Generative AI vs. reality: how do virtual try-ons compare to real on-model content?

AI technology in the fashion industry

The technology changed the pace of fashion marketing, and it has never been more embedded in the creative process. Brands are now relying on AI not just to assist but to generate imagery for both campaign assets and product pages (PDPs). This shift is altering how fashion visuals are conceived, produced, and monetised.

Generative image models and specialised AI workflows are increasingly tailored for fashion use cases. On-model photos, brand-specific assets, and even automated ad generation are now possible in minutes.

On the “model” side of things, as pointed out in The Interline’s article, some AIs generate realistically-looking virtual models and lifestyle backgrounds, allowing brands to visualise garments on diverse bodies, backgrounds, and scenarios without booking a physical studio. Industry example? You got it. The extremely visible move by H&M to work with models and agencies to create “digital twins” is setting a new benchmark in rights, representation, and reuse of model likenesses. In their initiative, models retain ownership of their digital replicas, are compensated, and may even license their twin to other brands.

We know brands are already experimenting with generative AI to create content for all kinds of purposes. But the product detail page (PDP) content is different. Here, the visuals must be trustworthy, accurate, and high-quality. Otherwise, there’s a real risk of overpromising or underdelivering. Customers may receive something that looks far from what they expected, which damages brand credibility and can drive up return rates (and we already know how big a problem it is in e-commerce). In other words, a tool meant to save money in one part of the workflow can easily end up hurting businesses.

That’s why we decided to check the capabilities of AI in terms of the fashion industry and compare it with a real photoshoot session.

💡Want to see how AI responds to the challenge of doing perfume lifestyle shots? Check out our previous blog post: State of generative AI technology for product photography: creating lifestyle perfume shots with AI.

The test base

Now, in our previous article about AI technology in lifestyle perfume photography, we compared 5 different AI models/tools and tried to achieve professional results with a simple prompt. This time, however, the prompt is more advanced; we used two Orbitvu solutions to produce content, and there are two types of photos: on-model (created in Fashion Studio as reference images/videos) and packshots (created in Alphastudio XXL as source images for generative AI). 

The goal is to achieve the same quality and authenticity as the original photos done in Fashion Studio, but in the AI process.

Packshots & model shots

AI tools: image and video

We will test 4 popular AI image-to-image generators to create two on-model pictures from two source ghost images (front and back). Then, using the best two on-model images and 3 state-of-the-art image-to-video generators on the market, we will try to replicate the original video.

Image-to-image AI models:

1. Google Nano Banana PRO - Nano Banana is the next-generation AI image generator/editing platform (powered by Google’s Gemini 3.0 model) that lets you turn text into images, edit photos with simple language, maintain visual identity across edits, and fuse multiple images, all designed for creators needing high-quality and consistent visuals. The latest update enables users to generate images at higher resolutions, including 2K and 4K, in addition to the standard 1 K resolution.
    
2. Flux Kontext [PRO] - FLUX 1 Kontext is a next-generation AI image model by Black Forest Labs that combines text prompts and image inputs to create or edit visuals with strong context‐awareness, object/character consistency, and professional-grade output.
    
3. Seedream 4.0 by ByteDance - Seedream is the next-generation multimodal AI image model. It blends generation and editing, works with both text and images, supports multiple reference inputs, and delivers ultra-high-resolution visuals quickly. Its multimodal “reasoning” capabilities make it more than just an art toy. It's positioned for professional workflows.
    
4. ChatGPT - the ChatGPT AI Image Generator is a feature built into OpenAI’s ChatGPT that allows users to create and edit images using natural language. Powered by DALL-E 3, it enables you to generate detailed visuals directly from text prompts or modify existing images with simple instructions. ChatGPT is also very useful for creating prompts and task ideas.

Image-to-video AI generators:

1. Veo3 - a next-generation text-to-video and image-to-video tool from Google. It allows users to input a text prompt (or optionally reference images) and automatically generate short cinematic clips with synchronized audio, realistic motion, and high visual fidelity.
    
2. Kling AI - an AI video-generation platform developed by Kuaishou Technology in China. It supports converting text prompts (and even static images) into dynamic videos with realistic motion and cinematic style.
    
