.9% favorability toward the AI-rendered burger, suggests that lay audiences are drawn to their idealized, clean, and stylized aesthetic. This likely reflects contemporary visual norms shaped by advertising and social media. However, this appeal contrasts with expert evaluations, which highlight AI’s shortcomings in accurately depicting color, texture, and structure. The gap points to both a marketing opportunity and a perceptual limitation: while GenAI excels in creating visually engaging content, it risks misleading consumers, especially with traditional or regulated products. Therefore, AI-generated visuals may be best used as a complement to photography, with clear labeling and reference-based refinement to balance appeal and accuracy.

Nevertheless, concerns remain regarding the professional adequacy of AI-generated images. In the expert evaluation, food-technology specialists clearly demonstrated that images generated with Adobe Firefly 3 frequently fail to meet the required standards for accurate color reproduction, structural detail, shape fidelity, and recognizability of characteristic features. Such deficiencies may confuse consumers, especially when dealing with protected or traditional products (e.g., Carniolan sausage), where visual accuracy is essential. Although AI-generated images can create a positive user experience, they can also contribute to misperceptions of the product, potentially leading to suboptimal purchasing decisions in certain contexts [25]. Results from the second study on lay preferences showed that younger age cohorts were more likely to favor images generated with Adobe Firefly 3, even when they were aware that the images were created by AI. Notably, older age groups did not uniformly reject these images; a significant proportion still found them visually appealing, indicating a shift in attitudes toward technology across all demographic segments.

The concept of the uncanny valley [15] offers a compelling lens through which to interpret the negative reactions to specific AI-generated images, particularly mortadella and beef tartare. These products were consistently rated poorly by both experts and lay participants, not because they were entirely unrealistic, but because they were almost correct in form, yet subtly off in critical ways (e.g., color, cross-section details, or texture). This near-realism likely triggered a perceptual dissonance similar to that observed in uncanny valley responses, where images that closely mimic real objects, but fall short in subtle ways, can evoke discomfort or distrust. In food imagery, such discrepancies may undermine appetitive responses, particularly for processed meats where visual cues strongly signal quality and safety. Recognizing this effect is essential for improving GenAI tools in domains where realism is a key trust factor.

A key technical insight from this study is that using reference photographs during AI image generation greatly improved visual fidelity and appeal. The burger and roasted chicken images, both created with sample photos, received the highest preference ratings (92.9%), showing that combining visual references with textual prompts helps GenAI better replicate crucial features like texture and color. This hybrid approach offers a practical solution for professionals aiming to balance aesthetic quality with product recognizability.

Moreover, cross-cultural differences in product perception were observed among survey participants. Individuals from different countries may hold varying expectations of how a given product should appear, which complicates the question of whether AI-generated images can achieve universal acceptance. Future research should therefore investigate how cultural backgrounds influence the reception of such content and how generative models might incorporate regional specificities.

In addition to technical and perceptual considerations, the ethical implications of using AI-generated food imagery deserve closer attention. As AI systems continue to evolve and approach greater levels of visual accuracy, particularly in terms of texture, structure, and culturally specific visual cues, it becomes increasingly important to implement robust regulatory practices. One of the most pressing issues is the transparent labeling of AI-generated images. Without a clear indication, consumers may be misled about the authenticity of the products depicted, especially when dealing with traditional or protected foods where visual characteristics are tightly linked to quality and origin. Proper labeling not only preserves consumer trust but also upholds ethical standards in advertising and product communication. As generative models improve, closing the gap between real and synthetic images, the responsibility to differentiate the two will grow accordingly. Moreover, future standards should address whether watermarking or provenance metadata is preserved in AI-generated images, as such mechanisms may play a crucial role in maintaining consumer trust once disclosure becomes mandatory. Ensuring visual appeal does not override transparency will be key to fostering responsible innovation and ethical communication in the food sector.

Looking ahead, future AI development may significantly reduce the current limitations in structural accuracy, textural realism, and visual consistency. Advances in multimodal training, higher-resolution modeling, and incorporation of photographic reference data could allow AI systems to generate food images that are nearly indistinguishable from real photographs. Such progress would expand the practical utility of generative models, not only for marketing but also for product design, education, and labeling, provided that ethical and regulatory safeguards evolve alongside these capabilities.

