Dralim: a Transition Noise is a photographic project that explores the role of noise within artificial intelligence tools known as Diffusion Models—systems used to generate synthetic images by reversing random noise into coherent visuals. The project investigates how this noise, while seemingly neutral and mathematical, becomes a carrier of cultural bias, inherited from the datasets on which these models are trained. Through this lens, the project aims to reflect on the hidden influence of AI in shaping visual culture, raising questions about authorship, representation, and the way technology reproduces human assumptions.

Diffusion Models are a class of generative AI systems that create images by progressively denoising a field of random Gaussian noise, guided by learned patterns from large-scale image-text datasets. During training, the model learns to reverse the process of adding noise to real images, enabling it to reconstruct visual content from pure statistical randomness. At inference time, a text prompt steers the denoising trajectory, allowing the model to generate coherent and detailed visuals that align semantically with the input.

Datasets are large collections of images paired with text descriptions and labeled by category, used to train the AI. They teach the model how to associate visual elements with language and how to transform random noise into coherent images. The broader and more diverse the dataset, the more capable the model becomes at generating complex and realistic visuals. Most of these datasets encode pre-existing cultural, aesthetic, and social biases, structured by classification systems and recurring visual patterns—factors that influence how AI “learns” to see.

Datasets used to train diffusion models are often labeled through manual or semi-automated processes, where human workers pair images with descriptive text or assign conceptual categories. This labor is frequently outsourced to large pools of remote workers, many of whom are based in low-income regions and receive minimal compensation for each micro-task. Despite being essential to AI training, this largely invisible and undervalued human contribution raises ethical concerns over precarious labor and the exploitation of cognitive effort.

However, the project raises then a critical question: what happens when these datasets—inevitably shaped by cultural biases—become the foundation for generating new images?

Rather than producing neutral representations, AI-generated visuals often reflect dominant narratives—mirroring societal norms around identity, beauty, and belonging. These images, shaped by selective training data, can subtly reinforce existing power structures, embedding pre-existing cultural frameworks into new visual forms.

Through a series of evocative photographs captured on film, the viewer is invited to critically reflect on the role of artificial intelligence in shaping visual perception and on how cultural biases are absorbed and reproduced through generative systems. In an era where the boundary between authentic and synthetic imagery is increasingly unstable, noise—paradoxically—emerges as a new language of representation: a generative element that not only structures images from randomness, but also carries with it the implicit assumptions of the data it stems from, subtly influencing how we interpret and relate to what we see.