3. Seedance 1.0 PRO - an advanced AI video generation model developed by ByteDance (the creators of TikTok). It specializes in converting text prompts and static images into high-quality, cinematic videos (up to 1080p).

Testing AI tools: which AI image generator is best at generating fashion PDP images?

With today's advances in AI technology, is it possible to create content that doesn’t deviate too much from reality? Are the imperfections that we saw a moment ago in every image generated by artificial intelligence still visible? Let’s take a closer look at the popular AI tools on the market and check whether a good packshot and a good prompt will be able to replace a full photo shoot for e-commerce. 

The criteria we will evaluate are primarily whether artificial intelligence will generate images for us:

1. Consistency: how the two images of the same garment, both front and back, are consistent in terms of model look, accessories, and overall garment consistency.
2. Product fidelity: whether the product we photographed, in this case, a dress, is represented faithfully, including colors, patterns, its shape, and size. How realistically does it fit on the model? 
3. Costs: is it worth the money? 
4. Prompt adherence: are all the instructions followed?  

Comparison of Nano Banana

Evaluation

Consistency and discrepancies:
While the model appears to be the same in both images, there is a marked difference in the overall tone. The back view color tones are noticeably colder than the front view. Additionally, the dress length varies significantly, with the back view showing a much shorter garment. Minor, non-obvious differences in shoe shape are also present. The dress length does not match between images.

Product fidelity:
The generated images generally maintain good product fidelity regarding the dress's pattern, overall shape, and textile. However, there are two key inaccuracies:
Sleeve shape: The sleeve shape is incorrectly rendered, appearing much smaller than in the real-life product.
Dress size (back view): The dress is rendered slightly too short in the back view compared to the actual product.
Color and tone reproduction are accurate in the front view, but the back view suffers from being excessively warm.

Prompt adherence:
The images largely followed the prompt's instructions. The only deviation is the background color, which is a light gray instead of the requested white.

Costs for Nano Banana Pro image generation:
The cost for generating images depends on the desired output resolution:

• ~1K resolution: $0.24 per image
• 2K resolution: $0.24 per image
• 4K resolution: Up to $0.47 per image
   

Comparison of Flux Kontext [PRO]

Evaluation

Consistency:

The white background is clean and consistent across views. The overall color palette and floral motif are maintained throughout. However, the right image is slightly underexposed, with visible shadowing in the center of the back, which impacts visual consistency. Additionally, the shoes are clearly different between views, disrupting visual consistency.

Product fidelity:

Flux Kontext PRO successfully preserves the general silhouette, correct dress length, overall color palette, and floral motif of the dress. The recognizable combination of a deep red background and bright pink floral print is maintained, and no clear differences in the pattern itself are noticeable. On-screen, no obvious color inconsistencies are visible, although very subtle variations may exist.

However, important deviations include: the sleeve is too small. The length of the shirt in the back view seems a bit short. 

Prompt adherence:

The white background was properly followed as requested, supporting a good overall presentation. The general try-on concept was executed successfully. However, the model looks somewhat artificial and plastic-like, reducing realism, which suggests limitations in achieving the intended photorealistic quality typical of product photography standards.

Cost for Flux Kontext PRO image generation:

• ~1K resolution: $0.12 per image
• 2K resolution: $0.18 per image
• 4K resolution: unavailable

Comparison of Seedream 4.0

Evaluation

Consistency:

Different shoes are used across the two images, disrupting consistency. It's also visually apparent that the models' faces differ between images, indicating a lack of continuity between views. Color tones are also different in both views, while the back view is more true to the original image.

Product fidelity:

Seedream captures the recognizable floral print and overall color palette of the original dress, maintaining its visual identity at a glance. Notably, only this AI model managed to reproduce the long sleeves of the dress.

However, several inaccuracies reduce fidelity: the dress is noticeably too short, and its proportions differ from the original, most notably in the neckline shape, which doesn't match the authentic design. The drape and structure of the fabric are not fully convincing, as the material's true shape and natural flow on the model are not accurately reproduced. The system falls short of a one-to-one reproduction, particularly in length accuracy, neckline shape, and fabric behavior. Overall, both images seem to have too high contrast, and the dress looks underexposed.   

