Slim Aarons Generative Studio

A style-preservation and reinterpretation workflow using LoRA fine-tuning.

This project develops a reproducible pipeline for preserving and extending Slim Aarons’s mid-century photographic aesthetic through ethically transparent generative AI workflows.

A curated dataset of high-resolution Slim Aarons photographs was used to train a LoRA that captures compositional structure, color palette, lighting, and subject framing without replicating specific originals. The trained model was deployed through ComfyUI workflows and a Gradio web app, enabling accessible experimentation across text-to-image, image-to-image, editing, animation, and 2D-to-3D outputs.

Overview

Methodology and System Design

Gradio Web App

A browser-based interface provides:

  • Multiple generation modes

  • Sliders for LoRA strength, guidance scale, denoise strength

  • Seed control for reproducibility

This enables use by non-technical users in teaching and research contexts.


Additional tools generate:

  • GIFs and cinemagraphs

  • Ken Burns-style pan and zoom

  • Multi-frame storyboard sequences

These extend still imagery into short-form narrative outputs.

2D - 3D Pipeline

Ethics & Provenance

  • Outputs explicitly marked as AI-generated

  • Dataset provenance documented

  • Internal trigger token retained for traceability

  • Project framed as educational and preservational, not substitutive


Project developed in collaboration with Biel Pitman. Original article can be found here.


Generated images were converted into textured 3D meshes using Hunyuan3D, producing lightweight assets suitable for VR, games, and spatial visualization.

Data Curation

1

Prompt and Trigger

2

Model Strategy

3

  • 150 high-resolution images selected across eras and locations

  • Each image paired with a detailed descriptive caption

  • Captions encoded composition, color relationships, lighting, subject positioning, and mood

A standardized caption structure was introduced, along with a dedicated trigger token:

“SLMRNS A Slim Aarons photograph of”

This ensured consistent invocation of the learned style during generation and provided a clear marker of stylistic conditioning.

Training Setup

4

Quality Control

5

  • Resolution: 1024 × 1024

  • Hardware: single A100 GPU

  • ~2,000 training steps

  • Tuned LoRA rank and text encoder parameters

  • Periodic checkpoints and sample review for quality assurance

  • Base model: FLUX.1-dev

  • Fine-tuning method: LoRA

  • Rationale: modular, lightweight style injection without altering the base model


Outputs were evaluated during training for:

  • Composition accuracy

  • Color fidelity

  • Architectural and clothing detail

  • Consistency across varied prompts


The final LoRA was exported as a compact safetensors file for reliable reuse.

The project delivers a reproducible Slim Aarons Generative Studio supporting research, teaching, and creative exploration. Planned extensions include usability testing, rights assessment for broader release, Hugging Face demo deployment, and expansion toward responsibly curated style-preservation toolkits.