Generative Art

Redacted Bodies

A custom-trained LoRA model exploring queer identity, censorship, surveillance, and digital embodiment

The Model

NoirmakStyle

Redacted Bodies is built around NoirmakStyle, a custom LoRA (Low-Rank Adaptation) model trained on FLUX and developed in ComfyUI. The model wasn't created simply to generate images. It was built as a way of asking what happens when identity becomes training data, when bodies are translated into metadata, and when machine learning systems begin interpreting and reproducing those bodies.

Every image used to train the model was selected by hand. The images surfaced from a live coding sketch created from a live performance Hydra code sketch. The dataset is intentionally small and carefully curated, allowing the model to learn a visual language rooted in themes of queer identity, censorship, visibility, surveillance, and transformation. Rather than treating AI as an autonomous creator, the model becomes another artistic medium, shaped by thousands of deliberate decisions about what is included, what is excluded, and what relationships the system is allowed to learn.

NoirmakStyle also powers Signal Bloom, my live-coded audiovisual performance system. The same model moves between two very different contexts: one exploring emergence, ritual, and transformation through performance, the other examining the politics of visibility and the ways machine learning reframes the human body. The technology remains constant while the artistic intent changes, revealing how context can reshape the meaning of the same visual language.

The generated images are only one layer of the work. Dataset curation, model training, prompt engineering, and the ongoing selection of generated outputs are all part of the artistic process. Every prompt is a negotiation between human intention and machine interpretation. Every image reflects both what the model has learned and the boundaries placed around that learning. The collection is less concerned with producing endless variations than with documenting an evolving conversation between artist, algorithm, and the politics of representation.

How the Work Is Made

The Pipeline

01

Dataset Curation

Every image was selected by hand, building a deliberately small dataset that teaches the model a visual language shaped by queer identity, visibility, censorship, and transformation.

02

Training

The curated dataset fine-tuned FLUX through LoRA in ComfyUI, allowing the model to learn a consistent visual vocabulary of chromatic bloom, signal corruption, and fractured identity.

03

Prompt Engineering

Generation begins with language. Prompts guide the model toward specific conceptual questions, making each image a negotiation between artistic intention and machine interpretation.

04

Inference & Curation

The model produces possibilities, not finished works. The collection emerges through selection, where curation becomes as important as generation itself.

A Note on What You're Seeing

Every image in this collection is generated by NoirmakStyle. None are photographs, and none are images from the training dataset presented as finished work.

Redacted Bodies is ultimately less interested in producing endless AI variations than in examining the relationship between artist, algorithm, and representation. The gallery documents that conversation: how a carefully constructed dataset, a trained model, and human judgment come together to produce images that question visibility, identity, and the ways machine learning learns to see us.