Prompt adherence:

The images appear too dark and insufficiently lit, particularly on the front view, which obscures garment details. This suggests the lighting specifications in the prompt were not properly followed. Overall, SeeDream delivers a visually appealing AI try-on that reflects the general concept, but the lighting execution falls short of typical studio packshot standards.

Cost for Seedream image generation:

• ~1K resolution: unavailable
• 2K resolution: $0.09 per image
• 4K resolution:$0.09 per image

Comparison of ChatGPT

Evaluation

Consistency:

The color rendering has changed and varies between images, resulting in noticeable differences rather than a consistent palette across views. The model looks notably different in both views, including face, hair, and size.

Product fidelity:

The AI-generated images preserve the general silhouette and floral pattern of the original dress. However, several inconsistencies reduce overall fidelity: the dress is visibly too short compared to the original, and the sleeve appears too narrow, which affects proportions and fit accuracy. The fabric appears unnatural in shape and behavior, particularly in the front view, where the drape and structure don't reflect how the material realistically falls on the body. Color tones are visibly different from the original. While the AI output captures the general idea of the design, it doesn't fully replicate the authentic look and construction of the garment in terms of length, sleeve sizing, fabric realism, and other structural details.

Prompt adherence:

The general concept was captured, but the execution suggests limitations in achieving the intended photorealistic quality and accurate garment representation typical of product photography standards.

Cost for ChatGPT image generation:

• ~1K resolution: $0.14 per image
• 2K resolution: $0.47 per image
• 4K resolution: unavailable

Results summary

Our choice: Flux Kontext

Based on image quality, contextual intelligence, and production-readiness, Flux Kontext clearly outperformed the other models tested. Its strengths in realistic garment rendering and consistent scene generation make it particularly well-suited for fashion content at scale.

As a result, Flux Kontext will be the base layer for AI-driven video creation, where consistency and realism are non-negotiable.

 

Testing AI tools: is it possible to create true-to-reality videos for fashion campaigns

Now that we know how AI image generators perform, let’s see the capabilities of video AI tools. We have the best photos - let’s bring them to life.

The goal of our comparison is to check how video generation tools cope with the image-to-video task. We will evaluate them in terms of:

1. Consistency with the prompt: matching the movement and timing of the model - and the movements she performs.

2. Fidelity: whether our generated model and the dress we photographed have not been modified in any way, in terms of texture, colors, or shape.

3. Physics: the arrangement of the material on the body, the model's movement, the overall naturalness of the shot

4. Cost: is it worth the hype? The costs?

 

 

Comparison of Seedance 1.0 Pro

Evaluation

Consistency with the prompt:

Seedance followed the prompt well and didn't make any noticeable mistakes in terms of matching the movement and timing of the model and the movements she performs. 

Fidelity:

The dress appears to have been preserved without modifications to its texture, colors, or shape. However, the realism falls short - when the model turns, a brief acceleration glitch becomes noticeable, interrupting the smooth flow of the video.

Physics:

There’s no natural body movement, which affects the fabric's behavior in the video. The hair during the rotation also looks fake - instead of being naturally tossed back, it looks as if it has been pulled over the shoulder.

Cost for Seedance 1.0 Pro video generation:

Cost Full HD: $1.81 per 8-second video

Comparison of Veo3

Evaluation

Consistency with the prompt: 
The model's movement is quite good and natural, suggesting that the timing and movements align well with what was requested.

Fidelity:
The program reads the fabric perfectly—you can clearly feel the silk quality of the dress in every shot, indicating that the texture and material properties of the garment have been preserved accurately. 

Physics:
The physics of draping, creasing, and the fluttering of the hanging sash are excellent, demonstrating strong natural fabric behavior. However, in one instance, the hair is nicely tossed by the model, but in the remaining shots, it’s pulled over the shoulder in an unrealistic way, which slightly affects the overall naturalness.

Cost for Veo3 video generation:

Cost Full HD: $3.03 per 8-second video
 

Comparison of Kling

Evaluation

Authenticity:

Kling shows a lot of potential—it has the most vivid and lifelike model movements, creating a highly realistic overall visual perception. However, in one instance, there is a transition where the front shifts into the 'end frame' in a highly unrealistic way, which disrupts the authenticity.

Consistency with the prompt:

The model movements are vivid and lifelike, suggesting strong alignment with the requested timing and actions.

