The European Commission is about to ban so-called nudify apps – tools that use neural networks to digitally remove clothing from photos of people. The measure, expected in the coming weeks, aims to hit a market of mobile and web applications that has raised privacy and abuse alarms for years. But a new report from the European non-profit AI Forensics tries to dismantle the regulatory framework before it even takes effect: the ban is designed for digital stores like the App Store and Google Play, while the real nudification tools live elsewhere – on Hugging Face, the platform that hosts open-source models, datasets, and machine learning pipelines.

The technical paradox is instructive for anyone designing on-premise AI architectures. Hugging Face is not an app store, but a model hub: anyone can download Stable Diffusion checkpoints, GANs, or variants specialized in so-called deepnude, and load them onto their own hardware. AI Forensics tested seven of the most popular image-editing tools on the platform, finding several capable of generating artificial nude images from clothed photos. The finding is not surprising for the technical community – fine-tuning diffusion models on specific datasets has been documented for years – but it shifts the debate from app regulation to model control, a level where legal boundaries are far blurrier.

From the perspective of those evaluating on-premise deployment, the episode highlights a short-circuit that goes beyond this single case. If a model is available in PyTorch or SafeTensors format, it can run on a workstation with a consumer GPU or on an enterprise server without going through any software intermediary. A distribution ban through stores hits the surface but leaves untouched the ability to perform inference locally, even on air-gapped machines. This means the control burden shifts downstream, to training datasets and weight publication – areas where the jurisdiction of a single regional bloc becomes fragile.

There is a second, more structural effect. The entire pipeline of open-weight generative models – from image synthesis to LLMs – rests on platforms like Hugging Face, which act as a neutral infrastructure layer. If the regulatory response focuses on packaged apps, it creates a perverse incentive: it pushes less careful users toward self-hosting, where guardrails are absent, and leaves model developers in a gray area. For organizations already weighing TCO and data control as primary criteria, this scenario reinforces the on-premise choice: a GPU with enough VRAM and an inference script are enough to replicate functionality that would be blocked elsewhere.

The tension between regulation and model access is not new, but the nudify app dynamic offers an almost textbook distillation. The European ban risks being circumvented not through legislative loopholes, but because of the very nature of contemporary AI infrastructure: distributed, replicable, and increasingly cheap to run locally. As AI Forensics continues its tests, the underlying question remains open: how long can packaged software be regulated when code and weights travel on channels that do not respond to the same logic of control?