AI-generated models offer fashion companies a new way to create product images. Deep Agency, a platform launched by Headlime founder Danny Postma, lets customers place virtual models in digital settings, with plans for uses such as social media content and ecommerce photography. Its debut has also brought a difficult question into focus: who benefits when fashion images can be generated without a photographer or a human model?
A digital studio with rough edges
Postma described Deep Agency as an “AI photo studio and modeling agency.” Customers can create models and arrange them against digital backdrops. The service was offered at a starting price of $29 per month for a limited time.
The platform was still a proof of concept when it launched. Images showed artifacts in models’ faces, and users had limited control over some generated details. For example, attempts to create a female model in a specific outfit, such as a police officer’s uniform, did not work. The platform also appeared to restrict which physiques it could produce.
Those shortcomings matter for a tool pitched as a way to make fashion imagery. Brands need images that match specific products, people and settings. If a model is difficult to steer or produces inconsistent results, the savings and convenience may be less useful than the concept suggests.
Who gains when a photo shoot becomes software?
Reaction to the launch was split. Some users welcomed the prospect of virtual models for clothing and apparel brands. Others accused Postma of building a business on other people’s photographs and likenesses, then selling the resulting service for profit.
That argument reaches beyond one startup. Fashion models and photographers often work as independent contractors for larger companies looking to cut costs. Models may also pay agency commissions and expenses such as travel, housing and promotional materials. Os Keyes, an ethical AI researcher at the University of Washington, said the creative workers most exposed to these changes can have little bargaining power.
AI model imagery could reduce demand for some paid assignments if companies use generated people in advertising or product photography. At the same time, the source article does not establish how widely brands will adopt the tools or how many jobs they might replace. The concern is about the direction of the business: a company may capture value from a system while workers whose labor and images contributed to its development receive no share.
Deep Agency had not announced a revenue-sharing plan for artists or other contributors. Shutterstock, by comparison, had created funds to share revenue from AI-generated art with artists and was experimenting with payments to creators whose work is used to train models.
Training data, consent and privacy
Image-generation systems such as diffusion models learn to produce pictures from text prompts using training data gathered from the web. Artists have argued that these systems can reproduce elements of copyrighted images and that their work may have been included without permission. Companies have maintained that fair use can protect training on licensed content, while the legal questions remain contested.
For Deep Agency, the source of the training images was not transparent. That makes it difficult for artists to determine whether their work was used. The platform also offered no way for artists who suspect their work was included to remove it from the training dataset. Other services, including DeviantArt and Stability AI, provide opt-out mechanisms.
Users face a separate privacy concern when making a digital twin. Deep Agency allowed customers to upload around 20 images of a person in different poses. According to the platform’s terms, those images could also be added to training data for higher-level models unless users deleted them. Its privacy policy did not specify exactly how uploaded photos were handled or where they were stored.
There was also no apparent way to stop someone from creating a virtual twin of another person without consent. That possibility is especially concerning given the use of image-generation models to make nonconsensual deepfake nudes. The terms further acknowledged that other users could generate similar or identical images, complicating the promise of a unique digital model.
Representation is part of the product
Training data can shape which people an image-generation system readily depicts. If the source images are not representative, generated results may reproduce existing racial, ethnic and gender stereotypes. Research cited in the source article found that models including Stable Diffusion and DALL-E 2 tended to depict people as white and male, especially in positions of authority.
Vice’s Chloe Xiang reported that Deep Agency generated images of women unless users bought a paid subscription. The platform also tended to favor blonde, white female models, even when a user selected a woman of a different race or likeness. Changing the appearance required adjustments that were not obvious.
These limitations complicate claims that virtual models will automatically broaden representation in fashion. Levi’s said it would use AI-generated models alongside human models and that the partnership would not affect hiring plans. Still, the move prompted questions about why brands would generate diversity digitally instead of hiring more models with the characteristics they want to show.
Deep Agency’s early problems may improve, but the larger issues are not limited to image quality. Fashion companies considering AI models must weigh convenience against the effects on creative workers, the consent and privacy of people in uploaded photos, and whether the systems can represent people fairly. The technology’s commercial promise is clear; how its costs and benefits are shared is less so.