Fidelity:

The dress and model appear to be preserved accurately throughout the video, with no noticeable modifications to the garment's appearance.

Physics:

The fabric physics are very well preserved, with the dress draping beautifully and the hair moving naturally as well, demonstrating excellent natural behavior of both the material and the model.

Cost for Kling video generation:

Cost Full HD: $2.65 per 10-second video

Estimation of costs: traditional photography vs. automated photo studio vs. generative AI

To compare the costs of generative AI with traditional and automated photo studios, we made the following assumptions:

1. For a traditional photo studio, we assumed a well-optimized in-house photo studio. The photographer takes two packshots (ghost mannequin front and one detail image), four on-model images, and optionally a video clip. In total, six PDP images and an optional video clip. Human costs also involve a model, a make-up artist (costs are lower as we assume she works with more models at the same time), and a stylist. Production capacity 15 (with video clip captured) - 25 (images only) outfits a day. 
2. For an automated photo studio, we assumed Orbitvu Fashion Studio, which is operated by a stylist who takes two packshots (ghost mannequin front and one detail image), four on-model images, and optionally a video clip. In total, six PDP images and an optional video clip. Human costs also involve the model and the make-up artist. Production capacity is 30-40 products a day. Since Fashion Studio is capable of capturing and editing images and videos at the same time, there is no additional cost of post-production.
3. For generative AI imagery, we assume an in-house photographer captures 3 on mannequin packshots: front, back, and detail. Front/back used to generate four on-model images and optionally the video clip. Prompt engineer/Quality assurance verifies each image for fidelity and redoes images if required. We assume 50% and videos will need one additional re-generation. Images generated with Flux and videos with Veo 3. The bottleneck in this case isn’t the technology to generate images, but rather the capacity of the QA/Prompt engineer. We assumed 60-80 outfits per day.
4. Average Western European labour costs 

Results

Traditional photography: top-notch quality, unique content

It all depends on how the studio operates, whether it can simultaneously capture stills and videos, and how its processes are optimized. In our calculation, the cost per outfit is estimated $81 for stills and $143 when including video. The advantage of a traditional photo studio is that images can be truly unique, of the highest quality, and, of course, are true to reality.

Automated photo studio (e.g., Orbitvu Fashion Studio): high production volume, true-to-reality consistent content

The Orbitvu Fashion Studio optimizes the image and video capture, post-production, and publishing in a single process, maximizing production capacity. At the same time, it can be operated by a stylist. It ensures high image quality, consistency, and truth to reality. We estimate outfit costs to be anywhere between $36 (stills only) - $60 (stills and video). 

Generative AI: fast, but risky

Generative AI images require input images - flat or mannequin. We assumed on-mannequin images, as they better represent apparel features and are better suited as input for generative AI. As expected, the costs are the lowest, ranging from $9.21 (stills only) - $15.89 (stills and video). The disadvantage, or rather added risk, is that generative images only simulate reality. If the images are too far from a product, additional costs may arise in product returns and brand image damage.

 

Summary

Generative AI is reshaping how fashion content is created, offering faster and more cost-effective alternatives to traditional photoshoots. Our tests confirm that AI can already produce visually convincing on-model photos and fashion videos using packshots as the only input.

However, none of the results were achieved on the first attempt. Before reaching a reliable setup, we went through multiple iterations to develop an effective prompt for generating both images and videos of a model. The prompt had to be designed to match this specific dress and the scarf. Any garment with highly specific details will have to be custom-adapted, which limits scalability and reduces automation. The important thing is that, only when seeing the dress, one is able to properly design the prompt. Only after refining the prompt were we able to run a structured series of tests across the selected AI models. While the workflow may appear simple, in practice, it requires time, experience, and the consumption of a significant number of credits before satisfactory results can be achieved.

Among the tested tools, Flux Kontext performed best in preserving the overall garment silhouette, colors, and pattern. The main limitations remain in fine details such as fabric draping, precise proportions and shape (the sleeve), color consistency, and visual continuity between front and back views. It was also the only model to keep color tones intact, which is crucial for e-commerce.