In summary, both studies indicate that AI-generated images of meat products still do not attain the professional precision and authenticity afforded by real photographs; as a result, traditional photography remains superior in certain contexts. However, AI-generated images are already proving useful in visual communication, especially in domains where aesthetic impact and creative freedom outweigh the need for absolute realism (for example, in advertising materials or rapid prototyping of design concepts). Moving forward, a thoughtful integration of both approaches will be essential, combining the visual appeal and flexibility of AI-generated images with the authenticity and informational value of real photographs. Such a hybrid strategy could yield an optimal balance between message effectiveness and content reliability while preserving consumer trust. It will also be critically important to implement consistent and transparent labeling of AI-generated content, ensuring that consumers always know the origin of the visual material. Visual appeal and aesthetic refinement must not override the need for transparency, as failure to disclose the source could lead to misleading interpretations, undermine credibility, and erode consumer trust. Therefore, the potential application of AI in the food sector will require clear guidelines, ethical orientation, and collaboration among industry experts, regulatory bodies, and marketing practitioners.

6. Conclusions

Within this article, we examined the potential applications of GenAI, specifically Adobe Firefly 3, for the visual presentation of meat products. We focused on two dimensions: the professional adequacy of AI-generated images and their aesthetic appeal relative to actual photographs. By conducting two distinct studies, we sought to determine whether GenAI can produce depictions of food items that are both technically accurate and visually attractive for use within the food industry.

The preliminary study enlisted food science experts to evaluate the professional adequacy of AI-generated images. Its results revealed significant limitations: while images of simpler products—such as a burger or prosciutto—were judged to be acceptable, more complex items (for example, mortadella and beef tartare) were rendered inaccurately, exhibiting unnatural colors, incorrect texture, and unrealistic cross-section composition. These findings confirm our initial hypothesis that, at present, GenAI is not sufficiently advanced to replace camera-captured photographs of meat products; in particular, Adobe Firefly 3 still cannot reliably reproduce the critical sensory attributes required for professional use [25].

As mentioned, some of the lowest-rated AI-generated images, such as mortadella and beef tartare, were visually close to reality but failed in subtle details like cross-section accuracy and color tone. This aligns with the “uncanny valley” effect introduced in the literature review: when food images are almost lifelike but contain minor abnormalities, viewers may experience discomfort or distrust [14]. Diel et al. (2025) [14] showed that people are particularly sensitive to such near-realistic distortions in food images, especially those with high disgust sensitivity or food neophobia. Our findings reflect this, as both experts and lay participants rated these “nearly correct” images poorly, suggesting that AI outputs that fail just slightly in realism may provoke stronger negative reactions than clearly stylized ones.

In the second study, lay respondents compared the aesthetic appeal of AI-generated images to that of corresponding photographs. Here, images generated with Adobe Firefly 3 were frequently rated as more attractive, especially when a sample photograph was provided during the generation process. Preference rates were higher among participants aged 25–44 years and among women, indicating generational and gender-based differences in the reception of digitally produced visuals. We further observed that participants’ ignorance of an image’s origin (AI versus photography) did not significantly influence their judgments: those who were not informed about the AI-generated nature of the images produced results similar to those who were. This suggests that while some bias against AI may exist, it is not decisive. Overall, 60.68% of non-expert respondents preferred images generated with Adobe Firefly 3 over photographs, confirming our second hypothesis that lay audiences find AI-created meat imagery more appealing than traditional food photographs.

Taken together, these results demonstrate that GenAI—in our case, Adobe Firefly 3—does not yet offer a complete substitute for professional photography in the food industry context, particularly in cases requiring exacting, expert-level depictions. Nonetheless, it holds considerable promise as a supplementary tool, especially for producing visually engaging content in advertising or conceptual design contexts where strict technical accuracy is less critical. Continued technological advancements and the strategic use of reference photographs can further enhance the quality and reliability of AI-generated food imagery.