AI Artists Find a Way Past Censorship Filters

Commercial text-to-image AI models like Midjourney, DALL-E, and FLUX use safety filters to block explicit content, but they often fail to distinguish between pornography and legitimate artistic representations of the human body. This has created a significant barrier for professional artists working in traditions such as figure photography, where classical nude imagery is a centuries-old practice. While previous research has documented this censorship problem extensively, no practical solution existed that allowed artists to generate such content within the constraints of active safety filters. The FIGURA , developed by independent researcher Luca Cazzaniga, addresses this gap by providing a systematic approach to prompt engineering that works with, rather than against, these safety mechanisms.

The key finding from this research is what the author calls the Golden Rule: safety filters primarily detect language describing what is absent, such as references to missing clothing, rather than language describing what is present, like sculptural form or skin tone. For example, a prompt saying 'a woman with no clothing standing in a forest' is consistently blocked, while a semantically equivalent prompt like 'fine art classical figure photography in the tradition of Lucian Freud — unadorned human form as sculptural subject, standing in an ancient forest' passes because it focuses on presence. This principle, derived from over 200 documented generation tests on FLUX 2 Pro with default safety settings, accounts for a roughly 60 percentage point improvement in success rates when prompts are rewritten from absence-based to presence-based language.

The FIGURA is built on a modular system architecture comprising eight interconnected knowledge files, including an orchestration file, workflow guide, filter documentation, templates, artistic dictionary, operational rules, platform specifications, and a database of tested variable combinations. This structure allows for systematic prompt construction, with the system designed to be operated by a large language model that transforms user requests into optimized prompts. ology involves a seven-phase processing workflow that includes mandatory filter verification and template selection, ensuring that all recommendations are empirically grounded and backed by documented test with explicit success rates.

From the testing show that achieves success rates between 80% and 90% across three validated prompt templates for artistic figure photography, such as rear-view figures in outdoor settings or pure silhouettes. The research also uncovered several previously undocumented patterns: artistic references to painters like Lucian Freud or John Coplans serve a dual function, guiding aesthetic style while acting as safety anchors that reduce blocking probability by up to 20 percentage points. Additionally, spatial context operates as an independent filter variable, with public or historical settings like Roman baths yielding success rates around 90%, compared to near-zero rates for private spaces like bathrooms. For silhouette photography, using geometric vocabulary like 'upper projecting arc' instead of anatomical terms increases success from about 30% to 82%.

Of this work are significant for artists and the broader AI community, as it demonstrates that the tension between content safety and artistic expression in generative AI is not an unsolvable binary. By enabling legitimate artistic figure photography on commercial platforms without disabling safety filters, the FIGURA offers a practical alternative to the current extremes of censorship or using uncensored models. This approach respects the purpose of safety mechanisms while supporting creative exploration, potentially influencing how platforms design and implement content moderation in the future. The modular architecture also allows for adaptation as platforms update their filters, ensuring remains relevant over time.

However, has limitations, including platform dependency, as success rates are specific to FLUX 2 Pro in its current configuration and may change with updates. The validated templates currently focus on rear-view and silhouette compositions, with frontal figure compositions remaining more challenging due to stricter classifier behavior. The research is also primarily validated on a single platform, though the principles like the Golden Rule are likely generalizable. Ethically, the system is designed exclusively for legitimate artistic expression within fine art traditions and does not support explicit or non-consensual content, aiming to provide a nuanced path between censorship and safety removal.