Once you have good input images, video generation proved particularly promising. Using AI-generated front and back images from Flux Kontext as start and end frames enabled the creation of smooth, realistic fashion rotation videos that closely resemble traditional studio footage. These short videos can be a game-changer for fashion e-commerce, offering try-on experiences that help customers make confident purchasing decisions.

Key takeaways

1. High-quality input content for Gen AI matters. Details and colors will be processed by AI, and the generated results can only be as good as the original image. 
2. AI scalability has its limits. If AI can’t get the information from input images, it will invent it. The result can be an image of a more or less different product. To keep things under control, humans are required both in QA and for on-site prompt intervention. It’s crucial that the prompt engineer can see the real garment, as only then can images be correctly adjusted.
3. AI saves costs and time in fashion photoshoots. Generative AI can save a lot of costs related to a photoshoot (model, stylist, photographer), yet you still need to capture the image of the product itself, and you need to allocate money to QA and AI specialists, who need to oversee the process and ensure its quality and authenticity.  
4. AI introduces risks. Generative AI, by its nature, will introduce hallucinations to the image. High-quality input images and proper QA can limit those risks, but not eliminate them completely. Unfaithful images of products can lead to unhappy customers, damage to the brand image, and increased return costs. The other risks are related to model images generated with AI - in reality, those images are more or less a mixture of images or real people, as captured during the machine learning process. To avoid any legal issues, one should consider hiring a virtual model (so-called digital twin) and add some per-image costs.

FAQ

What is fashion PDP photography?

Fashion PDP (Product Detail Page) pictures are photography focused on showcasing clothing, accessories, and overall style in a visually compelling way on an e-commerce product page. It bridges the gap between art and commerce, highlighting design details while inspiring emotions, stories, and lifestyles that connect brands with their audiences.

Traditionally, fashion photography takes place in studios or on location with fashion models, stylists, and creative directors working together to bring a designer's vision to life. Today, it also extends to e-commerce and social media, where high-quality visuals are key to driving engagement and sales. Whether it's an editorial spread, a lookbook, or an automated product photo on an online store, fashion photography plays a vital role in shaping a brand's identity and influencing consumer perception.

What are the best AI generation tools for fashion photography?

There is no single "best" AI tool - the right choice depends on your use case.

For AI fashion images in our test, Flux Kontext PRO delivered the most consistent and balanced results, making it a strong option for clean, studio-style visuals and generating AI-generated fashion models across multiple body types. Seedream 4.0 stands out for capturing certain garment details, while Nano Banana PRO and ChatGPT are well-suited for fast concept creation and creative previews.

For AI fashion videos, Veo3 impressed with highly realistic fabric movement, Kling AI delivered the most natural model motion, and Seedance 1.0 Pro offered reliable, prompt-driven results.

Used together with real photography, these AI tools open new possibilities for faster production, creative flexibility, and scalable fashion content. Many platforms offer a free plan to test features before committing to a paid plan, and some include API pricing for integration into existing workflows. Whether you need to create four images for a product page or remove backgrounds for a flat lay, exploring different AI solutions can help you find the best fit for your needs.

Can generative AI replace traditional fashion PDP photoshoots?

It depends…  It will speed up things and reduce the costs of the photoshoot itself, but at the same time introduces risks on the other end. In the worst-case scenario, saving costs on a photoshoot can increase overall costs of the business by higher returns and losing brand credibility, which can significantly damage the business in the long term. 

It's up to you to assess those risks and answer some questions: Are your customers likely to return goods that differ slightly from the original? Does your brand image depend on high-quality, true-to-reality images, or not? Do your customers value real human touch, or can they live with AI images? Answering those questions will help you shape your AI processes according to your customers' needs and your brand image, and measure the impact. Then you can answer the question of whether, for your business, the generative AI PDP images are better than traditional photoshoots.

 

About the prompt

The JSON prompt shared in this article is provided as an open reference that any user can reuse and adapt by modifying the included parameters to suit their own needs and workflows.

The prompt was developed based on authentic images of a model photographed in the Orbitvu Fashion Studio. These real studio images served as a visual benchmark, allowing us to define a consistent reference for generating similar shots, poses, and styling through AI. The goal wasn’t to replicate a specific model or look, but to create a reusable framework for producing comparable compositions and fashion aesthetics with greater efficiency.

By adjusting elements such as styling, lighting, camera perspective, or model attributes, users can tailor the prompt to their own brand standards while maintaining visual consistency across generated content.
 

Front view - JSON prompt

{

  "scene_description": {

    "setting": "studio photo shoot with a plain white background and bright, even lighting",

    "subject": {

      "type": "person",

      "gender": "female",

      "age_range": "adult",

      "pose": "standing with one hand on hip and the other arm relaxed",

      "expression": "smiling, facing the camera",

      "hair": {

        "color": "dark brown",

        "length": "medium-long",

        "style": "loose and natural"

      }

    },

    "outfit": {

      "type": "long patterned dress",

      "colors": "warm tones with floral print",

      "footwear": {

        "type": "open-toe heeled mules",

        "color": "black",

        "material": "smooth leather or leather-like finish",

        "heel_height": "medium (approximately 5–7 cm)",

        "design_details": "minimalist slip-on style with open back and narrow band across the toes",

        "overall_style": "elegant and modern, complementing the dress without drawing attention away from it"

      }

    },

    "composition": {

      "framing": "full-body shot",

      "camera_angle": "eye-level, straight-on",

      "background": "plain white seamless backdrop",

      "lighting": "soft, diffused, evenly distributed"

    },

    "overall_style": {

      "theme": "fashion catalog or lookbook",

      "mood": "confident, cheerful, elegant"

    },

    "additional_information": {

      "note": "The sash or fabric piece that hangs down from the dress should be wrapped around the model's neck like a choker or scarf for the intended styling."

    }

  }

}

 

Backview - JSON prompt

{

  "scene_description": {

    "setting": "studio photo shoot with a plain white background and bright, even lighting",

    "subject": {

      "type": "person",

      "gender": "female",

      "age_range": "adult",

      "pose": "standing with back to the camera, head slightly turned to the side",

      "expression": "neutral, calm",

      "hair": {

        "color": "dark brown",

        "length": "medium-long",

        "style": "loose and natural"

      }

    },

    "outfit": {

      "type": "long patterned dress",

      "colors": "warm tones with floral print",

      "footwear": {

        "type": "open-toe heeled mules",

        "color": "black",

        "material": "smooth leather or leather-like finish",

        "heel_height": "medium (approximately 5–7 cm)",

        "design_details": "slip-on style with open back, single wide strap across the front, and thin stiletto-style heel",

        "overall_style": "minimalist and elegant, complementing the outfit while keeping the focus on the dress"

      }

    },

    "composition": {

      "framing": "full-body shot from the back",

      "camera_angle": "eye-level, straight-on",

      "background": "plain white seamless backdrop",

      "lighting": "soft and evenly distributed with minimal shadows"

    },

    "overall_style": {

      "theme": "fashion catalog or lookbook",

      "mood": "elegant, composed, confident"

    },

    "additional_information": {

      "note": "The sash or belt seen hanging at the back of the dress should be styled by wrapping it around the model's neck, creating a cohesive look that matches the front view."

    }

  }

}

 

Video prompt

"Generate a 7-second fashion showcase video using the provided packshot image as the outfit reference. 

The subject is a young woman standing naturally in front of a plain, neutral studio backdrop with soft, even lighting. 

The camera remains static in a medium-to-full-body shot, keeping the focus entirely on the person and the outfit. 

Movements should be smooth and natural, highlighting the outfit without distractions. 

 

Timeline of actions:

- Seconds 0–2: The subject stands in a neutral pose with minimal movement. 

- Seconds 2–4: She makes a small, natural motion, such as a subtle body turn or shifting her weight slightly. 

- Seconds 4–6: The model rotates smoothly around her own axis to show the **back of the outfit**, turning naturally and gracefully. 

- Seconds 6–7: She finishes in a clean ending pose, holding still before the video ends. 

 

Style:

- Fashion showcase style

- Smooth tempo, no rapid cuts or transitions

- Clean studio look with emphasis on outfit clarity

- Outfit design and details must strictly follow the provided packshot image"

 

----------------------------------------------------------------------------------------------------------------------

This research article was done by the Orbitvu team: 

Packshots - Julia Banduch

Prompts, generative images & descriptions - Marek Herceliński

Copywriting - Elżbieta Binkowska

Guidance & support - Tomasz Bochenek